LeanScale CTO Jake Toepel opens the hood on the agent fleet behind a whole portfolio — and why one company's messy definitions become thirty boards' worth of wrong answers
Why this conversation matters — and who should spend the hour.
LeanScale runs revenue operations for dozens of fast-growing SaaS companies simultaneously — different CRMs, different data, different definitions of basically everything — with a small team and a stack of AI agents. In this companion to his AI-native GTM video, CTO Jake Toepel shows what that looks like under the hood at portfolio scale, then explains why none of it works out of the box.
The organising idea is that AI-native go-to-market does not run on vibes, it runs on operations. RevOps is supposed to be the AI operations function and at most companies it simply is not yet, so LeanScale started calling the work AI ops: the connective tissue of definitions, data model, skills and workflows that agents run on. Somebody has to build and maintain that foundation, and Jake frames it as the difference between AI that demos beautifully and AI you would trust to run the business.
Three plays follow. The day-one diagnostic replaces weeks of discovery — stakeholder interviews, spreadsheet archaeology, a deck at the end — with an agent pointed at connected data on day one, running the full teardown: funnel conversion by stage, where deals stall, rep coverage, pipeline that is quietly dead. It writes up against the playbook in the firm's brand voice, turning three weeks of a senior consultant's time into a first draft in twenty minutes. The consultant stays in the loop but is now editing judgement rather than assembling data, which is how a small team credibly covers dozens of accounts. The second play changes the QBR itself: instead of "great question, we'll follow up next week," a client asking why net revenue retention dipped in mid-market gets the segment, the churned accounts, the reason codes and the expansion that did not land pulled live on screen. The QBR stops being a report presented and becomes a conversation — and every answer uses the client's own definitions rather than a generic benchmark.
The third play is the one Jake is most excited about, because it is how the agency runs itself. Delivering for dozens of clients is operationally brutal: every call creates work, every account needs watching, every person needs coaching. Project-management agents read each call transcript, extract commitments and action items, and turn them into scoped, assigned tasks with nobody retyping anything. A customer-health agent reads every transcript and Slack channel for signals that an account is heating up or cooling off, long before a QBR would surface it. A team-evaluation agent watches delivery itself. What makes it compound is that everything those agents learn feeds back into the playbooks, so the agency improves itself while the work happens.
Then the same demonstration as before, with the stakes multiplied. Raw AI asked for pipeline coverage ratio returns a confident, instant, wrong answer, because four things are missing — shared definitions, identity resolution, plans and memory — and across dozens of companies each one multiplies. Every client defines pipeline, qualified and churned slightly differently, and raw AI blends them into mush. At one company that is annoying; at thirty it is chaos, and shipping confident wrong answers to thirty boards is not a business but a liability. The foundation is the context graph, built on Vasco: a semantic layer mapping each client's raw data to what it means and resolving every record to one source of truth across CRM, billing and product. The skills, plugins, workflows and interfaces clients use are then built on top. A foundation that is true, and the operation that runs on it.
The load-bearing ideas, each with the business implication and who should care.
Jake's organising claim: if you want agents running go-to-market, somebody has to build and maintain the foundation they run on — the definitions, the data model, the skills, the workflows.
Why it matters: Treat the foundation as staffed, ongoing work rather than a project that completes. It is the difference between AI that demos beautifully and AI you would trust with the business.
Jake names the gap directly and coins AI ops for the connective tissue the role should already own — which is why LeanScale had to name and staff it explicitly.
Why it matters: Assign ownership of the agent foundation to a named function. Where it is nobody's job, the definitions drift and the agents inherit the drift.
The old way was weeks of discovery — stakeholder interviews, spreadsheet archaeology, a deck at the end. The agent runs the full teardown on connected data on day one: funnel conversion by stage, where deals stall, rep coverage, pipeline that is quietly dead, written up in brand voice against the playbook.
Why it matters: The compression is in assembly, not judgement. What changes is where senior time goes, not whether senior time is needed.
Jake is explicit that the human stays in the loop. The agent produces a first draft; the expert's contribution shifts to the judgement layer on top of it.
Why it matters: Design the workflow around review rather than replacement, and expect the skill profile of the role to shift toward interpretation rather than collection.
The economics of portfolio delivery are the point: the diagnostic play is what makes it possible to serve many clients at senior quality without proportional headcount.
Why it matters: The leverage argument for agents is coverage at quality, not cost reduction — a different case to make and a different thing to measure.
"Great question, we'll follow up next week" becomes pulling the segment, the churned accounts, the reason codes and the expansion that did not land, live on screen while the client watches. The meeting stops being a report presented and becomes a conversation.
Why it matters: Live answers change what a client meeting is for. The preparation burden shifts from anticipating questions to having a foundation that can answer unanticipated ones.
Jake draws the distinction explicitly — the live QBR answer is theirs, from their data and their definitions, rather than an industry number pulled from somewhere.
Why it matters: A benchmark answers a different question than a client's own data does. Conflating them is how advice stops being specific enough to act on.
Every call transcript that lands is read by an agent that pulls out commitments and action items and turns them into real tasks and projects on the board, scoped and assigned, with nobody retyping anything.
Why it matters: The gap between what was committed on a call and what appears on a board is a reliable source of dropped work. Closing it mechanically removes a failure mode rather than speeding one up.
A customer-health agent reads every transcript and Slack channel for signals an account is heating up or cooling off, long before a QBR would surface it. A team-evaluation agent watches delivery — who is excelling, where someone is stuck, where coaching is needed.
Why it matters: Health signals already exist in conversation exhaust the company is generating anyway. The question is whether anything is reading them between formal reviews.
What the agents learn from transcripts, Slack channels and the board goes straight back into the playbooks. Every call run and project delivered sharpens the system, and the next client gets the benefit.
Why it matters: The compounding loop is the actual asset, above any individual agent. Without the feedback path the agents are efficiency; with it they are an accumulating advantage.
Every client defines pipeline, qualified and churned slightly differently, and raw AI blends them into mush. Running on raw AI across a portfolio means shipping confident wrong answers to thirty different boards.
Why it matters: The cost of missing definitions scales with the number of contexts served. What is tolerable inside one company becomes a business risk for anyone operating across many.
Shared definitions of pipeline, qualified and churned; identity resolution when billing says churned and the CRM says wide open; the plan each number is measured against; and the memory that makes an answer this client's rather than generic.
Why it matters: Use the four as the readiness test before agents touch revenue data, at any scale. Each absence produces its own predictable class of wrong answer.
Vasco provides the context graph — a semantic layer mapping raw data to meaning and resolving records to one source of truth across CRM, billing and product. The skills, plugins, workflows and interfaces clients use are built above it.
Why it matters: Sequence matters more than component choice. Interfaces built before the semantic layer inherit whatever ambiguity is underneath and present it confidently.
Jake's framing of the third play: not one assistant, but a whole operating system running the business behind the scenes — project management, customer health and team evaluation running simultaneously.
Why it matters: Evaluate agent investments as infrastructure with a maintenance burden, not as productivity tools adopted per person.
Every framework Jimmy names, defined and time-stamped.
The connective tissue agents run on — definitions, data model, skills and workflows — built and maintained as an operating function rather than assembled per project.
Named because RevOps is supposed to be the AI operations function and at most companies is not yet. Jake presents it as what separates AI that demos beautifully from AI you would trust to run the business.
An agent pointed at a new client's connected data on day one, running the full teardown — funnel conversion by stage, stalled deals, rep coverage, quietly dead pipeline — and writing it up against the playbook in brand voice.
Replaces weeks of stakeholder interviews and spreadsheet archaeology with a twenty-minute first draft. The consultant remains in the loop, editing judgement rather than assembling data.
A quarterly review where unanticipated client questions are answered live from the client's own data and definitions, rather than deferred to a follow-up.
Jake's example is a question about a mid-market NRR dip returning the segment, churned accounts, reason codes and missed expansion on screen. The meeting stops being a report and becomes a conversation.
Three background agents operating the delivery business itself: project management turning call transcripts into scoped tasks, customer health reading transcripts and Slack for account signals, and team evaluation watching delivery quality and coaching needs.
Described as an operating system rather than an assistant. Its defining property is the feedback loop — field learning returns to the playbooks, so the system sharpens with every call.
Build the context graph first — the semantic layer resolving definitions, motion, plan and identity across systems — then build the skills, plugins, workflows and interfaces on it.
The stated build order. Vasco supplies the foundation; everything clients touch day to day sits above it and inherits its trustworthiness.
Pulled verbatim. Copy or share any of them.
“We run revenue operations for dozens of fast-growing SaaS companies at the same time, different CRMs, different data, different definitions of basically everything, and we do it with a small team and a stack of AI agents.”
“It is the most critical piece of AI-native go-to-market, and it's the reason any of this can be trusted at scale.”
“Everyone wants AI-native GTM, but AI-native GTM does not run on Vibes. It runs on operations.”
“RevOps is supposed to be the AI operations function. At most companies, it just isn't yet.”
“If you want agents running your go-to-market, somebody has to build and maintain the foundation they run on.”
“That is the difference between AI that demos beautifully and AI you would actually trust to run the business.”
“What used to be three weeks of a senior consultant's time is a first draft in 20 minutes.”
“The consultant is still in the loop, but now they're editing judgment, not just assembling data.”
“That is how a small team credibly covers dozens of accounts.”
“It used to be the client asked question we didn't prep for, and we say great question, we'll follow up next week. But now the answer happens in the room.”
“The QBR stops being a report you present and becomes a conversation you can actually have.”
“Every answer is theirs, their definitions, their data, not a generic benchmark I just pulled from somewhere.”
“Delivering for dozens of clients at once is operationally brutal. Every call creates work, every account needs watching, every person needs coaching.”
“The work goes from the call straight to the board with nobody retyping a thing.”
“A customer health agent reads every transcript and every Slack channel for the signals that an account is heating up or cooling off, long before it would ever surface in a QBR.”
“Everything those agents learn in the field from the transcripts, the Slack channels, the board, feeds straight back into our playbooks.”
“The agency is improving itself while we work.”
“Not one assistant, but a whole operating system running the business behind the scenes.”
“Every client defines pipeline, qualified, and churned just a little bit differently. And raw AI blends them all into mush.”
“At one company, it's annoying. At 30 companies, it's chaos.”
“If we ran just on raw AI, we would be shipping confident, wrong answers to 30 different boards. That is not a business. That's a liability.”
“Vasco gives us a foundation that is true. We build the operation that runs on top of it.”
The playbook, split by the seat you sit in.
LeanScale's signature read on the AI-in-GTM question this episode wrestles with.
This is the portfolio-scale version of the context-graph argument, and scale is what makes it an operating risk rather than an inconvenience. Every client defines pipeline, qualified and churned slightly differently; raw AI blends them into mush; and the output is confident wrong answers delivered to thirty separate boards. The interesting design choice is that LeanScale turned its own delivery into the proving ground — the agents run the agency before they run anything else.
The same conversation, filtered for RevOps, pipeline/marketing ops, and customer ops.
The portfolio the agent fleet supports with a small team — and the multiplier that turns definitional ambiguity from an annoyance into a liability.
Senior consultant time to produce a first-draft diagnostic, replaced by an agent working from connected data on day one.
Jake's characterisation of the risk of running on raw AI across the client base — confident wrong answers reaching thirty different boards.
Auto-extracted and linked into the knowledge graph.
The context graph LeanScale builds on — the semantic layer mapping each client's raw data to their definitions, motion and plan, and resolving every record to one source of truth across CRM, billing and product. Named as the foundation everything else is built above.
One of the two surfaces the customer-health agent reads continuously, alongside call transcripts, to detect accounts heating up or cooling off before a QBR would reveal it.
Cited as the licence stack teams assume is sufficient — Jake's point is that handing out seats gets you none of the portfolio-scale plays without the foundation underneath.
Agents and workflows that multiply operator throughput — the 'make humans superhuman' thesis, operationalized.
The episode is a direct tour of the AI-native operating model — the agent plays and the semantic layer that makes their answers trustworthy at portfolio scale.
Embedded, field-facing RevOps leadership and execution — the operator model Jimmy describes, delivered as a service.
The day-one diagnostic and the delivery agent fleet are the mechanics of how embedded RevOps is delivered across many accounts at once.
Define a raise-worthy ARR and build a reliable consumption forecast across finance, data science, and go-to-market.
Pipeline coverage, stalled deals and the definitions underneath them are the forecasting substrate the context graph exists to make reliable.
Generated from the conversation, marked up for search and AI extraction.
AI ops is the connective tissue agents run on: the shared definitions, the data model, the skills and the workflows, built and maintained as an ongoing function. Jake Toepel's position is that RevOps is supposed to be this function already, and at most companies it simply is not yet — so LeanScale named the work explicitly. The distinction matters because if nobody owns the foundation, definitions drift and every agent running on them inherits the drift.
At LeanScale, roughly three weeks of a senior consultant's time becomes a first draft in about twenty minutes. The agent is pointed at a new client's connected data on day one and runs the full teardown — funnel conversion by stage, where deals stall, rep coverage, pipeline that is quietly dead — written up against the firm's playbook in its brand voice. The saving is in assembly rather than judgement: the consultant remains in the loop, but edits interpretation instead of gathering data.
It stops being a report you present and becomes a conversation. Previously an unanticipated client question — why did net revenue retention dip in mid-market — produced a note and a follow-up next week. With a trustworthy semantic layer underneath, the agent pulls the segment, the churned accounts, the reason codes and the expansion that did not land onto the screen during the meeting. Crucially the answer uses the client's own definitions and data rather than a generic industry benchmark.
Three, in LeanScale's case. A project-management agent reads every call transcript, extracts commitments and action items, and creates scoped, assigned tasks on the board without anyone retyping them. A customer-health agent reads transcripts and Slack channels continuously for signals that an account is heating up or cooling off, ahead of any formal review. A team-evaluation agent watches delivery to surface who is excelling, where someone is stuck, and where coaching would help. What ties them together is that their field learning feeds back into the playbooks.
Because the error multiplies by the number of contexts. Every client defines pipeline, qualified and churned slightly differently, and raw AI blends those definitions together. Inside a single company an inconsistent definition is annoying and usually caught by someone who knows the business. Across thirty companies it means shipping confident, wrong answers to thirty different boards — which Jake describes not as a business but as a liability.
Foundation first. The context graph — a semantic layer that maps raw data to what it means in the business, carries the definitions, motion and plan, and resolves every record to one source of truth across CRM, billing and product — comes before anything else. LeanScale builds this on Vasco. Only then do the skills, plugins, workflows and interfaces that people use day to day get built on top. Building interfaces first means they inherit whatever ambiguity sits underneath and present it confidently.
Click any chapter to open it in the transcript.
Broken into chapters, searchable, verbatim from the audio. Speakers inferred (not diarized).
0:00 We run revenue operations for dozens of fast-growing SaaS companies at the same time, different CRMs, different data, different definitions of basically everything, and we do it with a small team and a stack of AI agents.
0:15 Most people have never seen what that actually looks like under the hood.
0:19 AI-native go-to-market run not for one company, but across an entire portfolio at once.
0:25 That's what I'm going to show you today.
0:27 I'm going to show you the exact place we run, and then because none of this comes out of the box, I'm going to show you exactly how it's built, and it all comes down to one thing, the context graph.
0:38 It is the most critical piece of AI-native go-to-market, and it's the reason any of this can be trusted at scale.
0:45 So go through the place with me first, but stay to the end because that's where I break it down.
0:51 Let's get into it.
0:52 Here's the idea that reorganized how we think about all this.
0:56 Everyone wants AI-native GTM, but AI-native GTM does not run on Vibes.
1:01 It runs on operations.
1:03 RevOps is supposed to be the AI operations function.
1:07 At most companies, it just isn't yet.
1:11 So we started calling the thing we actually do AI ops, the connective tissue.
1:17 If you want agents running your go-to-market, somebody has to build and maintain the foundation they run on.
1:23 The definitions, the data model, the skills, the workflows, that is AI ops.
1:29 And that is the difference between AI that demos beautifully and AI you would actually trust to run the business.
1:35 I'm Jake, the CTO at LeanScale, and this is how we do it across a whole book of clients.
1:41 Three plays, here's the first.
1:43 Play number one, the diagnostic.
1:46 When a new SaaS company comes on board, the old way was weeks of discovery.
1:51 Stakeholder interviews, spreadsheet archaeology, and finally a deck at the end.
1:57 Now on day one, I point our diagnostic agent at their connected data and ask it to find the leaks.
2:03 It runs the whole teardown, funnel conversion by stage, where deals stall, rep coverage, pipeline that is quietly dead.
2:11 And it writes it up against our playbook in our brand voice as a finished diagnostic.
2:17 What used to be three weeks of a senior consultant's time is a first draft in 20 minutes.
2:22 The consultant is still in the loop, but now they're editing judgment, not just assembling data.
2:28 That is how a small team credibly covers dozens of accounts.
2:33 Now play number two is the QBR, and this is where it gets fun because the meeting itself changes.
2:39 It used to be the client asked question we didn't prep for, and we say great question, we'll follow up next week.
2:47 But now the answer happens in the room.
2:50 The client asks, why did net revenue retention dip in the mid market?
2:54 Instead of taking a note, I just ask.
2:57 It pulls the segment, the churned accounts, the reason codes, the expansion that didn't land, live on screen while they watch.
3:05 The QBR stops being a report you present and becomes a conversation you can actually have.
3:10 And every answer is theirs, their definitions, their data, not a generic benchmark I just pulled from somewhere.
3:18 Play number three, and this is the one I'm most excited about because it is how we run the agency itself.
3:24 Delivering for dozens of clients at once is operationally brutal.
3:28 Every call creates work, every account needs watching, every person needs coaching.
3:33 So we built a layer of agents to run it.
3:36 First is project management agents.
3:38 Every call transcript that lands, an agent reads it, pulls out the commitments and the action items,
3:43 and turns them into real tasks and projects on our PM board, scoped and assigned.
3:49 The work goes from the call straight to the board with nobody retyping a thing.
3:54 Then two more agents run constantly in the background.
3:57 A customer health agent reads every transcript and every Slack channel for the signals that an account is heating up or cooling off,
4:05 long before it would ever surface in a QBR.
4:08 And a team evaluation agent watches our own delivery.
4:11 Who's excelling?
4:12 Where is someone stuck?
4:13 And where can we coach?
4:15 And here's the whole thing that makes it compound.
4:18 Everything those agents learn in the field from the transcripts, the Slack channels, the board,
4:22 feeds straight back into our playbooks. Every call we run, every project we deliver,
4:27 the system gets a little bit sharper and the next client gets the benefit.
4:31 The agency is improving itself while we work.
4:34 That is what operationalizing delivery with agents actually looks like.
4:39 Not one assistant, but a whole operating system running the business behind the scenes.
4:44 Now, everything I just showed you, you cannot do by handing your team a stack of clawed licenses.
4:50 And at our scale, the reason is brutal.
4:53 Watch raw AI on real data.
4:55 Same simple question.
4:57 What is my pipeline coverage ratio?
4:59 Confident, instant, and wrong.
5:02 Because four things are missing.
5:04 And across dozens of companies, every one of them multiplies.
5:08 The first is shared definitions.
5:11 Every client defines pipeline, qualified, and churned just a little bit differently.
5:16 And raw AI blends them all into mush.
5:19 Identity resolution.
5:21 Billing says this account churned, the CRM says it's wide open.
5:24 Is this the same account across all three systems?
5:27 And then plans.
5:29 What is the actual target each number is measured against?
5:33 Finally is memory.
5:35 The context that makes it this client's answer and not a generic one.
5:39 At one company, it's annoying.
5:41 At 30 companies, it's chaos.
5:43 If we ran just on raw AI, we would be shipping confident, wrong answers to 30 different boards.
5:49 That is not a business. That's a liability.
5:52 So here is how it's actually built.
5:54 The thing I told you to stay for and the part I am genuinely proud of.
5:58 The foundation is the context graph.
6:00 And for that we use Vasco.
6:02 It is a semantic layer that maps every client's raw data to what it actually means.
6:07 Their definitions, their motion, their plan, and resolves every record to one source of truth across CRM, billing, and product.
6:17 That is the layer that makes the answers trustworthy.
6:20 And then on top of it, we build the skills, the plugins, the workflows, the interfaces our clients actually use day to day.
6:28 So think of it this way.
6:30 Vasco gives us a foundation that is true.
6:33 We build the operation that runs on top of it.
6:36 A trustworthy context graph plus the operations built on top of it.
6:40 Now that is AI ops.
6:42 That is the thing that lets a small team run go to market for an entire portfolio without it falling apart.
6:49 So whether you're a SaaS company that wants this for yourself, an agency trying to scale past the hours for dollar trap,
6:55 or a fund that wants this across your entire portfolio, the mechanics are the same and they're real.
7:01 And the place to start is the same and honest assessment of where your foundation is today.
7:07 Link is in the description below.
7:10 Now, if you want a deeper dive on all of this, Guillaume from Vasco and I are doing a session on what it actually takes to become an AI native agency.
7:18 That one is linked below.
7:19 Feel free to register now.
7:21 And if you want to see what this looks like for a single SaaS company instead of a whole portfolio,
7:26 I have a separate video that walks through exactly that.
7:29 It'll be in the description as well.
7:31 I'm Jake from Lean Scale.
7:33 Thank you so much for watching and we'll see you in the next one.