Subscribe now

DEAR STAGE 2: We’re a Series B company scaling our GTM team and trying to evolve our interview process to make sure we’re maximizing time with candidates, getting to decisions faster, and actually screening for the skills we need. How has interviewing or hiring changed in an AI-native world? ~INTERVIEW FATIGUE

DEAR INTERVIEW FATIGUE: Talent is always top of mind, but even more so right now. We’re all feeling the push to find efficiency, drive AI adoption, evaluate our teams and then train, retrain and uplevel. Every board meeting I’ve been in this year has had a variation of this topic - do we have the right people for the next stage of growth? Do we need to change our hiring profile? How should our hiring process evolve?

This week I chatted with Scott Peyser, a Stage 2 LP and Head of Revenue Strategy, Operations and Readiness at Asana. Scott’s team runs RevOps, GTM Excellence, Solutions Engineering and full cycle SMB sales. He’s built and rebuilt sales interview processes at scale, and as we dug into this topic, he was adamant: “the tools may have changed, but the traits that make someone great at this job did not.”

Many people are assuming we need to start again and completely rethink how we hire for an “AI-native team”, but Scott’s push is really to get back to the basics - define the job, be clear in what you are hiring for, understand how you are going to evaluate it and then do that consistently at scale and over time.

I want to share 2 takeaways from our conversation that are particularly actionable - the first about Asana’s interview process and the second about Scott’s interview learnings throughout his career:

1. Define the traits (and what good looks like)

Scott is in the process of revamping the interview process for GTM hiring at Asana. He started by bringing the sales leaders together for a whiteboard exercise: each leader listed everything their best sellers do well and generated a list of about 45 items total. Then everyone marked their top 5. The lists converged on five qualities:

1. Curiosity and questioning

2. Grit and accountability

3. GM mindset

4. Resourcefulness

5. Business acumen

With a short, specific list everyone agrees on, the interview process becomes more focused. As Scott puts it: “You can be prescriptive, but if the prescription is wrong, it doesn’t matter.” Once you have the list, you need to take the time to write a shared definition of what good looks like. This work is critical to laying the foundation for a great interview process and candidate evaluation.

If your team isn’t screening for anything specific it’s because nobody ever defined what a strong answer sounds like. When five interviewers each bring their own favorite question with no shared definition of what good looks like, you get (near-total) overlap and almost no signal on the candidate.

Asana also moved to live working prompts. One example: “a customer is at risk and you’ve only sold into a single buying center. How do you triage the risk and expand into a second buying center, live, in the room?” For leadership hires, they hand over a take-home data exercise: here’s our data, we’re planning for 2027, what’s missing, what would you want to make a real structured decision? That puts GM mindset and business acumen under actual pressure, instead of asking for a story about the last time someone showed them.

2. Hiring like Scott

Scott has hired a lot of people in his career and I wanted to hear what has worked for him. Regardless of role or company, Scott personally screens for two things when hiring a new member of his team: intellect and drive. And he has 2 questions he likes to ask every candidate:

1. “You’ve obviously been really successful. What do you think’s prevented you from being more successful than you’ve already been?” People who point to something internal (a skill gap, a habit, a blind spot) have a shot. People who point to a bad territory or an unsupportive manager don’t. This question is a read on accountability - does the candidate look to internal or external factors? Do they look for things they can control or play victim?

2. He also asks candidates to imagine the job worked out: “In a couple weeks you’re with family and friends and they ask what you’ll be doing. How do you answer?” People who’ve actually done the work on your company and your market answer specifically. They’ve thought about what the job will entail and can envision themselves doing it. This one really surprised me, but it’s easy to imagine someone going through the motion of an interview process without actually internalizing the work.

This was a very grounding conversation for me because it brought back the basics. I can almost guarantee someone in your interview panel is asking “how do you use AI in your sales process?” as if the answer is a proxy for talent. Yes you want someone who is using AI, but that’s because they are curious and motivated and will use whatever the current tools are available to succeed.

Until next week!

Subscribe now

DEAR PLG v. SLG: Before we dig in, let’s acknowledge the reality that you have the hardest thing solved: real demand. Plenty of founders would trade places with you in a heartbeat.

I caught up with Prem Iyer last week, who has deep experience building enterprise motions inside PLG businesses. We talked about what happens when strong bottom-up adoption starts creating pressure to turn that usage into larger contracts, deeper commitments, and enterprise revenue. Here are a few ideas from our conversation:

Reframe what “not a customer” means

No contract does not mean you are starting from zero. In fact, Prem described the opposite based on his recent experience at Cursor. “We have 50,000 customers and ~70% of the Fortune 500. And the ones that are not customers? They might have 80 developers paying for and using it on their own. They’re just not on an enterprise agreement with us yet.” His language for those accounts is worth noting: “I always call them underpenetrated, not zero LTV.” On the enterprise side, the job is conversion. You’re turning usage you already have into a signed commitment.

Timing matters

The reason to move now with a sales-led motion is that your organic lead is on borrowed time. Prem was direct about the dynamic he saw at Cursor: developers picked Cursor and “just did it, whether their CIO or anybody signed off or not.” But once competitors show up, “they’re going to come and try to take that from us. Now we’re moving to SLG so that we can get enterprise commitment”. My takeaway? Every month a company runs on individual self-funded seats is a month you’re susceptible to a competitor. Enterprise contracts allow you to convert user love into locked-in agreements.

Your first sales hire is not any old salesperson

This is where PLG companies often go wrong. They hire a polished enterprise rep who knows how to manage a deal process, and then wonder why it’s not working. What you need is someone who can create a process where there is none. Prem described the candidates he looked for as “glass eaters.” His description: “These are true hunters. They’re not farming accounts. They’re not waiting for a PO. We have to go build it, go into the account, talk to 20 personas, and build the whole thing ground up.” You have to screen for the person who lights up at the idea of multi-threading and working an account bottoms-up and tops-down in parallel until they find a path in.

Bring the CFO in early, and lead with ROI

The magic of PLG is that your ROI case already exists within your user base and usage patterns. Prem shared a story about walking into a Fortune 150 account where he got less than a minute with the CFO. His whole pitch: “If I had a way to show you we could save 30% on your software development economics and add 40% to productive time for the team, would you have that conversation?” He said, ‘we’d have to’. If you are a true PLG product, you have the usage data and can tell a story. Find it, make it defensible, and lead with it!

Making the leap from PLG to SLG is about taking what’s working and wrapping a commercial layer around it. Keep the self-serve on-ramp exactly as it is (don’t mess with a top of funnel that’s working), and layer a small team of genuine hunters whose only job is to turn “everybody here already uses this” into a signed, multi-year yes.

We’ll leave you with one question: do you actually know which of your logos are most underpenetrated right now? If you can’t answer that right now, that’s where you start. You need to have the data infrastructure in place to fuel the sales-led motion.

Until next week!

Subscribe now

DEAR ENGINEERING GTM: Let me start with a slightly spicy take: the “GTM engineer” is pretty much just rev ops. Anyone who did rev ops well (aka not order-taking/administrative) was already doing a lot of this work: instrumenting the funnel, wiring the tools together, running go-to-market like a system instead of a gut feel, and acting as a strategic partner. But AI has given this role real leverage. Title aside, you probably need the function, and the real questions are 1) when and 2) how to make the right hire for your unique setup.

When I saw this question come in, I immediately thought of Kyle Henson, founder of Arda (a Stage 2 Portfolio company). Kyle recently hired a GTM Engineer and he graciously agreed to take you through their journey:

When is the right time to bring in a GTM Engineer?

The “why now?” for Arda happened when the teammate who was driving top of funnel went on paternity leave last year, and Arda didn’t close a single sale that month. Their entire pipeline generation motion lived with one person, which was a glaring single point of failure in their GTM motion. Kyle got to work building with Unify and Clay, and put together a scrappy two-touch sequence, and …it worked. But he needed the motion to run without his involvement.

It’s time to start thinking about this hire once you’re seeing signs of your go-to-market working and you need to turn it into a system.

Scope the GTM Engineer role as a builder, not a report-runner

Kyle’s north star is a fully hands-off funnel for SMB manufacturers, and that means someone needs to own orchestration, testing, and optimization.

Kyle wanted “a scrum master for our go-to-market motion,” someone to run GTM like a product practice: form a hypothesis, build it, measure it, kill it, repeat. Two non-negotiables: fluency with tools like Clay and Unify, and the ability to vibe-code. At a recent conference, instead of buying the overpriced lead-retrieval app, his hire bought API access and built a purpose-built tool on top of it for a fraction of the cost - that exemplifies the job.

Write a GTM Engineer job description that actually attracts builders

Kyle wrote a job description that covered the standard elements of the role. But he also added an unconventional line at the end: “the best candidates will likely proactively figure out how to get our attention.” Applicants took this seriously. One person built a Yelp scraper to surface local manufacturers who aren’t on LinkedIn, a channel Kyle hadn’t even mentioned. Beyond creative and shows real initiative.

The final step in the interview was a paid project. Kyle paid the finalist for a day of work as a contractor, gave them HubSpot access, and worked together on a real problem. There are two things about this approach that I particularly liked: First, Kyle aligned on scope with a quick call before the candidate started, so nobody burned time moving in the wrong direction. Second, the prompts were vague on purpose - that’s the real test. Kyle wanted to understand how candidates would work in an ambiguous environment.

Here are three actual scenarios Kyle challenged the candidate to problem-solve:

Scenario 1: “I just landed in [CITY] at [CONFERENCE]. I want to text attendees of the conference and tell them to stop by the booth while the conference is going on. Didn’t have the bandwidth to prep that.”

Scenario 2: “We want to incorporate texting into outbound. We pay for it on HubSpot. Haven’t set it up.”

Scenario 3: “During [CONFERENCE] last year we had the conference send an email blast that got a lot of opens. We got a ton of email addresses and did nothing with them because we were wall-to-wall with demos for three weeks after the show.”

Notice what these have in common: each is a real, unfinished job, there’s no guidance or instructions, and there’s no “right” answer.

I wish more founders and hiring managers would approach designing hiring exercises this way. Your scenarios don’t need to be fully scoped problems. Some of the best tests are ones where you’re not even sure if the work is worth doing. For example, one scenario floated texting as a channel, hoping the candidate would either run with it or push back with a reason. Drew (the eventual hire!), pushed back with a clear “not right for us now, here’s why.” That’s the kind of relationship Kyle was looking for. Drew wasn’t an order-taker and brought real opinions to the table that were backed by experience.

How a paid working session can reveal the right GTM Engineer

A working session does something that interviews can’t: it can show you who acts like part of the team even before they’re hired. Before Drew’s paid project officially kicked off with the prompts above, he’d already been in the systems and come back with feedback. Just a few examples he actually sent:

• Broken Workflow: One sequence wasn’t firing at full volume because a misconfiguration was excluding its own audience. He caught it and fixed it.

• Dropped Leads: A live prospect had replied, and there was no trail of a follow-up in either HubSpot or Unify. A real, potentially good account was…ghosted. He flagged it as a symptom of broken field mappings between the two tools and proposed a bidirectional sync.

• Missed Opportunity: He started scrubbing the dormant conference lists, found about a third of the contacts fell outside ICP, and proposed a Unify play built around what the team learned at the event to route the rest into free trials.

He literally did the job: found the leaks, built the plays, closed the loops. You cannot get this from a resume or a “walk me through your funnel” conversation.

My advice? Don’t hire “a GTM engineer” just because your feed told you to. Instead, look at your own funnel and identify gaps and who can move the needle the fastest. Then make sure you scope the role as a builder who owns outcomes, screen with a paid project, and give them the leeway to make a real impact from day one.

Until next week!

Many Founders describe their ICP the way they’d describe it to a friend at dinner: “mid-market SaaS companies that are growing fast.” But hand that to a rep staring at a list of 200 leads and they probably won’t know whether a 40-person company or a 4,000-person company counts as “mid-market.”

AI is good at the two things that make this exercise less painful: combing through a pile of closed-won and closed-lost deals for the actual pattern, and refusing to let you move on when your answer is still too soft to use. That’s what a Claude skill built around Stage 2’s Green/Yellow/Red ICP worksheet does. The skill walks you row by row and won’t sign off on a cell until it’s a rule. To show how it works, I ran the skill against a fictional company with illustrative deal data. The output gives you a sense of what this looks like.

Subscribe now

Download the Claude Skill

The setup

You’ll need Claude with the ICP Worksheet Builder skill enabled, and whatever deal data you’ve got. A CRM export is best, but you could also just use a list of your best-fit customers and worst-fit customers (either churned or closed lost).

Step 1: Load the skill and bring your deal data

If you have a CRM, export your last 30 closed-won and closed-lost deals and attach it. Then ask:

“Help me fill out my ICP worksheet. Here’s an export of our closed deals from the last two quarters — some are wins, some are losses, and I want to understand the pattern behind each.”

Step 2: Let it find the Green pattern first

Green is usually the easy row — it’s how you already talk about your best customers. Claude will ask what your closed-won accounts actually have in common depending on industry, size, buyer, and use case, and push you past a label (“mid-market”) to a number (“50–500 employees”).

Step 3: Interrogate the losses for Yellow vs. Red

This is where many ICP docs fall apart. Everything that isn't a clean Green gets dumped into 'not ideal' and left there. Claude will ask what was different about the deals that didn’t work, and specifically whether that difference cost you something real. Churn, refunds, a support fire (that’s Red). Customers that may work due to something specific (that’s Yellow).

“These three accounts churned within 90 days of closing — what did they have in common that our best customers don’t?”

Step 4: Assign Priority — and let it push back

Not every row matters equally. Claude will ask which 2-3 rows would actually change a rep’s go/no-go call, versus which are just nice-to-know context. If you can’t name a reason a row deserves top priority, it probably doesn’t.

Step 5: Pressure-test against real accounts before you ship it

Take 5 real customers (including one that surprised you, good or bad), and check whether the finished worksheet sorts them where they actually landed. If a customer you’d call a great fit falls out as Yellow, the criteria aren’t calibrated yet. Fix the row, don’t ignore the miss.

What good looks like

A finished row reads something like “50–500 employees, Series A–C” in Green, “500- 1,000 employees” in Yellow, and “pre-seed, no dedicated RevOps hire” in Red, not “small companies” versus “companies that aren’t ready yet.” Every cell should be something a rep could apply cold, in ten seconds, without calling you.

Watch-outs

• Green creep. If more than half your customer base technically qualifies as Green, the row isn’t discriminating between anything.

• Red without teeth. Red should cost you something specific — churn, refunds, a support burden. If the honest reason is “just not our favorite,” that’s Yellow, not Red.

The takeaway

The hard part of this process is writing answers specific enough that a rep can check them against a real prospect and know right away if it should be green or not. That's where AI helps: it keeps reps focused on the green prospects, the ones most likely to actually close.

Subscribe now

DEAR MEASURING AI IMPACT: You’re asking the right question, but it’s hard to jump from “we’re experimenting” to “measurable ROI” overnight. Start by measuring adoption. You can’t connect AI to revenue until you know who’s actually using it, for what, and how often.

Last week I chatted with Liza Rothschild, Stage 2 LP and VP of Sales at Workday. Their entire GTM organization and her team are in the thick of solving this exact question.

Start With Adoption, Not ROI

Here’s how she’s handling it:

She pulls a weekly usage report on her team that covers prompt volume and time spent in the tools, by rep and by manager. She’s not reading anyone’s individual prompts, and she’s not using it to police. The point is to see where the activity is and surface what’s working, and identify where there are gaps so coaching can happen one-on-one. The second your team thinks usage tracking is a stick, they’ll either game it or resent it.

Then she does something unique: she makes the learning social. Once a month, her top performers share the specific prompts/agents/learnings that drove results, by use case (prospecting, live meetings, expansion). Those get packaged into “plug and play” prompts and campaigns the rest of the team can run without having to recreate or build on their own. This step means you’re not waiting for everyone to figure it out alone. You’re taking what your best rep already cracked and handing it to everyone else - it’s a motion that can scale across an entire team/org.

Sequence What You Measure by Funnel Stage

On what to measure, Liza recommends sequencing it by funnel stage. She started top-of-funnel, where the lift is easiest to see, and the work is most repeatable: building account lists, running look-back analysis, pulling renewal timing, drafting value cases. Get a clear read there before you push AI into the messier middle of the funnel (discovery, negotiation, close). Trying to measure everything at once is nearly impossible.

Early on, the signal is likely going to be soft. Her read after a few months of this was “seemingly a lift, but nothing’s blown out of the water.” That is normal. You don’t want to invent a precise ROI number to make a board slide look good. What she’s doing instead is running parallel testing through rev ops to define the real KPIs over the next few months, rather than forcing a metric before the data supports one.

AI Fluency Is a Coachable Skill

One thing I want to call out is that her top AI adopters are, almost one for one, her top performers. It’s a little chicken-and-egg (do the best reps adopt because they’re already curious and coachable, or does adoption make them better?), but the same pattern shows up elsewhere too. The reps who lean into AI are the same ones who show up to every enablement session, do the practice pitches, and sell the new product first. AI proficiency is just the next tool in the toolbox, so treat it like any other coachable behavior.

Why the Tech Stack Is Shrinking, Not Growing

One last observation from our conversation, and this is a trend I’ve seen across teams, Liza’s tech stack is shrinking, not growing right now. Contracts are consolidating, redundant features and use cases are being cut to ensure budget is freed up for new experiments, and they are often leaning into the AI features of established platforms first (think Gong v. startup GTM tech). The reasoning: data integrity and single-vendor accountability matter when you’re running real customer data through these tools.

Where we landed? Measure adoption before you measure impact, learn out loud from your best rep, sequence by funnel stage, resist forcing a premature ROI number, and treat AI fluency as a coachable performance behavior. Do that for a quarter, and you’ll actually know what’s working instead of guessing.

Until next week!

Subscribe now

DEAR STAGE 2: We’re seed-stage with some inbound trickling in. Everyone’s pitching me AI SDRs. Can I actually replace a human rep with one, or is this hype? ~EXPLORING AI SDRS

• Inbound AI SDRs (engaging warm, high-intent traffic) are working in production

• Cold-outbound AI SDRs are mostly not.

• Don’t buy an AI SDR to manufacture demand you don’t have yet.

• Salesforce buying Qualified (an inbound agent) is a great signal

• OpenAI and Anthropic are hiring hundreds of SDRs - they are NOT automating it away

DEAR EXPLORING AI SDRS: Strong conviction loosely held: the AI SDR is not ready for prime time (aka outbound). Find me a case study or show me a customer all-in on AI SDRs and I’m ready to change my mind, but I just don’t think we are there yet.

I’m seeing *great* applications for inbound (Salesforce acquiring Qualified is a huge signal here), where agents engage warm, high-intent traffic that has already raised a hand. They are not there yet for cold outbound. I compared notes on this with Alton Heller, who works with early-stage founders on building their first outbound motion through SDR Academy (sdrtraining.com), and neither of us has seen a single cold-outbound AI SDR we would point to as working.

Yes AI makes you way more efficient and we’re both big believers in supporting your GTM workflow with AI, but SDRs aren’t going anywhere. Want proof? OpenAI and Anthropic are actively hiring SDRs (OpenAI is actually looking for “Strategic Business Development Representaatives…SBDR anyone?) and Heads of Sales Development right now - go check out those job boards. If AI could do this job today, the leading AI companies would already be doing it. Alton said it well, “if Claude could replace the role, Anthropic would be the first to replace it with Claude.

So the real question isn’t “human or AI.” It’s “what do I point AI at?” Here’s our take:

• Account prioritization and research: Pull your target account list and run it through a tool (or straight through Claude) to rank by the criteria that actually matter and to surface the news and trigger events worth reaching out about. That’s hours of grunt work done in minutes.

• Personalization hook: Let AI find the reason to reach out. Then have a human write the first two sentences and template the rest of the message.

And here’s where it falls down, so you don’t get burned:

• Outbound email at scale: Point AI at “write and send more emails” and it all starts to sound the same. You don’t get more good emails, you get (a lot) more mediocre ones. SMB is an exception here - if your market is large enough and your deal size is small enough, it is likely worth it to play the volume game.

• Context: AI can’t account for the near-win last quarter, the conversation you had at a conference, or the competitor an account just churned from. That context matters for sounding informed and relevant and is where replies come from, and it lives with your rep, not your tool.

• Order of operations: Before you can point AI at anything, you need a proven motion. AI scales what already works - it doesn’t build the motion from scratch. Most early-stage founders don’t have their ICP, a sequence that converts, or a clear definition of what a good meeting looks like. AI won’t figure that out for you. Hire the human and build the playbook before automating with AI.

People still buy from people. For MM/Ent businesses, the right outbound setup is a human SDR armed with AI: the AI prioritizes accounts, does the research, and finds the hook, and the human owns the message and the relationship.

So, to the founder deciding where to spend: don’t buy an AI SDR to manufacture demand you don’t have yet. If you have real inbound, an inbound AI agent can absolutely help you keep up. For outbound, hire the human, and make them dangerous with AI.

Until next week!

Liz

Subscribe now

One of the most common challenges I see our founders face isn’t generating traffic to their site. It’s turning that traffic into qualified meetings. And when you’re running a lean team, especially if you don’t have an inbound SDR function yet, keeping that conversion engine running usually falls to nobody. Visitors show up, poke around, and leave.

A well-built AI agent can fix that without adding headcount. Charles Vandercook, Head of Growth at Qwick, a staffing marketplace for the hospitality industry, built one that does exactly this for about $40 a month. Here’s how he did it.

Why Website Traffic Doesn’t Convert Into Meetings

Most B2B websites do one of two things with inbound visitors: push them to a form (high friction, low conversion) or hope they book a demo themselves (optimistic). Meanwhile, your SDRs are spending time on manual follow-up for leads that were never qualified in the first place.

Charles built an AI agent called Avery to solve exactly this. In a single month, Avery handled 1,400 visitor interactions, routed them by persona, answered support questions, and pushed qualified leads directly to an SDR’s calendar. The whole thing costs roughly $40 a month to run.

This guide walks you through how he did it: the tool, the architecture, and the learnings from four or five iterations.

The Setup: An AI Agent Built in Voiceflow

The tool Charles uses is Voiceflow, a low/no-code platform for building conversational AI agents. No backend dev required. If you can map a flowchart and write a decent prompt, you can build this.

Avery lives as a pop-up widget in the bottom right corner of Qwick’s website. It’s trained on Qwick’s full site map (FAQs, blog content, onboarding docs) and its primary job is to route visitors down the right path and get the right ones to book a meeting.

Step 1: Define Your Routing Logic Before You Touch the Tool

Before you think about which AI model to use, figure out who visits your site and what each persona actually needs. This is the decision that matters most.

For Qwick, there were two distinct audiences: workers looking for shifts, and businesses looking for staff. Building one generic experience for both was a mess. Building two separate agent paths, one per persona, cleaned everything up.

Before you open Voiceflow, map this out on paper:

• Who are the 2-3 types of people that land on your site?

• What does each one need? (Information? Qualification? A meeting?)

• What’s your desired end state for each? (Meeting booked, form submitted, support deflected?)

Example prompt to get yourself started in Claude: “I sell [product] to [ICP]. My website gets traffic from [persona A] and [persona B]. Help me map out what each visitor type needs and what the ideal outcome is for each one.”

Don’t skip this step. Every iteration Qwick went through — and there were several — came back to getting the routing right.

Step 2: Build the Agent in Voiceflow

Once your routing logic is mapped, stand up the agent in Voiceflow.

The architecture is simpler than it sounds: Voiceflow lets you chain agents together. Think of each agent as a node in a conversation flow. The first agent collects context (who are you?), passes that to the next agent (what do you need?), and so on. Each subsequent agent inherits all prior context and adds to it.

Key build decisions:

• Train it on your content. Voiceflow connects to your knowledge base: site map, FAQ docs, blogs. Feed it everything relevant. Avery can answer detailed questions about Qwick’s business because it’s been trained on the full content library.

• Use buttons, not open text fields. Qwick tested both. Users prefer buttons. They’re faster, reduce drop-off, and give you cleaner data. Pre-populate the options you care about (business type, city, etc.) rather than asking visitors to type free-form answers.

• Set a question limit. Avery asks up to 10 qualifying questions, then routes. If a visitor has follow-up questions, it answers up to two more before redirecting to a human. This keeps conversations moving and prevents the agent from becoming a chatbot people abandon.

Step 3: Connect It to Your CRM and Calendar

This is where the lead gen actually happens.

Voiceflow integrates natively with Salesforce, HubSpot, and Google Sheets. Once a business visitor has been loosely qualified, Avery routes them directly to an SDR’s Calendly (or equivalent). No form. No intermediate step. Just: book the meeting.

“We really just want to get a meeting booked — so we don’t even want to go to a form.” — Charles Vandercook, Head of Growth, Qwick

The context collected during the conversation (business type, market, staffing needs) gets passed to the SDR before the call. The meeting is already warm. The SDR isn’t starting from zero.

What to connect:

• CRM (Salesforce or HubSpot) — so every interaction creates or updates a contact record

• Calendar booking link — so the agent can drop a meeting directly on an SDR’s calendar

• Google Sheets — if you want a lightweight backup or want to do batch analysis separately

Step 4: Use the Chat Transcripts to Improve the Agent and Your GTM

This is the part most teams skip, and it’s where a lot of the value lives.

Voiceflow captures every conversation transcript. At Qwick’s volume of 1,400 interactions per month, that’s a significant signal set about what your visitors are actually asking, where they’re dropping off, and what they don’t understand about your product.

Charles runs batch evaluations on the transcripts: deflection rates, satisfaction scores, drop-off points by step. But you can also just pipe a batch of transcripts into Claude and ask:

Prompt: “Here are 20 chat transcripts from our AI agent. What are the top 5 questions visitors are asking that we’re not addressing well? Where in the conversation are they dropping off? What content gaps does this reveal?”

This is direct, unfiltered feedback from your customers, the kind they’d never give you in an NPS survey. Use it.

Step 5: Embed It and Iterate

Deployment is the easy part. Voiceflow outputs a simple HTML snippet. Drop it into your site, and Avery shows up.

The harder part is committing to iteration. Qwick’s first version assumed mostly businesses would interact with Avery. Turns out it was mostly workers. Opening a separate lane for workers changed everything. It improved the business-side conversion rate and gave the operations team a support deflection tool they hadn’t planned for.

Iteration checklist:

• Review transcripts monthly and look for drop-off patterns and unanswered questions

• A/B test button options (Qwick rotates their city list based on which markets are hot)

• Add new agent nodes as your use cases expand (Qwick is building a post-sale version of Avery now)

• Watch your conversion rate. Qwick converts 14% of interactions into a meeting, and that’s a real benchmark to aim for.

What Good Looks Like

A well-built lead gen agent does three things simultaneously:

1. Routes visitors to the right experience without friction

2. Qualifies leads passively, before a human ever touches them

3. Generates a continuous stream of product intelligence from real visitor conversations

At ~$40/month and 14% meeting conversion across 1,400 monthly interactions, Avery is producing roughly 196 warm, pre-qualified meetings per month for Qwick’s SDR team. In practice, that’s the output of a full-time SDR with zero ramp time.

Watch-Outs

Don’t start with open-ended chat. It seems more conversational, but users drop off faster. Buttons win. Start with a constrained, button-driven flow and open it up only when you have a clear reason to.

Don’t try to close in the chat. The agent’s job is routing and qualification, not selling. The second you ask Avery to do too much, it does everything worse. Get the meeting. Let your SDR / AE do the rest.

The Design Work Matters More Than the Build

The barrier to building a lead gen agent is lower than you think — in cost, in code, and in time. What’s actually hard is the design work that happens before you open the tool: mapping your personas, defining your routing logic, and deciding what “good output” looks like.

Get that right, and the build is straightforward. Then let the transcripts teach you. The agent gets better every month because your visitors are telling you exactly how to improve it.

Special thanks to Charles Vandercook, Head of Growth at Qwick, for walking the Stage 2 community through Avery’s build and sharing real metrics from the field.

Subscribe now

DEAR SEEKING EFFICIENCY GAINS: Whether you’re considering building or buying, your instinct is right: you have to figure out where time is going and what work can be replaced *before* investing in the tech stack.

I had a great chat with Shaun Porcar, Stage 2 LP, Advisor and multi-time Head of Customer Success on this exact topic. We agreed we’re seeing a lot of CEOs saying “use AI to create capacity,” then the CS leader starts demoing tools, pilots get started, and a few months later, there’s still no measurable change. Plus, all the overhead that comes with implementing and managing a new tool.

It feels like we’re not spending enough time on diagnosis and diving right into solutioning.

Vendors aren’t lying when they quote results like “30% time savings”. But 30% of what, exactly? If your team is spending most of their time on work that isn’t driving retention or expansion, you don’t have a capacity problem. You have a work allocation problem. And buying a tool to do the wrong work faster doesn’t fix that.

So the real challenge you run into in diagnosing this is that your team’s self-reported data is nearly useless.

Ask your CSMs where time goes, and they’ll all say the same thing: customer calls and activities related to meeting prep or next steps. That’s true, but it also tells you nothing. What kind of calls? At what stage of the customer journey? Are they proactive or reactive? Is the customer getting value from them or tolerating them? You need real data.

Run a time audit

The audit Shaun recommends is simpler than you’d think.

Start by connecting to your team’s Google Workspace or 365 calendars. Classify every meeting and logged activity as external or internal. Then try to map each external meeting and important activity to a stage or key inflection points in the customer journey: onboarding, adoption, renewal, expansion, escalations, stakeholder changes, etc...The goal is to understand where time is spent and what work is actually being done.

As Shaun puts it: “Where is time actually being spent? You can literally build something that will provide you with a dashboard. You’ll start to see where you’re bottlenecking around onboarding, renewals, or something vaguely classified as a check-in call. Something more is happening here that people are spending a lot of time in.”

Start lightweight and don’t let tools be the barrier. A spreadsheet works, or Cowork or Monocle if you want something automated. Shaun’s framing: “You don’t need a lot of fancy tools to do this. You just need something that’s going to help you break down and understand where people are spending time and what they are actually doing with that time.”

Once you see where time concentrates, ask one more question about that category of work. As Shaun frames it: “Is this something that you’re doing to be proactive, is it something that you’re doing to be reactive, or are you actually collaborating with that customer to action a specific task on their end? You can figure out what sort of value is being driven, if any, in that conversation.”

Proactive work is usually worth protecting. Reactive firefighting is almost always a signal of a process gap upstream (incomplete onboarding, unclear success criteria, a handoff that isn’t working). And customer collaboration calls are context-dependent. Sometimes high-leverage, sometimes a sign that self-serve is missing. Going further, sometimes the biggest opportunity isn’t automation at all. It’s eliminating or rethinking that part of the workflow altogether.

While the audit runs, talk to your customers

It’s easy to skip this part, but don’t! Shaun is direct about it: you need to spend real and meaningful time hearing from your customers. Shaun calls out a few specific topics to dive into: “Ask them about their experience with your company. What is the path they followed to see value? How are they thinking about ROI? The amount of time they’re spending with us, should they spend less? What are their preferred communication channels?”

The answer most CS leaders find when they ask customers directly: they actually want less of your time, not more. They want the right answer in the moment they need it, in whatever channel they’re already in.

White glove doesn’t mean a CSM on the phone 3 days from now. It means fast, accurate, in-context help (whether that’s a chat response, a triggered email, or something surfaced in the product).

Would you build it this way today?

One additional question worth asking: if we were designing this process today with AI available from day one, would we design it the same way?

For many companies, the answer isn’t finding ways to automate an existing workflow. It’s redesigning specific parts of the workflow entirely. An onboarding process, customer communication cadence, or escalation path that made sense a few years ago, or even a few quarters ago, may no longer be the highest-leverage way to create customer value.

That doesn’t mean every company should rip and replace existing processes and tools. Mature organizations often benefit from targeted AI improvements first, just as early-stage organizations should look for targeted opportunities to build agentic workflows with a human in the loop. The goal should be thoughtful workflow design, not simply bolting AI onto legacy operating models.

Now you can buy the tool

Once you know where time is bottlenecked and what kind of work is happening there, vendor evaluation becomes straightforward. You’re not buying “AI for CS.” You’re buying a specific solution to a specific problem, and you can hold the vendor accountable for the results in that slice of your workflow. Just as important, this gives you a clear business case and measurable ROI for the investment.

That’s the fastest path to actual leverage. Skipping it feels faster, but almost always lands you in the failed pilot graveyard.

Until next week!

Ready to get started? Grab the Target Account Playbook on GitHub and put it to work right away. Otherwise, read on for how it works.

Most GTM organizations have more data than they know what to do with. Your CRM, your website analytics, an intent data subscription, LinkedIn alerts, news feeds, and your reps’ own observations. Theoretically, all of it tells you something useful. In practice, most of it sits in a tab someone forgot to close.

The challenge is knowing which signals actually matter for your business, and how to act on them in a way that doesn’t sound like every other rep who saw the same funding announcement.

AI doesn’t solve this problem by adding more data. It solves it by helping you think through what you already have, sort it into something actionable, and build targeted plays — fast.

Subscribe now

The Problem

Most teams skip straight to buying intent data or building scoring models before they’ve done the harder foundational work: figuring out which signals they already have access to, what each one actually tells them, and which ones are most impactful. The result is a 100-point composite score built on guesses — false precision that feels rigorous but breaks down the moment you try to explain it to a rep. When the score doesn’t hold up, you can’t untangle what went wrong.

And even when signals exist, surfacing them to reps in a usable way is a whole separate challenge. Slack alerts across every account in a rep’s book get overwhelming fast. Reps stop paying attention, which defeats the purpose.

What AI Makes Possible

AI can help you do the foundational work that most teams skip. Specifically: take stock of what signals you actually have, sort each one by what it tells you, and help you design focused experiments — one signal, one cohort, one message hypothesis at a time. You don’t need a scoring model. You need a signal map and a place to start.

Here’s how to build it.

Step 1: Give AI Your Starting Context

Before AI can help you with anything here, it needs to understand your business. This is the setup that makes everything downstream more useful.

Open a conversation with Claude (or your AI of choice) and give it the following:

The more specific you are here, the better. If you sell to mid-market SaaS companies and your CRM is HubSpot and you have a free trial motion, say that. AI can’t surface relevant signals if it’s working with a generic profile.

Step 2: Map What You Actually Have

Most teams undercount their first-party signals and overspend on third-party intent data. Before you go hunting for new tools, take stock.

A useful organizing framework is to sort your signals by source:

Use this prompt to brainstorm with AI:

This conversation should take 10–15 minutes and end with a list you probably didn’t have written down anywhere. Third-party signals are worth exploring, but don’t start there. Everyone has access to the same job change alerts and funding announcements. Your first-party data is the one thing competitors can’t replicate.

Step 3: Sort Each Signal by What It Actually Tells You

Not all signals mean the same thing, and the action each one suggests should be different. Here’s the sorting framework:

Some signals map to more than one category. A leadership change is both a person fit signal (new person in the role you care about) and a timing signal (new leaders make vendor decisions early in their tenure). That’s fine — what matters is knowing what each signal tells you so you can act on it appropriately.

After your signal inventory from Step 2, prompt AI to run the sort:

The output is your signal map. It doesn’t have to be a formal document — a table in your notes or a shared doc is enough. The point is having a clear picture of what each signal tells you before you try to act on any of them.

Step 4: Build Focused Plays, Not a Composite Score

Here’s where most teams go wrong: they try to assign weights to everything and build a single prioritization score. It looks scientific. It gives you a neat ranked list. The problem is you’re assigning weights before you know which signals actually predict outcomes for your business.

A score also doesn’t give a rep what they actually need: context. Why should I prioritize this account? How should I engage? A number doesn’t answer either question. Context does.

Instead, pick one or two signals you have real conviction in — ideally ones that are unique to your business or at least not signals everyone else is acting on — and run focused plays around them.

For each play, you need four things:

Prompt AI to help you design the play:

Run the play. Measure what happens. Then use AI to help you interpret results:

This is how you build conviction in specific signals over time — not by buying a scoring tool, but by running experiments and measuring outcomes.

Step 5: Get Signals in Front of Your Reps in a Usable Way

Once you know which signals matter, the question is delivery. How do reps actually see and act on them?

The instinct is to build real-time alerts and integrations. Resist it, at least at first. Real-time Slack alerts for news signals across every account in a rep’s territory get overwhelming fast. Reps stop paying attention, and you’ve built a notification system no one uses.

A better starting point: a weekly digest. One message or report that answers “which accounts in my territory showed activity this week and what did they do.” That’s enough context for a rep to go into Monday morning with a prioritized list.

Prompt AI to help you design it:

As your plays prove out, you can invest in tighter integrations — signals surfaced directly in your CRM, enrichment workflows in Clay, Slack alerts for your highest-conviction triggers only. But the digest is how you validate that reps actually find the signal useful before you build infrastructure around it.

What Good Looks Like

A well-run signal map and play system doesn’t look like a dashboard. It looks like a rep who can answer: “I’m prioritizing these five accounts this week because of X, and here’s the angle I’m leading with.”

The signal map gives you the inventory. The plays give you the experiments. The measurement gives you the feedback loop. And AI compresses the time it takes to go from “I have a lot of data” to “I know which accounts matter and why, and I have a specific way to engage.”

The teams that do this well aren’t using more signals than everyone else. They’re using fewer, better ones — and they’ve done the work to understand why each one matters for their specific buyers and motion.

Watch-Outs

• Timing signals decay fast. A company switching away from a competitor creates a window. A hiring spike suggests investment now. These signals have a shelf life of days or weeks, not months. Build your plays with urgency in mind, or the signal will have passed before you act on it.

• Don’t skip the reasoning layer. A data point is not a signal. A signal is a data point paired with a clear reason why it matters for your business and what you should do when you see it. If you can’t articulate the reasoning, you’re collecting noise, not signals. Use AI to pressure-test your logic before you build a play around it.

Tools Worth Exploring

The specific tools matter less than having a clear plan for what each signal tells you and what action it should trigger. That said, here’s where to look by signal type:

• Website and engagement tracking: Factors.ai, Clearbit Reveal

• Intent and news signals: Clay, Exa

• Job changes and hiring patterns: LinkedIn Sales Navigator, Clay enrichment

• Competitive signals: G2, TrustRadius

• Signal orchestration and plays: Clay, your CRM’s sequence tools

The Takeaway

The signal problem isn’t that you don’t have enough data - you just haven’t built the reasoning layer on top of it yet. What makes a data point a signal is the judgment about why it matters and what to do when you see it.

AI speeds up that foundational work dramatically. Instead of spending weeks building a scoring model, you can spend an afternoon mapping your signals, sorting them by what they tell you, and designing your first play. Then you measure, learn, and sharpen.

Special thanks to Brian Aoyama and Mallory Blocker from the MedScout team for sharing their workflow

Subscribe now

DEAR SELLING TO FINANCE: You’re in lock step with your champion. Your pricing is fair. And you’re still losing at the finish line. Turns out we are all selling to finance now and what’s happening at the CFO stage is usually more about proof than pricing. The CFO is (rightfully) skeptical of your math (think of the time savings and revenue if all of our ROI calculator promises actually came to be!!). So when the math can’t be verified, the deal collapses into a negotiation about discounts instead of a conversation about the value.

This week I chatted with Kenny Scannell, Stage 2 LP and CRO at Otter.ai who has spent 20+ years scaling revenue teams at companies like Rocketlane, Zoom, and Klaviyo and he shared a real life example of building out a business case and evolving it from a google sheet into a live product using Lovable.

The mistake most sales teams make is treating discovery and business-case creation as separate activities. The strongest business cases are built during discovery itself. Every question becomes an input. Every workflow becomes evidence. By the time you present to the CFO, you are not introducing a new narrative. You are organizing information the prospect already helped create.

Build the Case From the Prospect’s Own Data

Most business cases are built on case studies. “Company X got 10x ROI. You will too.” As Kenny puts it: ‘Every time you get to a CFO, they call BS.’ The issue is not that CFOs dislike ROI. It is that generic vendor assumptions rarely hold up. Once the CFO decides the numbers are not grounded in their business, the conversation shifts from value to price, and the seller loses credibility.

At Rocketlane, Kenny’s team identified roughly 35 common activities performed by implementation teams inside their existing PSA or project management platform. For each activity, they captured the current time required, then compared it to the same workflow inside Rocketlane.”

The data didn’t come from case studies. It came from the prospect. As Kenny describes it: “We’d actually go through an interview process with an IC and a manager.” Running these interviews in parallel to the deal is the key move. Discovery isn’t just generating rapport or qualifying the deal. It’s generating the inputs for the business case at the same time. By the time you get to the CFO, you’re not presenting your assumptions. You’re presenting their data back to them.

The math becomes verifiable. “It’s as simple as saying, hey, to migrate this template and do this and that, it takes you five minutes today on this system. And then we start to compare and contrast. Within Rocketlane, it takes you one minute because we have AI. The inputs from Rocketlane on how long these tasks take are all provable through demos.”

A CFO can push back on a benchmark from a vendor’s case study. It’s a lot harder to push back on a time estimate from your own IC.

What It Looks Like When It Works

Kenny’s team started with a Google Sheet. Eventually they built a business case creator in Lovable so prospects could fill it out themselves and get a real-time ROI score. The Lovable version made the case more interactive. Instead of a rep manually maintaining a spreadsheet, prospects could enter their own operating assumptions, see the calculations transparently, and adjust variables in real time. The result: win rate went from 18% to 34% over four quarters.

Now at Otter.ai, they’ve taken it a step further: discovery transcripts become the raw material for the business case. The team connects Otter’s MCP server to Claude, pulls the relevant discovery call context, combines it with a Google Drive business-case template, and generates a first draft in about two minutes. Rather than starting from a blank page, reps start with a case populated from the prospect’s own words.

In fact they’ve now created 2 swim lanes - one Lovable product that can be used with the prospect to gather the interview feedback noted above and one to create the finished product.

The rep starts here:

Works through a series of inputs with their prospect: The output of the workflow looks less like a static ROI calculator and more like an operational assessment. The prospect can see the inputs: team size, meetings per person, average meeting length, manual note-taking, follow-up work, manager review time, and hours that can be reallocated. The model then translates those inputs into reclaimed hours, reduced admin time, more customer-facing capacity, and projected business impact. That matters because the CFO can inspect the logic. The case is not “trust our 10x benchmark.” It is “here is how your own workflow changes, task by task.”:

A Word of Wisdom

AI can assemble the case, but it’s only working with the data you provide. This business case is ultimately only as good as your discovery. As Kenny says: “You have to run a great discovery to actually do that. So if we’re missing information, we have to double down. But it’s a good start.”

And on big deals, the AI output is a starting point, not a final answer. “For larger strategic opportunities, I’m still reviewing the business case personally. AI accelerates the process, but judgment still matters. I’ve seen ROI estimates come in both too conservative and too aggressive. The quality of the output depends on the questions we asked and how well we quantified the answers.”

The team still needs to ask the right questions in discovery. Reps still need to know their product’s time-savings well enough to fill in accurate inputs. And someone needs to sanity-check the output before it goes to the CFO on a big deal.

CFOs are looking for proof. Give them math they can verify using their own data and the conversation changes. Go log in to Lovable and try this out for yourself.

Until next week!

Sourcing contacts inside an account is where most GTM teams waste a lot of time. And it usually starts with a data filter. "VP of Sales" at a 50-person company and "VP of Sales" at a 5,000-person company are different jobs. Same title, different universe. But most teams run the same filter on both, load reps with contacts that were never going to convert, and then wonder why they aren’t seeing results.

Here’s what’s actually happening: you’re outsourcing judgment to a database. And databases don’t know what good looks like at your company - you do.

Ready to get started? Grab the Contact Qualification & Enrichment skill on GitHub and put it to work right away. Otherwise, read on for how it works.

The Problem

Most teams treat contact sourcing like a data pull. They set some title and seniority filters in Sales Nav or ZoomInfo, export the list, push it to reps. Reps then spend hours clicking through profiles trying to figure out who’s actually worth reaching out to. There’s no documented logic, no consistency across the team, and the criteria live in one senior person’s head. When that person leaves or scales past the point of doing it themselves, the judgment goes with them.

What AI Makes Possible

You can extract that judgment, document it, and run it programmatically — so every contact in every account gets evaluated against the same criteria, at scale, without burning rep time.

Here’s how to do it.

Step 1: Validate Your Contact Theory Against Reality

Before you go evaluate individual people, pressure-test your assumptions about who you’re actually selling to.

Most teams have a working theory about which personas matter — job functions, seniority levels, types of roles. Write that down. Then export two things from your CRM: contacts who came inbound in the last 6–12 months, and contacts associated with closed deals. Upload both to Claude or ChatGPT and ask it to synthesize the patterns.

Example prompt:

“Here are two CSV exports: recent inbound contacts and contacts linked to closed deals. Identify which job functions, titles, and seniority levels appear most frequently. Flag any titles or functions that appear often in closed deals but not in our working persona list.”

You’re looking for surprises. Roles you didn’t expect. Functions that keep showing up in your deals that aren’t in your target list. This gives you a validated filter set — grounded in what actually worked, not just what sounds right.

Step 2: Do a Manual Review with Dictation Running

Pull contact lists from a handful of accounts — mix of segments and verticals if you have them — and have your most experienced person go through each profile on LinkedIn, out loud.

The key is capturing the why in real time. Use a dictation tool (Wispr Flow and Monologue both work well) and have them narrate as they click: what catches their eye, what makes them hesitate, who they’d reach out to first and why. Don’t filter it — you want the raw reasoning, not a polished summary.

To get the contacts: build a search in Sales Navigator using the filters from Step 1, then export the leads using a browser automation or export tool — PhantomBuster, Evaboot, and Captain Data all handle this well.

What you’re listening for:

• Language on the profile that signals forward-thinking vs. entrenched — are they describing outcomes or tasks?

• Seniority and growth trajectory — have they moved up within a company over time?

• Whether the title reflects real ownership or a narrow individual contributor role

• How company size changes what the title actually means

That last one matters more than most teams account for. A “Director of Sales” at 50 people probably owns the entire function. Same title at 2,000 people might manage one region. You may end up needing separate criteria by segment.

Step 3: Synthesize Patterns with AI

After the review, upload your notes (or the dictation transcript) to Claude and ask it to synthesize what you found.

Example prompt:

“Here are my notes from reviewing [X] contact profiles. Identify the consistent signals that made someone a ‘yes’ vs. a ‘no.’ Group them by theme — seniority signals, language patterns, role ownership, etc. Then flag any patterns that were specific to a company size or segment.”

You’re not asking AI to invent your criteria. You’re asking it to find the shape of something you already know. The output should read like a documented version of what your best reviewer has been doing in their head.

Review the synthesis carefully. Push back where it oversimplified. Add nuance where it missed it. What you end up with is a qualification framework — your actual criteria, written down, ready to be turned into a prompt.

Step 4: Build Your Classification and Qualification Agents — Separately

Here’s something learned the hard way at MedScout: don’t try to classify and qualify in one step. When you combine them, the assessment gets muddy — you end up trying to figure out what someone does and whether they’re relevant at the same time, and both answers suffer.

Split it:

Classify first

Build a prompt that maps each contact to a job function based on their LinkedIn profile. What bucket do they actually belong in — sales, ops, marketing, finance, etc.? The same title can mean different things at different companies. Getting the function right makes the fit judgment cleaner.

Example prompt:

“Based on this LinkedIn profile, classify this person’s primary job function into one of the following categories: [your function list]. Use their title, current responsibilities description, and career history. If ambiguous, default to the function most represented in their recent roles.”

Qualify second

Now that you know what they do, assess fit. Are they the kind of person who would drive or shape an evaluation of a tool like yours? The answer is a holistic judgment — seniority, scope, function ownership, how they describe their work.

Structure the output in tiers that map to action, not just quality:

• Best Fit: worth prioritizing in outreach — they’d drive or heavily influence the evaluation

• Good Fit: worth having in the CRM and engaging, but not the person you lead with

• Not a Fit: not worth the outreach

If your segments are meaningfully different (say, SMB vs. enterprise), build separate qualification prompts for each. The buying dynamics are different enough that one set of criteria won’t cover both cleanly.

Step 5: Operationalize the workflow

Once your prompts are tested and trusted, load them into a workflow automation platform. Clay, Apollo, and Cargo are all built for this kind of multi-step enrichment and qualification pipeline.

The workflow: build your contact list in your contact database → export the leads → enrich with LinkedIn profile data → run your classification prompt → run your qualification prompt → push qualified contacts to your CRM with source and date tagged.

A few things to get right on the back end:

• Deduplicate against existing CRM records before pushing new contacts

• Tag enrichment source and date on every record — you’ll thank yourself when you need to audit data quality later

• Don’t enrich every contact at every account. Enrich the people who matter at accounts that are already qualified. Sequencing matters: account qualification first, then contact qualification, then enrichment

The goal is flipping the work. Instead of each rep spending hours in Sales Nav trying to figure out who to reach out to, the system surfaces the right people with enough context for reps to react and make the final call. That time savings compounds fast.

What Good Looks Like

A well-built contact qualification agent returns something like this for each contact:

Function classification: RevOps / Sales Operations

Fit tier: Best Fit

Reasoning: Director-level with explicit ownership language — describes “building the forecasting process from scratch” and “owning the tech stack.” Company is 300 people, so this title reflects real scope. Career history shows consistent growth within the function.

Suggested action: Prioritize in outreach. Likely economic or technical evaluator.

That output is what you want reps reacting to — not raw Sales Nav profiles.

Watch-Outs

Don’t skip the manual review phase. It’s tempting to jump straight to building prompts. But your AI agent is only as good as the criteria you give it — and those criteria have to come from real judgment applied to real profiles. The manual step isn’t inefficient. It’s the input.

Segment earlier than you think you need to. If your ICP spans SMB and enterprise, don’t try to use one prompt for both. The signals that make someone a “Best Fit” at a 50-person company look different than at a 5,000-person company. Build for that early, or you’ll rebuild later.

Tools Referenced in This Guide

This guide focuses on the methodology — how to extract, document, and operationalize your team’s contact qualification judgment. The specific tools are interchangeable. Here are the categories and options mentioned:

- Contact database (search and filter potential contacts by title, function, seniority, company): LinkedIn Sales Navigator, ZoomInfo, Apollo

- Contact export (pull contact lists from your database into a workable format): PhantomBuster, Evaboot, Captain Data

- Enrichment & workflow (enrich profiles with additional data, run AI prompts at scale, manage the pipeline): Clay, Apollo, Cargo

- Dictation (capture spoken reasoning during manual review): Wispr Flow, Monologue

- AI assistant (synthesize patterns, build classification and qualification prompts): Claude, ChatGPT

Pick what works for your stack. The frameworks in this guide work across all of them.

Special thanks to Brian Aoyama and Mallory Blocker from the MedScout team for sharing their contact qualification workflow, which shaped several of the frameworks in this guide.

Subscribe now

Subscribe now

DEAR OVEREXTENDED MANAGER: If you missed it, here’s the article on how AEs should using AI in live deals. To be clear, there are lot of great solutions out there, but I was trying to share some actionable advice that an individual could implement (aka not company-wide, no business case or new budget). Addressing the same question at the manager level, it comes down to coaching. Here are 3 ideas to get you started with any mainstream LLM:

1. Pattern Recognition Across Calls

Instead of reviewing one call at a time, upload 5–10 transcripts and ask:

• What patterns show up in lost deals?

• Where is discovery shallow?

• Are reps multi-threading effectively?

• What objections are recurring?

Most managers rely on anecdotal memory when AI can surface trends quickly. You still validate the conclusions, but you start from signal instead of guesswork.

→ Key Takeaway #1: Use AI to find patterns. Use your experience to interpret them.

2. Create Personalized Coaching Plans

After reviewing a rep’s recent calls, prompt:

• What 3 skill gaps are most evident?

• What drills or role-play scenarios would strengthen those areas?

• What metrics should I track over the next 30 days?

Instead of vague feedback (“ask better questions”), you can walk into a 1:1 with:

• Specific moments from calls

• A clear development focus

• A measurable improvement plan

That changes the quality of the conversation and allows you to scale yourself as a manager. One thing worth knowing: managers tend to lean into the back-and-forth style v. being told what to do. That’s exactly why coaching-plan workflows like this work best!

→ Key Takeaway #2: Coaching should feel hyper-customized.

3. Upgrade Your 1:1 Prep

Before your weekly 1:1:

• Input CRM notes and pipeline data (connect the data sources to make this easier if you can!).

• Ask for a summary of leading indicators.

• Identify deals at risk based on stage aging

This gives you sharper questions:

• Why has this deal stalled?

• What’s the next customer-committed action?

• Where are we single-threaded?

AI sharpens your agenda and saves you meaningful time on admin work. You can even take this a step further and start automating these workflows to land in your inbox morning of a 1:1 or summaries at the end/start of a week.

→Key Takeaway #3: Managers spend less time on administrative prep.

Until next week!

Most investor updates get skimmed for two minutes and filed away. A few get saved, forwarded, and referenced in board meetings. The difference isn’t the quality of your quarter — it’s the quality of your communication.

Here’s how to use AI to write a monthly update that builds investor trust even when the numbers are mixed.

Download the Skill on GitHub

The Problem

Founders treat investor updates like a chore and often miss including information investors need to quickly assess your situation and provide support.

Often founders write paragraphs of information without including key metrics that help round out the narrative. The data-dump version makes investors do the interpretation work themselves, which means they fill in the story with whatever assumptions they have.

What you actually want: lead with the number, add the context that matters, skip everything else.

What AI Makes Possible

AI won’t make your quarter better. But it will help you structure what you know into an update that gives your investors the right context and builds credibility over time, even when the news is mixed.

Specifically, the founder-monthly-update skill in Claude structures every update the same way, so your investors always know where to look for what they care about. It also pushes you to include the things founders tend to skip: upfront summary, the pipeline coverage ratio, the real reason the logo churned.

Step 1: Gather Your Raw Numbers First

Before you open Claude, pull everything into one place. You need:

• Revenue: Current ARR, new ARR this month, expansion, contraction, churn, as well as the plan

• Cash: Bank balance, last month burn, 3-month average burn

• GTM detail: New logos + ACV, expansions, renewals, churn with reason

• Pipeline: Current quarter target, pipeline coverage ratio, any named deals worth flagging

• Product: What shipped, what’s coming next month

• Team: Headcount, new hires, departures, open roles

You can either pull these numbers together or connect Claude to a spreadsheet and/or your systems. Don’t editorialize yet. Just get the raw inputs together. The skill will help you turn them into a narrative.

Step 2: Give Claude Your Data and Ask for the Update

Paste your raw numbers into Claude and say:

“Write my monthly investor update using the founder-monthly-update skill. Here’s my data: [paste your inputs]”

The skill will produce a complete email draft with this structure, every time:

• Subject line — metric-anchored, built to get opened

• TL;DR — 3–5 sentence narrative that hits every section at a glance

• Key Revenue Metrics — clean data block including goal and % to goal

• Cash — balance, burn, runway, flagged if elevated

• GTM Update — logos, expansions, churn with reasons, pipeline coverage

• Product — what shipped and what’s coming

• Team & Culture — hires, departures, open roles, leadership updates

The % to goal calculation matters. Investors track it. The skill always shows both your actual number and your goal — not just one or the other.

Step 3: Read the TL;DR Out Loud

Before you send anything, read the TL;DR paragraph out loud. It should answer these five questions in 3–5 sentences:

• What was your biggest win (or miss) this month?

• What’s the key product or strategic move?

• What’s the team headline?

• What’s the one honest flag — burn, pipeline, a risk?

• What’s the energy on the team right now?

If you can’t answer all five from the TL;DR alone, ask Claude to tighten it. The TL;DR does the heavy lifting. Everything below it is for investors who want the detail.

Step 4: Check the Flags

The skill will prompt you to include things that are going wrong — not because investors enjoy bad news, but because founders who only share wins lose credibility over time.

Before you finalize, verify you’ve named:

• Burn if it’s elevated (and why)

• Pipeline coverage if it’s tight (and what you’re doing about it)

• Any churn, with the actual reason — not “strategic misalignment,” the real reason

• Runway if it’s under 18 months

If all of these are genuinely positive, great — say so. But don’t skip the section because it feels uncomfortable.

What Good Looks Like

A strong investor update from a founder who closed 103% of their Q1 target might open like this:

We closed Q1 at 103% of our new business target ($X against $Y), added 5 new logos including X and Y, and made the decision to shift our entire product org toward building an <insert detail here>.. March burn was elevated at $X, and Q2 pipeline coverage is tight at ~1x against a $X target — both areas we’re focused on. The team is ready to climb a new mountain.

That’s five things in four sentences. The win, the logos, the strategic move, two honest flags, and closing sentiment. Investors can file that away and know exactly what the company stands for that month.

Investor Update Mistakes to Watch Out For

Don’t let AI write the strategic narrative.

The skill structures everything, but the company strategy section — why you’re making a big bet, what the market is telling you — has to be in your own words. Investors can tell when a founder didn’t write their own conviction.

Don’t use % to goal without showing the goal.

“We hit 87% of plan” is meaningless without knowing the plan. The skill enforces this, but double-check the metrics block before you send.

Don’t skip the Customer Love section.

One real customer quote or anecdote — the VP who told you your product changed how their team works — does more for investor confidence than three more metrics. Include it when you have it.

The Takeaway

The goal of a monthly investor update isn’t to impress your investors. It’s to keep them calibrated. Investors who are well-calibrated ask better questions, make better introductions, and better understand how to support you.

The founder-monthly-update skill gives you the structure to do that consistently — so your updates become an asset over time, not just a monthly obligation.

Download the Skill on GitHub

Paste your raw metrics into Claude and say: “Write my monthly investor update using the founder-monthly-update skill.”

Thanks for reading GTM in Practice with Stage 2 Capital! Subscribe to receive new posts to your inbox weekly.

Subscribe now

DEAR HACKATHON TO EXECUTION GAP: Running a hackathon is a great idea in theory, but turning the winning ideas into something that changes how your team works is where things break down.

We recently brought together a group of Stage 2 LPs, including Jeb Dasteel, former Chief Customer Officer at Oracle, David Hwang, CCO and Interim CRO at Grammarly, and Colin Murphy, CCO at Zendesk to dig into this exact challenge. Below is a summary of their collective takeaways. Their read: these projects fail because they weren’t designed or scoped to survive governance, resourcing, and cross-functional realities.

If you want your hackathon to produce something that actually ships, you have to build the operating model before you build the prototype.

Start with the business outcome, not the experiment

A common pattern: teams come out of a hackathon with great ideas, then realize they don’t have the business units they need (IT, data science, security) to actually unlock them. The energy dissipates, and the idea dies in a backlog somewhere. Getting those stakeholders into the conversation early, before the hackathon, not after, is what separates the experiments that ship from the ones that don’t.

Match oversight to risk

Not every AI initiative requires the same level of governance. Internal workflow automation moves much faster because the exposure is lower. Once automation touches customers directly, the oversight model changes.

Separate internal productivity use cases from customer-facing automation. Internal use cases often include knowledge retrieval, drafting communications, or demo preparation. These create quick wins and build momentum. Customer-facing automation requires deeper collaboration with security and data teams.

If governance enters the conversation after the hackathon, it feels like friction. If governance is built into the evaluation criteria from the start, it becomes part of the design.

Pressure test your foundation before you automate it

Many teams assume their documentation is ready, then discover it isn’t updated, isn’t structured, and is full of assumptions that were never written down.

Preparing for AI-powered onboarding automation often means going back and designing a prescriptive experience from scratch. That means defining the journey, creating the content, and filling in the data gaps before any automation is possible.

Many teams discover that the real work is not building the AI layer. It is cleaning up what sits underneath it. If that’s the case, take a beat to do this foundational work first.

Plan for maintenance, not just momentum

The build versus buy discussion is different today than it was even a year ago. Tools evolve quickly, and internal teams are more capable than ever. But the real question is not whether you can build something. The question is whether you are prepared to support it long term.

After the hackathon, assign a clear owner and ask the hard resourcing questions:

• Who is going to maintain this?

• Does your team have the capacity to keep up with it as the technology keeps moving?

This was the make-or-break factor that came up again and again - not whether you can get something built, but whether you have the people to sustain it. If you don’t have a clear answer, that’s worth resolving before you commit.

Until next week!

Most discovery call debriefs go like this: the AE gives a verbal summary or fills in a few CRM fields, sends a follow-up email, and the deal moves to the next stage with half the MEDDIC blank.

Subscribe now

The Problem

MEDDIC works as a framework. What breaks down is the debrief. Managers don’t have time to listen to every call. AEs self-report optimistically. And even when someone does listen, translating a 45-minute transcript into a crisp qualification gap analysis takes real cognitive work.

The result: deals enter the pipeline underqualified, forecasts are wrong, and reps spend cycles on accounts that aren’t closing.

What AI Makes Possible

Drop a call transcript into Claude with a structured MEDDIC debrief skill, and you get a scored qualification analysis, a sales leader commentary, a prioritized next steps list, and a ready-to-send mutual action plan email — in under two minutes.

This isn’t AI summarizing the call. It’s AI doing the analytical work a good sales manager would do: reading between the lines, flagging what wasn’t said, identifying the moment the seller should have pushed harder, and turning that into something the AE can act on immediately.

🛠️ New: Stage 2 GTM AI Skills on GitHub

We’ve launched a Stage 2 GitHub repo containing all of our GTM AI Skills — including the MEDDIC debrief skill referenced in this post — so you can fast-track these how-tos for your own team. Visit our GitHub to download the skill and plug it into your Claude workflow.

Get the MEDDIC Analyzer Skill on GitHub

The How To

Step 1: Get your transcript

Most sales teams record calls via Gong, Chorus, Fathom, or Fireflies. Export the transcript as plain text — you don’t need the recording, just the words. If you’re working from notes, that works too, but the more specific the input, the more specific the analysis.

If you have Claude connected to your recorded calls automatically, you can just ask it to run the MEDDIC analysis against your calls during a specific time and it should pull them automatically.

What to watch out for: Transcripts with heavy small talk at the top (sports, tech issues, catching up) can dilute the signal. However, this skill will skip past preamble automatically and look for the moment the substantive conversation starts.

Step 2: Run the MEDDIC debrief skill

Paste the transcript into Claude and trigger the skill. The skill will:

• Parse the transcript for deal context: stage, competitive situation, timeline, key stakeholders

• Score each MEDDIC element (Strong / Partial / Missing) with specific evidence from the call

• Render a visual scorecard so the gaps are obvious at a glance

• Write a sales leader commentary identifying the critical misses and broader patterns

• Generate 4–6 prioritized next steps ordered by urgency and deal impact

The prompt is simple:

Analyze this discovery call against MEDDIC and give me my next steps. [paste transcript]

Step 3: Review the scorecard and commentary

The scorecard is your quick read — any red (Missing) element is a live deal risk. But the real value is in the commentary. A good debrief will tell you:

• The specific moment the seller should have pushed and didn’t

• Whether your ‘champion’ is actually a champion or just a friendly coach

• How the competitive situation actually looks vs. how the AE described it

• The pattern in how this AE runs calls — not just this deal

Real example: On a recent portfolio company call, the AI flagged that the rep let the buyer name the actual decision-maker and then never asked how to get to him. The call ended with ‘let’s reconnect in May’ — which isn’t a next step, it’s a prayer. That’s the kind of miss that’s obvious in hindsight but easy to miss in the moment.

Step 4: Send the mutual action plan email

The skill generates a ready-to-send MAP email for the AE to send to their buyer contact within 24 hours. The email:

• Opens with something specific the buyer said — not generic ‘great talking to you’

• Summarizes what was covered from the buyer’s perspective, not the seller’s

• Lists what the seller is sending first, before making any asks

• Includes a clean action table with owner and target date for each next step

• Closes with a frictionless CTA — easy for the buyer to confirm

This isn’t just a nice-to-have. A MAP email forces the AE to commit to specific deliverables, gives the buyer something to react to (which flushes out ghosting early), and creates a written record of what was agreed. Deals with documented next steps close faster.

Quick Reference: What to Surface in Each Element

What Good Looks Like

A strong AI-assisted MEDDIC debrief will be uncomfortable to read. It should name things the AE didn’t want to flag — the champion who hasn’t committed to anything, the budget cycle that probably rules out this fiscal year, the ‘free competitor’ objection that was addressed technically but never economically neutralized.

It’s a great opportunity for the AE to continue to learn the art of discovery without impacting their current pipeline.

Watch-Outs

• Garbage in, garbage out: If the AE’s notes are thin or the transcript is heavily garbled, the analysis will be thin too. This skill rewards teams who record and transcribe calls.

• Don’t skip the review: AI will flag gaps based on what was said. It can’t know what the AE knows from offline conversations. Always sanity-check the debrief against the full relationship context.

• The MAP email needs a human read before sending: The skill will draft commitments based on what was discussed on the call. Make sure the AE isn’t promising something they didn’t actually agree to send.

The Takeaway

MEDDIC doesn’t fail because reps don’t know the framework. It fails because the debrief process doesn’t give it teeth — there’s no one with the time and the transcript access to do the analysis consistently, at scale, after every call.

AI closes that gap. Not by replacing sales judgment, but by doing the analytical legwork so the manager can spend their time on coaching, not call review. The sellers who adopt this workflow will have better-qualified pipelines, fewer stalled deals, and follow-up emails that actually get responses.

Subscribe now

DEAR CAUTIOUS CRO: AI doesn’t change the quality of your sales team. It amplifies it. Thoughtful reps get sharper. Average reps get faster at producing average work.

So the question isn’t “Should they use it?” It’s really “Where does it improve thinking inside a live deal?”

In working with our portfolio, the trend I notice: AI is most powerful when it strengthens preparation and inspection, not when it replaces judgment.

NOTE: I’m not going to comment on specific LLMs or products for the purposes of this article. This advice can just as easily be applied to Gemini, Claude, OpenAI, etc…

Let’s dig in…

Before Discovery: Raise the Quality of Preparation

Strong reps don’t walk into discovery hoping the buyer tells them what’s wrong. They walk in with a strong hypothesis based on research.

AI can accelerate the formation of that hypothesis. It can help synthesize likely business priorities, surface industry pressures, and suggest angles the rep may not have considered. That’s useful, especially for newer AEs or when the team is feeling time-constrained.

But you have to set the standard that the rep refines the thinking before it ever reaches the buyer.

If they can’t articulate why they’re asking a question, it shouldn’t be asked. AI-generated discovery questions are raw material and the rep’s job is still to connect the dots.

The real win here isn’t speed. It’s depth of preparation in less time.

Between Stages: Improve Deal Inspection

This is where AI can create real leverage.

You’ve seen deals stall in the pipeline or surprisingly go silent after weeks of engagement. Deals collapse when risk goes unnoticed. Risk comes in many forms - weak access to power, unclear decision process, lack of urgency, etc...

After a call, a rep can input their notes and ask for a structured critique. What buying gaps exist? Where is the economic buyer missing? What objections are likely to appear late? Here’s a sample prompt to get you started:

System Role: You are a senior B2B sales manager with deep expertise in enterprise deal qualification and pipeline risk assessment. You think like a skeptical sales leader preparing for a forecast call. [Add any industry or company specific insight]

Instructions: Review the call notes or transcript below and provide a structured deal inspection critique. Be direct and do not soften findings. Your job is to identify risk, not validate optimism.

Evaluate the following:

1. Access to Power - Is the economic buyer identified, engaged, and on record supporting this deal? If not, what is the path and timeline to get there?

2. Decision Process - Does the rep have a clear, confirmed understanding of how this company buys? Who is involved, what are the steps, and what could derail it?

3. Compelling Event / Urgency - Is there a real, documented reason this deal closes in the stated timeframe? Would this deal slip if the rep stopped pushing?

4. Competition and Risk - What competitive threats or internal alternatives exist? What has not been asked or answered that a competitor could exploit?

5. Buying Gaps - What objections have not surfaced yet but are likely to appear late? What does the rep not know that they should know by now?

6. Recommended Actions - List 3 to 5 specific next steps the rep should take before the next pipeline review to reduce deal risk.

End with a one-line overall risk rating: Low / Medium / High and a single sentence explaining why.

Call Notes / Transcript: [INSERT HERE]

That exercise forces clarity and introspection, and the more data you provide the better the output will be.

To be clear, this doesn’t replace your pipeline review, but can make your pipeline review sharper. Reps who consistently pressure-test their own deals show up differently in forecast conversations because they are actively thinking about risk mitigation.

Late Stage: Elevate the Champion

In enterprise sales, your champion is selling internally when you’re not in the room. We’ve been arming them with material for years and AI can help here too.

AI can absolutely help draft an executive summary, outline a business case, or structure a mutual action plan –as long as you read it, edit it and treat it this as a draft. That’s an appropriate use. It saves time on format so the rep can focus on substance.

Bonus points if you roll this out at a team or org-wide level. You can build a skill or agent trained on the best business cases (or exec summaries, or action plans, etc…) that your org has used over the years. This creates consistency for the team and removes duplicate work.

I’ll say it again…no late-stage communication should be sent without thoughtful editing. Tone, sequencing, and nuance matter more at the executive level. The rep always owns the final message.

Set the Standard Before the Tool Sets It

If you simply tell your team to “use AI,” quality will vary wildly. Instead, define what good usage looks like. For the examples shared today, you can boil it down to:

• AI drafts. The rep rewrites.

• AI surfaces risk. The rep decides action.

• AI suggests ideas. The rep owns judgment.

AI will widen the gap between thoughtful sellers and transactional ones. The reps who win will use it to prepare sharper hypotheses, identify risk earlier, and strengthen their champions, not to send faster templated emails.

As a leader, it’s on you to define what “great” looks like when your team uses AI.

Until next week!

Subscribe now

Most target account lists are built the same way: export from ZoomInfo, filter by firmographics, assign out to your reps. The list looks defensible, but it can be missing the most useful signals that match your ICP.

The problem isn’t the data provider. It’s that the criteria that actually determine fit — the stuff your best reps have been quietly applying for years — never makes it into the list at all. It lives in someone’s head. Filters can’t capture it. And until you write it down, you can’t train to it, scale it, or ask AI to help with it.

Here’s a process for building an account list that captures real judgment, not just spreadsheet logic.

Step 1: Acknowledge what your filters can’t do

Start by pulling your initial list from ZoomInfo, Sales Nav, or whatever data provider you use. Filter by the firmographics you know matter — size, industry, geography. This gives you a working universe.

Then sit with it honestly. There are companies on that list that don’t belong, and companies missing that should be there. That’s not a failure of the tool — it’s the ceiling of filters. Industry classifications don’t map neatly to your ICP. Revenue ranges don’t tell you anything about buying behavior. The raw list is a starting point, not a judgment.

This skill walks you through the entire process end to end. If you want to test each section first, we have included prompts to try along the way.

Prompt to try in Claude: “Here is my current ICP criteria: [paste your firmographic filters]. What are the types of companies that might pass these filters but still be a bad fit? What types of companies might not pass these filters but still be worth targeting? Explain your reasoning for each.”

This is a useful gut-check. It starts surfacing the criteria you’re already applying informally — the ones that never made it into your filter settings.

Step 2: Shadow your best rep(s) — and document what you see

This is the most important step, and the most skipped.

The team at MedScout figured this out early. Brian Aoyama and Mallory Blocker brought their entire team into a room, grabbed a few accounts from the CRM, hooked one of their best AEs up to the projector, hit record on Fathom, and asked him to walk through account evaluation out loud in real time. The team watched what tabs he opened, noticed where he lingered, and paid attention to what made him say “this is interesting” versus “this one’s not for us.”

They wrote up what they observed. Then they took it back to him and asked: what did we miss? The first pass always misses things that were instinctive. That conversation surfaced criteria he’d been applying for years without consciously thinking about them.

Once you have a recording of a session like this, feed the transcript into Claude:

Prompt to try: “Here’s a transcript of a rep walking through account evaluation. Pull out the criteria they applied, the signals they looked for, and anything they weighted heavily without naming explicitly. Flag any areas of uncertainty or unresolved questions. Structure it so we can bring it back to the rep to confirm and fill in what’s missing.”

The output of this step is a first documented draft of how your best person actually evaluates accounts — not invented top-down, but captured from real behavior. That’s worth building for two reasons: training new reps, and eventually serving as the foundation for AI to help qualify accounts at scale.

Step 3: Run your best and worst customers through the same lens

Pull at least three of your best-fit customers and three of your worst-fit or churned customers. For each one, ask: what made this a good or bad fit? What did we know — or could have known — before the deal closed?

You’re looking for the underlying traits beneath the surface patterns. Not just “they were a small company” but why that mattered. Was it budget constraints? Decision-making process? Risk tolerance? “Small company” is a data point. “Low risk appetite in a category they see as non-core” is a trait.

Prompt to try in Claude: “Here are three of our best-fit customers: [names + 2-3 sentences on each - what made them a good or bad fit, how the deal played out, anything you wish you’d known or asked about sooner]. Here are three that churned or were bad fits: [same]. Based on what I’ve shared, what traits seem to distinguish the two groups? What questions would you ask me to sharpen the distinction?”

Run this in rounds — one good, one bad, then another pair, then another. Each round tunes the lens. You end up with a set of real patterns grounded in your actual customer history, not your assumptions about who should be a good fit.

Step 4: Turn patterns into traits, and traits into signals

Now you have a list of good and bad patterns. The next step is leveling up: what are the underlying traits these patterns point to?

A trait is something like “openness to innovation” or “organizational stability” — not a single data point, but an attribute that matters for whether an account will succeed with your product. One trait might have multiple signals. Your job here is to get specific: for each trait, what would actually tell you it’s present?

Prompt to try in Claude: “Here are the patterns we identified across our best and worst customers: [list]. Help me group these into 4-6 underlying traits. For each trait, suggest 2-3 observable signals that would indicate it’s present or absent.”

This is where vague instincts become something you can actually work with — and eventually, something you can hand off.

Step 5: Map each signal to when and how you can find it

For each signal, you need to answer two questions: could you find evidence of this from outside-in research before a conversation? Or does it only surface during the sales process itself?

This matters because it shapes your list-building logic and your qualification workflow. Some signals are research-accessible — job postings, LinkedIn, news, a company’s website. Others only show up in discovery. Both are useful, but you need to know which is which before you can build a repeatable process around them.

Prompt to try in Claude: “Here are our qualification signals: [list]. For each one, help me identify: (1) whether this is findable through outside-in research or only through direct conversation, (2) where specifically you’d look to find it, and (3) what ‘good’ vs. ‘concerning’ looks like for each signal.”

The output is your qualification rubric: a structured document a new rep could follow, or that Claude could use to research and pre-qualify accounts at scale.

Ready to give it a go?

Download the skill here. It guides you through the process to produce a rubric for account qualification tailored to your business.

Download the Claude Skills

What good looks like

A finished rubric has 4-6 traits, each with 2-3 signals. For each signal, you know: what it is, why it matters to your business, what good looks like versus what’s concerning, and where to find it. It’s specific enough that two different people reviewing the same account would reach the same conclusion.

The practical test Brian and Mallory used at MedScout: could a new rep use this rubric to evaluate an account the way your best rep would? If yes, you’re done. If not, keep iterating.

Watch-outs

The first pass will feel incomplete. That’s normal — the goal of the shadowing step and the best/worst customer analysis is iteration, not perfection on round one. Don’t try to build the definitive rubric in a single session.

Don’t let the rubric drift into a yes/no checklist either. The value is in documented judgment with reasoning, not a binary score. You want to be able to defend why a signal matters and what you’re looking for — not just whether a box got checked.

The takeaway

A target account list built on filters alone is a starting point. What turns it into something your team actually trusts is the reasoning underneath it — both your ICP guidelines and the criteria your best reps apply instinctively, documented clearly enough that anyone can follow it.

That documentation is worth building for its own sake: it trains new reps, aligns the team, and creates a shared standard for what good looks like. And once it exists, it becomes the foundation for Claude to help — not replacing your judgment, but doing the research legwork against criteria you’ve already defined.

Special thanks to Brian Aoyama, VP Growth & Applied AI, and Mallory Blocker, Growth Marketing Lead, at MedScout for sharing their process and the hard-won lessons behind it.

Your rep is on a customer call and the customer asks a detailed product question. The AE doesn’t know the answer, says they’ll follow up, and then spends 20 minutes tracking down an engineer or sorting through a database to find the answer. Sound familiar?

Subscribe now

The problem

Your team is splitting their attention across CRMs, dashboards, Slack threads, and internal wikis — and still can’t get fast, accurate answers about your own product. Engineers field the same questions on repeat. CS teams interrupt each other to find customer-specific data. Product knowledge lives in too many places to be useful in the moment. The best information your company has is locked behind people, not accessible through systems.

The setup

A Claude Code agent that lives in Slack with read access to your codebase and key tools can answer most of those questions in under 3 minutes — while the rep is still on the call. Jeff Silver, CTO and Co-Founder at Opus, built exactly this. They call it Product Bestie.

And he’s sharing the recipe so you can build it, too.

Step 1: Find an engineering resource

If you’re a RevOps leader or revenue leader reading this and thinking I want this for my team — great, that’s the right instinct. But the actual build requires a technical founder, senior engineer, or CTO to set up. Your job is to define the use case and make the case for the investment. Their job is to wire it together.

The question to answer before you start: What tools does an engineer need to answer the questions my team asks most often? That list becomes the blueprint.

Step 2: Set up the four components

Product Bestie has four moving parts. You don’t need all of them on day one — but you do need to understand what each one does.

1. A Slack App — Create a Slack app with bot token scopes for reading mentions and posting replies. Subscribe to the app_mention event. This is what lets the agent “hear” when someone tags it.

2. An execution environment — A machine where Claude Code is installed with your codebase cloned and dependencies loaded. Opus uses Fly.io containers with pre-built “golden volumes” so agents spin up in seconds. A VM, EC2 instance, or even a local machine works for V1.

3. An agent definition file — A markdown file (.claude/agents/your-agent-name.md) that tells the agent what to do. Write it like you’re briefing a new hire: here’s how to read a Slack thread, here’s when to query the database, here’s what a good answer looks like, here are the formatting rules.

4. Skill files — Markdown files that teach the agent how to use specific tools. Opus gives theirs: a Slack CLI (to read and post messages), access to some anonymized customer data, , Datadog (for log and session search), Linear, and GitHub. Each skill file is essentially a how-to guide for the agent.

Example agent prompt structure:

Read the full Slack thread context.

Identify what question is being asked.

Check the codebase for relevant logic.

Query the production DB if needed for customer-specific data.

Post a reply in the thread. Be specific. If you found exact data, show it.

If you’re an engineer reading this, include what you did at the bottom.

Step 3: Lock down security before you go further

This is the step people skip. Don’t.

If your agent has read access to customer data — and it should, if you want it to answer real questions — you need to own the infrastructure. The reason Opus built on Fly.io instead of using Cursor or an off-the-shelf Claude integration is control: they know exactly what’s in the container, they control the secrets, and the container gets destroyed when the job is done. The key security principles your technical partner should be focusing on are: access scoping (no access to PII), secret management, and ephemeral execution.

Know what data your agent can touch. Know who can query it. Build the lockdown before you give reps access.

Step 4: Add observability so it gets smarter over time

Product Bestie isn’t a set-it-and-forget-it tool. Opus built an internal dashboard (they call it Magnum) where engineers can see everything the agent has done — every question, every response, every action. Reps can give a thumbs up or thumbs down on responses. A thumbs down prompts feedback, and the agent learns not to repeat that mistake.

This is the part that makes it genuinely useful over time instead of just a neat demo. If you’re not watching what it does, you’re not managing it. Build in the feedback loop early.

What good looks like

Here’s a real example from Opus. A CS rep was on a call with a new account, and was asked a question about a specific product feature. Rather than a “let me get back to you” they were able to ask Bestie and get an answer a few minutes later, before the call ended.

Or imagine a scenario where an AE is prepping for a renewal call and pings Bestie to surface the account’s usage trends, recent support tickets, and any feature requests from the past quarter. In just a few minutes, they have a tailored talking-points doc.

Watch-outs

• Security first, always. The moment your agent can touch production data, you’re in different territory. This is not a place to take shortcuts.

• Don’t try to measure ROI too early. The value shows up in rep confidence, faster answers, fewer interruptions for your engineers. It’s real — but it’s diffuse. If you’re managing closely enough to know your team, you’ll know if it’s working. Trust that.

Start simpler if you need to

If building on Fly.io feels like too much for right now, there’s a V1 path: Cursor connected to Slack, pointed at your codebase. You lose the tools that make Bestie extremely valuable, but you gain a fast, low-lift way to start answering product questions in Slack but without relevant customer information.

Jeff’s framing: “You can do this pretty easily if you want to just get answers from the codebase. Things get more complex when you start to introduce more tools, but introducing those tools is where the magic happens..”

Use the simpler version to prove the value internally. Once your team is hooked on getting answers in Slack, the case for the full build becomes easy to make.

The takeaway

The buy vs. build decision is changing fast. For years, SaaS tools won because internal builds weren’t worth the engineering investment. That math is shifting. When a technical founder can build a working V1 in an evening, and when the customization ceiling on off-the-shelf tools keeps frustrating your team, the calculus looks different.

For Opus, Product Bestie isn’t just about answering questions faster. It’s about building infrastructure where every employee spends less of their day searching for answers and more time actually solving customer problems. The companies that figure this out first are going to have a real productivity edge.

Special thanks to Jeff Silver, CTO and Co-Founder at Opus, for sharing how Product Bestie works and letting us build this How-To from his write-up and our conversation.

Subscribe now

DEAR PRIORITIZING CSAT IN AN AI-WORLD: Every founder and support leader is feeling this tension right now.

You see the opportunity. Automate repetitive tickets. Free up your team. Reallocate headcount to strategic work. Improve margins.

But you also know one bad automated experience can undo months of goodwill with a customer.

We asked Mary-Beth Donovan, former Chief Customer Officer at EDB and Stage 2 LP, to share her take on how operators should think about automation without compromising trust. Her framing was simple and practical. There are three lenses to evaluate every automation decision: operational efficiency, customer experience, and product value.

Here is how to approach it thoughtfully.

Start with capacity, not cost savings

The cleanest use case for automation is not eliminating people. It is reallocating their time.

Mary-Beth emphasized that the first win most teams should chase is freeing up one full-time equivalent from low-value interactions. If AI can consistently handle repetitive, predictable questions, your team can move upstream into more proactive and strategic conversations.

This shift matters. Customers will tolerate automation if it leads to better human engagement where it counts.

Before setting an automation target, ask:

• Which interactions are repetitive and rules-based?

• Which conversations require judgment and context?

• What higher-value work would we do if we reclaimed this time?

If automation does not unlock a better customer experience somewhere else, it is the wrong goal.

Trust and tolerance are the real metrics

Mary-Beth introduced a powerful framing around trust and tolerance.

You do not need 100 percent perfection to automate. In many environments, good enough is acceptable if you preserve trust. The danger is over-automation that traps customers in loops or prevents escalation when they need it.

Every automated workflow should clearly answer one question. When does a human step in? That moment should not feel hidden or punitive. It should feel seamless.

Practical guardrails include:

• A clear path to escalate to a human

• Monitoring negative feedback patterns

• Regular review of cases that required escalation

Automation without a visible safety net erodes confidence quickly.

You cannot automate bad documentation

One of the most common surprises teams encounter is that their knowledge base is not ready for automation.Outdated documentation, inconsistent articles, or unclear processes will surface immediately once AI begins referencing them. Automation does not fix weak foundations. It amplifies them.

As an aside, I recently had a pretty funny intersection with an AI chatbot. I asked a “how to” question, and was immediately told it wasn’t possible. After playing around for 5 minutes, I found a way to do it and replied to that effect in the chat. The AI asked “can you walk me through how you did it so I can give better feedback to the next question?”. Rather than insisting I was wrong, I was surprised and delighted to see that the company had thoughtfully architected this moment for learning and growth.

Before increasing your automation rate, invest in tightening your knowledge base and documenting edge cases. The quality of your answers will determine whether customers view automation as helpful or frustrating.

This is often less about technology and more about discipline.

Set targets, and revisit them quarterly

There is no universal “right” automation percentage. The correct level depends on your customer base, complexity of product, and support model.

Mary-Beth cautions against chasing an arbitrary number like 80 percent automation without considering the downstream impact. Instead, set a goal tied to a clear business outcome. For example, freeing up capacity for proactive outreach or improving time to resolution on complex cases.

Then measure not only deflection rates, but also customer satisfaction and expansion indicators. If satisfaction drops or renewal conversations become harder, reassess.

Automation is not a one-time project. It is an evolving balance between efficiency and experience. When done well, it does not replace your team. It elevates them.

Until next week!

The Starter Play: How to Build a Marketing Attribution Report in 10 Minutes Using Claude + HubSpot (or your CRM)

Marketing attribution has been the thing everyone wants but it’s hard to generate. You know the conversation — “we need to understand what’s driving pipeline” — but it can take weeks to pull, clean and analyze the data to get a directionally correct answer.

Here’s the thing: if your data is in HubSpot (or any CRM that has a connector to Claude), you can have a working attribution report today that tells you what marketing activities are actually correlating with closed revenue.

Subscribe now

Before You Start

• Claude account with the HubSpot connector enabled (this is key — without the connector, Claude can’t read your data). If you are using a different CRM, just make sure that it is also connected to your Claude account and follow these same instructions).

• HubSpot access with closed/won deal data and at least some marketing activity tracked against contacts

• A rough idea of the time window you want to analyze (last 6-9 months is a good starting point)

• Deals should have contacts associated — Claude reads contact-level properties, not just deal records

Step 1: Connect Claude to HubSpot

In Claude, open a new conversation and enable the HubSpot connector from the tools panel. You don’t need to export anything or build a prompt file ahead of time. Claude reads your HubSpot data directly once the connector is active.

Step 2: Run the Attribution Prompt

Start with this prompt, adjusted for your time window:

“Look at every deal that closed won in the last 9 months. Analyze the activity of each contact associated with those deals. Determine what marketing activities generated closed revenue. Give me an executive summary with: total revenue, average deal size, time to close, and the top marketing touchpoints correlated with winning. Then break it down by deal tier if you can.”

Claude will work through your data and return a structured report — not just text. Expect an executive summary, a breakdown by marketing activity, and a recommendations section. Brant Morton at Rally got a Markov chain analysis on his first try, which he wasn’t even expecting.

Your output looks something like this:

Step 3: Iterate on the Report

Don’t stop at the first output. Keep prompting to go deeper. Try:

“Now show me first touch and last touch attribution separately.”

“Show me the most common activity sequences for your top 20% of deals by deal size.”

“What marketing activities appeared in more than 50% of closed won deals?”

Each follow-up refines the picture. The first report is a starting point, not the final answer.

Step 4: Read the Output Critically

Claude can only read HubSpot contact properties right now — not full activity timelines or custom objects. That means it’s working from fields like first conversion, most recent conversion, and MQL date. The insights are real, but they’re directional. If your report tells you “one-on-one sales emails are your top driver,” that’s probably true — and also not very actionable. Push Claude to surface the marketing signals that aren’t obvious.

Step 5: Screenshot and Document What Surprised You

Before you close the conversation, take screenshots of the sections that surfaced something unexpected. This is your working version. You’ll build on it.

What Good Looks Like

A solid first-pass report gives you: total pipeline influenced, average deal size by segment, the top 3-5 marketing touchpoints correlated with winning, and at least one insight you didn’t already know. Brant’s report surfaced that in-person dinners were showing up in enterprise deals while content downloads were driving smaller ones — a genuine signal that shaped their demand gen priorities.

Watch-Outs

Claude is reading your HubSpot data, which means it’s only as good as what’s actually logged. If your team isn’t consistent about recording activity — or if marketing touches aren’t tracked at the contact level — the report will reflect that. Garbage in, limited insight out.

Also: custom objects in HubSpot are not currently readable by Claude’s connector. If you’ve built custom conversion tracking, you won’t see it here. That’s what Version 2 is for.

The Takeaway

This would have taken a data analyst weeks to build. You did it in one conversation. The report isn’t perfect — but it’s real, it’s based on your actual data, and it gives you something to react to. That’s more than most teams have. Start here, pressure-test the output with your sales and marketing leads, and then decide what you want to go deeper on.

The Advanced Play

How to Build a Full Marketing Attribution App Using HubSpot, Supabase, and Lovable.

Claude’s HubSpot connector is a great starting point. But it has a ceiling: 200K tokens, no custom objects, no full activity timelines. If you want attribution across your entire history — every touchpoint, every channel, every contact — you need to pull the data out of HubSpot and build something on top of it.

The good news: you don’t need a product team to do it. Brant Morton at Rally built this with a $30/month Lovable subscription, a free Supabase table, and a few weekends. It’s running in production.

This is Version 2. Do Version 1 first — it’ll tell you what questions to go answer with this.

Before You Start

• HubSpot access with API permissions (you’ll be pulling data via automation, not the Claude connector)

• A Make.com or n8n account to set up the HubSpot → Supabase workflow

• A Supabase account (free tier works to start)

• A Lovable account ($30/month starter tier is enough)

• At least one quarter of closed/won deal data in HubSpot with marketing activities logged

• Basic comfort with no-code automation — you don’t need to write code, but you need to follow a workflow

Step 1: Set Up Your Supabase Data Table

Create a table in Supabase to receive marketing touchpoint data. You’ll want columns for: contact ID, company ID, event type, event date, campaign, channel, and deal association (if any). This is your attribution database. Everything else gets built on top of it.

Step 2: Build the HubSpot → Supabase Sync in Make or n8n

Create a workflow that fires every time a marketing interaction is logged in HubSpot — email delivery, form submission, ad interaction, webinar attendance, sales call — and writes it to your Supabase table. This is the key architectural decision: instead of asking Claude to re-read everything every time, you’re building a running log that never hits a token limit.

Brant pulled in: marketing email deliveries, ad interactions, sales emails, calls, form submissions, and webinars. That gave him 442,000 touchpoints — way beyond what Claude’s HubSpot connector can handle in a single session.

Step 3: Build the Lovable App on Top of Supabase

Open Lovable and connect it to your Supabase table. Start with a prompt like:

“Build a marketing attribution dashboard connected to my Supabase table [table name]. I want to see: total touchpoints by channel, attribution by model (first touch, last touch, linear, time decay, W-shaped), and a contact-level view that shows every interaction a contact had before a deal closed.”

Lovable will build the app. Iterate from there. Brant added a company rollup view (so you can see activity across all contacts at an account, not just the one on the deal), a paid ads page pulling LinkedIn and Google Ads data, and a “levers to pull” page that surfaces recommendations by attribution model.

Step 4: Add Attribution Models

The real power is running multiple attribution models against the same dataset. Start with first touch and last touch — those are easy. Then add linear and time decay. Markov chain is the most accurate but requires a Python calculation running in the background; Lovable can set it up but you’ll need to fine-tune the logic.

The point isn’t to pick one model as “right.” It’s to see where they agree and disagree — and investigate the gaps.

Step 5: Use It to Ask Better Questions

The dashboard is a question-generator, not a final answer. Use it to surface: which two or three channels show up in almost every closed/won deal, where marketing and sales activity overlap (and where they don’t), and which touchpoints appear in enterprise deals vs. smaller ones. Brant found that in-person dinners correlated with his largest deals — something that never would have surfaced in a standard HubSpot attribution report.

What Good Looks Like

A fully working version of this gives you a live dashboard with six attribution models, contact-level activity timelines, company-level rollups, and a paid ads ROI view — all from your own data, without a third-party attribution platform. Brant built Rally’s version in a weekend

Watch-Outs

Your data quality matters more here than in Version 1 — you’re running actual statistical models. If your team isn’t logging activities consistently, the models will surface noise as signal. Audit your HubSpot data hygiene before you build.

Also: Markov chain analysis requires some tuning. Don’t trust the first output of that model without cross-referencing it against your simpler models.

The Takeaway

Three months ago, standing up a marketing attribution system like this would have required a dedicated PM, a team of engineers, and months of roadmap time. Juliette Kopecky said it clearly: at HubSpot, they had an entire product team devoted to tools like this. Brant built the equivalent over a weekend. That’s not hyperbole — that’s what this moment actually is. The question isn’t whether you can afford to build this. It’s whether you can afford not to know what’s actually driving your revenue.

Special thanks to Brant Morton, GTM Ops, and Juliette Kopecky, CMO, both at Rally, for building this in public and sharing every detail.

Subscribe now