---
title: "Your AI Strategy Is Really a Data Model Problem"
episode: 129
podcast: "The LeanScale Podcast"
publisher: "LeanScale"
guest: "Anthony Enrico"
guest_title: "Co-Founder"
date_published: 2026-09-30
date_modified: 2026-10-21
duration: 00:06:15
word_count: 995
topics: ["ai-in-gtm", "revenue-operations", "gtm-strategy", "forecasting"]
canonical_url: https://www.leanscale.team/knowledge/podcast/anthony-enrico-leanscale-ai-strategy-really-data-model-problem/
source: "LeanScale Knowledge Hub — https://www.leanscale.team/knowledge"
license: "Free to quote and cite with attribution to The LeanScale Podcast."
---

# Your AI Strategy Is Really a Data Model Problem

_Anthony Enrico, solo, on why everyone rents the same Formula One engine and what actually separates you: the semantic layer, the context graph, the one-owner rule, never letting the model write your business logic, and three things to do on Monday_

**Episode 129 · The LeanScale Podcast**  
Anthony Enrico, Co-Founder (LeanScale)  
Published September 30, 2026 · Updated October 21, 2026 · 00:06:15  
Canonical: https://www.leanscale.team/knowledge/podcast/anthony-enrico-leanscale-ai-strategy-really-data-model-problem/

**Topics:** AI in GTM · Revenue Operations · GTM Strategy · Forecasting


## Executive summary

Anthony Enrico, co-founder and CEO of LeanScale, opens this six-minute solo episode with a test. Ask Claude, connected to Salesforce and Gong, which customers are likely to expand this quarter. The answer comes back confident and well written, and there is no way to tell whether it is true. His argument is that this is not the model's fault. An AI strategy is really a data model problem.

The simple question hides at least three definitions. What counts as a customer: the account field, pilots, a company that has already given notice? What counts as expansion: more committed contract or more usage? And what does likely mean? Anthony has watched four people on one leadership team give four different answers. When the AI doesn't know, it infers the middle. He frames the competitive picture with two metaphors. The frontier labs have built a Formula One engine and you and your closest competitor rent the same one, so nobody has a horsepower problem; what is missing is the map. And an Einstein hired on Monday still can't tell you which deals are at risk if Acme Corp, Acme Inc. and Acme Limited sit in three systems with nothing saying they are one company. Intelligence is not the bottleneck. Access is.

The reason to do the work is yield from the go-to-market machine a company has already paid for. On a $20 million plan, one point of conversion, 20 to 21 percent, is worth $1 million, and one day off the sales cycle brings in roughly another quarter million within the year, with no new headcount, no price change and no extra leads. Anthony adds that while frontier model capability climbs every quarter, what a company can actually use has been flat for two years, because it is capped by context rather than by the model.

Closing the gap takes two layers, and neither needs a data science team or a fine-tuned model. The semantic layer, which LeanScale calls its GTM brain, is business logic written in plain English in one place AI can read; LeanScale's lives in GitHub, but a Google shared drive will do. The rule that matters more than tooling is that every metric has exactly one human owner, the person who has to defend the number in the board meeting. The context graph connects contracts to opportunities to accounts to people, and for most companies the CRM is already the first version. Without it, line items fan out so a $10,000 deal forecasts as $30,000, and 150 touch points that look like engagement turn out to be unanswered support tickets. One hard rule: never let the model author business logic, because it will ship generic industry practice. Write it yourself and let AI convert the document to YAML.

Anthony closes with his own experience. Running customer success at Emailage with more than 40 CSMs and hundreds of customers, a trustworthy health score needed a full data science team. With both layers in place it is days of work, and the agents become the easy part. Implementation is about a ten-week project, half of it getting leaders to agree on what words mean. The no-budget start for Monday: write down the 20 definitions every reported number depends on, put one name against each, and record each number's source of truth. The audience is founders, revenue executives and RevOps leaders deciding where AI investment should go first.


## Key takeaways

1. **A confident AI answer you can't verify is a data model failure, not a model failure** — Ask an AI connected to Salesforce and Gong which customers are likely to expand this quarter and it returns something confident and well written, with no way to check it. Anthony says that isn't the model's fault.
   _Why it matters:_ Before blaming or switching models, look at whether the definitions and connections the question depends on exist anywhere the AI can read.
   _For:_ Revenue Executives, Founders, RevOps Leaders

2. **Simple questions hide definitions nobody has agreed on** — Which customers will expand depends on what a customer is (the account field, pilots, a company that has given notice), what expansion is (committed contract or usage), and what likely means. Anthony has watched four people on one leadership team give four different answers to what a customer is.
   _Why it matters:_ Where definitions are missing, the AI infers the middle. If your leadership team disagrees on a term, the AI's answer is a guess dressed as an analysis.
   _For:_ Revenue Executives, RevOps Leaders, Customer Success

3. **Everyone rents the same engine; the map is the differentiator** — The frontier labs have built a Formula One engine that improves every quarter, and you and your closest competitor rent the same one. Nobody in your market has a horsepower problem. What you lack is the map: the corner ahead and the line that won the last ten races.
   _Why it matters:_ Model choice will not separate you from competitors. Your own definitions, owners and connected data will. Without them the car reaches the same corner faster.
   _For:_ Founders, Revenue Executives

4. **Intelligence is not the bottleneck; access is** — Hire Einstein on Monday and ask which deals are at risk, and he can't answer. He sees Acme Corp in the CRM, Acme Inc. in marketing automation and Acme Limited in billing, with no idea whether that is one company or four.
   _Why it matters:_ Identity resolution across systems is a prerequisite for any account-level AI question. A smarter model does not fix records it cannot tie together.
   _For:_ RevOps Leaders, Revenue Executives

5. **The point is yield from the machine you already paid for** — Most companies put millions, often tens of millions, into a go-to-market organization and have never measured what it yields. On a $20 million plan, one point of conversion, 20 to 21 percent, is worth $1 million, and one day off the sales cycle brings in roughly another quarter million within the year.
   _Why it matters:_ That gain needs no new headcount, no price change and no extra leads, and the two levers compound. This is the business case for the data work.
   _For:_ Founders, Revenue Executives, Sales Leaders

6. **Model capability climbs; usable capability has been flat for two years** — What frontier models can do rises every quarter, but what a company can actually use has been flat for two years, because it is capped by the company's context rather than by the model.
   _Why it matters:_ Every point of that gap is the company's to close. Waiting for the next model release will not close it.
   _For:_ Founders, Revenue Executives, RevOps Leaders

7. **Layer one: a semantic layer in plain English, in one place** — The semantic layer, which LeanScale calls its GTM brain, is business logic written in plain English in one place the AI can read: what win rate means, how bookings are counted, what qualified means. LeanScale's lives in GitHub so every model can connect cleanly, but a Google shared drive will do.
   _Why it matters:_ Neither this layer nor the context graph needs a data science team or a fine-tuned model. The tooling bar is low; the discipline is the work.
   _For:_ RevOps Leaders, Founders

8. **Every metric gets exactly one human owner, not a team** — A definition owned by RevOps is owned by nobody. Within a quarter there are two versions of win rate and an argument about which is right. The owner is whoever has to defend the number in the board meeting.
   _Why it matters:_ Anthony says this rule matters more than any tooling choice. Assign names, not departments.
   _For:_ Revenue Executives, RevOps Leaders, Founders

9. **Layer two: the context graph, and your CRM is already the first one** — The context graph connects the dots: contracts to the right opportunities, to the right accounts, with the right people on them. For almost everyone, the CRM is already the first context graph, so cleaning it up gives a solid version zero.
   _Why it matters:_ Skip it and errors pass unnoticed: products fan out across line items so a $10,000 deal shows up as $30,000 in the forecast, or an account with 150 touch points looks healthy until they turn out to be unanswered support tickets.
   _For:_ RevOps Leaders, Sales Leaders, Customer Success

10. **Never let the model author your business logic** — Ask AI what a good renewal looks like and it will grab standard industry practice and ship it, because that is the fastest path from A to B. Unless you run the most vanilla SaaS company alive, that isn't your business.
   _Why it matters:_ People write the definitions; AI converts the Google Doc into YAML, which it does well. The split keeps the logic yours and the formatting cheap.
   _For:_ RevOps Leaders, Revenue Executives

11. **Once the data model exists, the agents are the easy part** — At Emailage Anthony ran customer success with more than 40 CSMs and hundreds of customers, and a health score he could trust was effectively impossible without a full data science team. With the two layers underneath, he says, it is days of work.
   _Why it matters:_ Start with agents instead and you reverse-engineer a private little data model for every agent you build. Build the shared layers first.
   _For:_ Customer Success, RevOps Leaders, Revenue Executives

12. **A ten-week project, and half of it is agreement** — Implementing this is realistically a ten-week project, and half of it isn't engineering. It is getting leaders to agree on what words mean, which has always been the hard part and matters more now.
   _Why it matters:_ The Monday start needs no budget: write down the 20 definitions every reported number depends on, put one name against each, and write down each number's source of truth.
   _For:_ Founders, Revenue Executives, RevOps Leaders


## Frameworks

### Same Engine, Different Map (01:03)

**Definition:** Frontier AI models are a Formula One engine every company rents on equal terms. Competitive advantage comes from the map, a company's own definitions, owners and connected data, not from horsepower.

Anthony's point is that nobody in a market has a horsepower problem, because competitors rent the exact same model. What nobody sells you is the corner ahead and the line that won the last ten races. Without that map, a faster car only reaches the same corner sooner.

### Go-to-Market Yield (01:58)

**Definition:** Measure the revenue a go-to-market organization produces per unit of what has already been spent on it, and improve it through conversion and cycle time rather than new headcount, pricing or leads.

On a $20 million plan, one point of conversion (20% to 21%) is worth $1 million and one day off the sales cycle brings in roughly another quarter million within the year. The two levers compound. Anthony says most companies fund the machine with millions and never measure its yield.

### The Semantic Layer (GTM Brain) (03:08)

**Definition:** A company's business logic written down in plain English in one place AI can read: what win rate means, how bookings are counted, what qualified means.

LeanScale calls its semantic layer the GTM brain and keeps it in GitHub so every model can connect cleanly, but Anthony says a Google shared drive works too. It needs no data science team or fine-tuned model. It is the first of the two layers that close the gap between what models can do and what a company can use.

### The One-Owner Rule (03:27)

**Definition:** Every metric gets exactly one owner, a named human rather than a team: whoever has to defend the number in the board meeting.

A definition owned by RevOps is owned by nobody, and within a quarter there are two versions of win rate and an argument about which is right. Anthony says this rule matters more than the tooling the semantic layer lives in.

### The Context Graph (03:50)

**Definition:** The connections between a company's data: contracts to the right opportunities, to the right accounts, with the right people on them.

Anthony calls it a fancy way of saying connect the dots, and says the CRM is already most companies' first context graph, so cleaning it up gives a solid version zero. Without it, line items fan out (a $10,000 deal forecast as $30,000) and activity counts mislead (150 touch points that are unanswered support tickets).

### Humans Write the Logic, AI Writes the YAML (04:31)

**Definition:** Never let the model author business logic. People write definitions in plain language and AI converts them into structured formats such as YAML.

Asked what a good renewal looks like, a model will ship standard industry practice because it is the fastest path from A to B. Unless the business is entirely generic, that answer is wrong for it. Converting a Google Doc into YAML is the part AI is good at.

### Monday: 20 Definitions, 20 Owners, 20 Sources (05:47)

**Definition:** A zero-budget start: write down the 20 definitions every reported number depends on, put one name against each, and write down where each number comes from.

Anthony offers this as the first step of what is realistically a ten-week implementation, half of which is getting leaders to agree on what words mean rather than engineering.


## Quotes

_Speakers inferred from an undiarized transcript — verify before attributing._

> "And what comes back is confident, well-written, and you'll have no idea whether it's true or not."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (00:12)

> "I don't think that's the model's fault. Your AI strategy is really a data model problem."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (00:19)

> "I've watched four people on one leadership team give four different answers."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (00:39)

> "AI doesn't know, so it does what it always does. It makes up the middle and infers as best as it can."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (00:56)

> "You and your closest competitor are renting the exact same one. Nobody in your market has a horsepower problem."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (01:10)

> "Nobody sells you that, and without it, the car just reaches the same corner faster."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (01:24)

> "Intelligence is not the bottleneck, access is."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (01:53)

> "Do the math on a $20 million plan. One point of conversion, 20% to 21%, is worth a million dollars."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (02:16)

> "No new headcount, no price change, not one extra lead. Same exact machine, higher yield."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (02:31)

> "What you can actually use has been flat for two years because it's capped by your context, not by your model."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (02:50)

> "Every metric gets exactly one owner, a human, not a team. A definition owned by RevOps is owned by nobody."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (03:29)

> "Who's the owner? Whoever has to defend the number in the board meeting."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (03:46)

> "For almost everybody watching, your CRM is actually your first context graph."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (04:01)

> "One hard rule, never let the model author your business logic."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (04:31)

> "What really surprised me is that once the data model exists, the agents are the easy part."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (04:55)

> "It's a 10 week project and half of it isn't even engineering. It's getting your leaders to agree on what words mean."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (05:34)

> "Your model is fine, go build the map."
>
> — Anthony Enrico, The LeanScale Podcast Ep. 129 (06:09)


## Practical advice by role

### Founders

- Stop treating model choice as the AI strategy. Competitors rent the same models; invest in your own definitions, owners and connected data.
- Make the business case in yield terms: on a $20M plan, one point of conversion is $1M and a day off the sales cycle is roughly $250K, with no new headcount, pricing change or leads.
- Plan for about ten weeks to stand up the semantic layer and context graph, and expect half of that time to go to leadership agreement rather than engineering.

### RevOps Leaders

- Write the business logic down in plain English in one place AI can read. GitHub works well; a Google shared drive is enough to start.
- Assign every metric to exactly one named person, the one who defends it in the board meeting, never to a team.
- Clean up the CRM as the first context graph: contracts linked to the right opportunities, accounts and people.
- Resolve company identity across CRM, marketing automation and billing so Acme Corp, Acme Inc. and Acme Limited become one account.
- Check for line-item fan-out in forecasts and for activity counts inflated by support tickets before trusting AI analysis on top of them.
- Write definitions yourself and use AI only to convert them into YAML or other structured formats.

### Revenue Executives

- Test your leadership team: ask each member to define customer and expansion, and resolve the differences before asking AI those questions.
- Do the Monday exercise: list the 20 definitions every reported number depends on, one owner for each, and the source of truth for each.
- Treat a confident AI answer you can't verify as a signal that the data model is missing, not as an insight.

### Customer Success

- Agree on what counts as a customer (pilots, accounts that have given notice) and what counts as expansion (committed contract or usage) before building expansion or health models.
- Build health scores on top of the semantic layer and context graph; with both in place Anthony puts it at days of work rather than a data science project.
- Inspect engagement counts. A high number of touch points can be unanswered support tickets rather than healthy activity.


## AI takeaways

**Thesis:** The model is fine. Frontier capability is a rented engine every competitor shares, and what a company can actually use is capped by its own context. A semantic layer with one human owner per metric and a context graph that connects contracts, opportunities, accounts and people close that gap without a data science team, and once they exist the agents are the easy part.

- **** — 
- **** — 
- **** — 
- **** — 
- **** — 
- **** — 

**Agent & automation ideas**

- An expansion-likelihood agent that answers only against written, owned definitions of customer, expansion and likely.
- A customer health score agent built on the semantic layer and context graph rather than a bespoke data science model.
- An identity-resolution agent that flags the same company under different names across CRM, marketing automation and billing.
- A forecast QA agent that detects line-item fan-out, such as a $10,000 deal reporting as $30,000.
- An engagement-quality agent that separates real buyer activity from unanswered support tickets in account touch-point counts.
- A converter that turns a plain-language definitions document into YAML for the semantic layer, with the definitions themselves written by named owners.


## Operations takeaways

### Revenue operations

- **.** 
- **.** 
- **.** 
- **.** 
- **.** 

### Pipeline & marketing ops

- **.** 
- **.** 
- **.** 
- **.** 

### Customer operations

- **.** 
- **.** 
- **.** 


## Metrics mentioned

| Value | Metric | Context |
| --- | --- | --- |
| $1M on a $20M plan | Value of one point of conversion | Moving conversion from 20% to 21% on a $20 million plan, with no new headcount, price change or extra leads. |
| About $250K within the year | Value of one day off the sales cycle | Anthony's estimate on the same $20 million plan. He says the conversion and cycle-time levers compound. |
| Flat for two years | Usable AI capability | What companies can actually use from AI, capped by their context, while frontier model capability climbs every quarter. |
| 4 people, 4 answers | Leadership definitions of customer | What Anthony has seen on one leadership team asked what a customer is. |
| $10,000 deal forecast as $30,000 | Line-item fan-out | What happens when products fan out across line items without a clean context graph, and nobody catches it. |
| 150 touch points | False engagement | An account that looks healthy on activity until the touch points turn out to be unanswered support tickets. |
| 40+ CSMs, hundreds of customers | Customer success team at Emailage | The team Anthony ran, where a trustworthy health score needed a full data science team. With both layers in place he puts it at days of work. |
| About 10 weeks | Implementation timeline | Anthony's realistic estimate for implementing the semantic layer and context graph, half of it leadership alignment rather than engineering. |
| 20 definitions, 20 owners, 20 sources | Monday starting list | The zero-budget first step: the definitions every reported number depends on, one name against each, and each number's source of truth. |


## Entities mentioned

- **LeanScale** (company) — Anthony Enrico's firm. Its semantic layer, the GTM brain, lives in GitHub so every model can connect cleanly. The episode's deck, worksheet and GTM agents are offered at leanscale.team/mcp. · https://www.leanscale.team/knowledge/company/leanscale/
- **Emailage** (company) — Where Anthony ran customer success with more than 40 CSMs around the world and hundreds of customers. A health score he could trust was effectively impossible there without a full data science team, his example of what the two layers now make a matter of days. · https://www.leanscale.team/knowledge/company/emailage/
- **Anthony Enrico** (person, host) — Co-founder of LeanScale and host of The LeanScale Podcast. · https://www.leanscale.team/knowledge/guest/anthony-enrico/
- **Claude** (tool, AI Assistant) — The AI assistant in the opening test: connected to Salesforce and Gong and asked which customers are likely to expand this quarter, it returns a confident, well-written answer that can't be verified.
- **Salesforce** (tool, CRM) — The CRM in the opening test. Anthony says the CRM is already most companies' first context graph and cleaning it up gives a solid version zero.
- **Gong** (tool, Revenue Intelligence) — Connected alongside Salesforce in the opening test, to show that access to data sources alone doesn't make the AI's answer trustworthy.
- **GitHub** (tool, Developer / Code Hosting) — Where LeanScale keeps its semantic layer, the GTM brain, so every model can connect to it cleanly.
- **Google Drive** (tool, File Storage / Collaboration) — Anthony's low-tooling alternative for the semantic layer: even a Google shared drive will do, because the one-owner rule matters more than the tooling. A Google Doc of definitions is also what he says AI can convert into YAML.


## FAQ

**Q: Why does AI give confident but unreliable answers to go-to-market questions?**

A: According to LeanScale co-founder Anthony Enrico, it is usually a data model problem rather than a model problem. A question like which customers are likely to expand this quarter depends on definitions most companies have never agreed on: whether pilots or accounts that have given notice count as customers, whether expansion means committed contract or usage, and what likely means. He has seen four leaders on one team give four different answers. Where definitions are missing, the AI infers the middle, so even with Salesforce and Gong connected the answer is confident and well written but can't be verified.

**Q: What is a semantic layer for go-to-market teams?**

A: A semantic layer is a company's business logic written down in plain English in one place AI can read: what win rate means, how bookings are counted, what qualified means. LeanScale calls its own the GTM brain and keeps it in GitHub so every model can connect cleanly, but Anthony Enrico says a Google shared drive will do. It needs no data science team or fine-tuned model. The rule he says matters more than the tooling is that every metric has exactly one human owner: whoever has to defend the number in the board meeting.

**Q: What is a context graph, and does a company need new software to build one?**

A: A context graph is the set of connections between a company's data: contracts linked to the right opportunities, linked to the right accounts, with the right people on them. Anthony Enrico says most companies don't need new software to start, because the CRM is already their first context graph and cleaning it up gives a solid version zero. Without those connections, products can fan out across line items so a $10,000 deal shows up as $30,000 in the forecast, and an account with 150 touch points can look healthy when the touch points are unanswered support tickets.

**Q: Who should own a GTM metric definition?**

A: Exactly one named person, not a team. Anthony Enrico's rule is that a definition owned by RevOps is owned by nobody, and within a quarter the company ends up with two versions of win rate and an argument about which is right. The owner should be whoever has to defend that number in the board meeting.

**Q: Should AI write a company's business logic and metric definitions?**

A: No. Anthony Enrico calls this a hard rule. Asked what a good renewal looks like, a model will ship standard industry practice because that is the fastest path to an answer, and unless the company is entirely generic that isn't its business. People should write the definitions in a plain document and let AI convert that document into a structured format such as YAML, which it does well.

**Q: What is the financial case for fixing the GTM data model before deploying AI?**

A: Anthony Enrico frames it as yield from the go-to-market machine a company has already paid for. On a $20 million plan, raising conversion one point, from 20% to 21%, is worth $1 million, and taking one day off the sales cycle brings in roughly another quarter million within the year. Neither needs new headcount, a price change or extra leads, and the two levers compound. He adds that usable AI capability has been flat for two years because it is capped by a company's context, not by the model.

**Q: How long does it take to build a semantic layer and context graph, and where should a team start?**

A: Anthony Enrico estimates about a ten-week project, and says half of it isn't engineering but getting leaders to agree on what words mean. The zero-budget start is three steps for Monday: write down the 20 definitions every reported number depends on, put one person's name against each, and write down where each number comes from, its source of truth.


## Timeline

- **00:00** — Ask your AI which customers will expand
- **00:26** — What's a customer? What's expansion? The definitions problem
- **01:03** — Everyone rents the same Formula One car
- **01:28** — Einstein on your payroll: Acme Corp, Acme Inc., Acme Limited
- **01:58** — Why bother: getting more yield out of the GTM machine
- **02:16** — One point of conversion is $1M on a $20M plan
- **02:43** — The model keeps climbing. Your usable context is flat.
- **03:02** — Layer one: the semantic layer (the GTM brain)
- **03:27** — The one-owner rule
- **03:50** — Layer two: the context graph
- **04:13** — Line-item fan-out and false engagement
- **04:31** — Never let the model author your business logic
- **04:55** — The health score that used to need a data science team
- **05:34** — A 10-week project, and three things to do Monday


## Related episodes

- **Ep. 122: Your AI Strategy Is a Data Model Problem** (Jake Toepel (LeanScale)) — The full 45-minute version of this argument, co-presented by Jake Toepel and Anthony at the Updata Partners CEO Summit, with the maturity curve, tokenomics and the ten-week roadmap in detail.
- **Ep. 98: AI-Native GTM: 3 Agent Plays and the Layer That Makes Them True** (Jake Toepel (LeanScale)) — The agent plays that sit on top of the layer this episode says to build first.
- **Ep. 126: The Agents Are the Easy Part** (Bastien Botella (Fountain)) — An operator's version of Anthony's conclusion: in a 55-agent marketing system, documenting and joining the data was the hard part and the agents were easy.
- **Ep. 119: The Agent Was Never the Hard Part** (Kushal Sharma (Circle)) — Another case that the agent is easy and the foundation underneath it decides whether it works.
- **Ep. 96: After the Series A: The Capital Clock** (Anthony Enrico (solo)) — Anthony's earlier solo episode, which introduces the GTM Brain as the instrumentation layer AI depends on.
- **Ep. 58: Why RevOps Shouldn't Have to Beg Engineering for Data** (Ghalib Suleiman (Polytomic)) — Getting CRM, billing and product data into one connected place, the context-graph work this episode describes. · https://www.leanscale.team/knowledge/podcast/ghalib-suleiman-empathy-data-revenue/
- **Ep. 28: Run CS Ops like a Pro: The 5 Things Every CS Operation Needs to Have** (Adrian Diaz (Findem)) — The customer success operations foundation behind the trustworthy health score Anthony couldn't build at Emailage. · https://www.leanscale.team/knowledge/podcast/adrian-diaz-cs-ops/


## Full transcript

_Machine-transcribed and not diarized; speaker attribution is inferred._  
_Transcript only, as a separate file: https://www.leanscale.team/knowledge/podcast/anthony-enrico-leanscale-ai-strategy-really-data-model-problem/transcript.md_

### 00:00 — Ask your AI which customers will expand

**[0:00]** Which of my customers are likely to expand this quarter?

**[0:04]** Go type that into cloud right now.

**[0:06]** You're connected to Salesforce, you're connected to Gong.

**[0:09]** It should be easy.

**[0:12]** And what comes back is confident, well-written,

**[0:15]** and you'll have no idea whether it's true or not.

**[0:19]** I don't think that's the model's fault.

**[0:21]** Your AI strategy is really a data model problem.

### 00:26 — What's a customer? What's expansion? The definitions problem

**[0:26]** Think about what it has to know

**[0:27]** to answer that seemingly simple question.

**[0:30]** What's a customer?

**[0:32]** Sounds like a silly question.

**[0:33]** Is it the field on the account?

**[0:35]** Do pilots count?

**[0:36]** Does a company that already gave notice

**[0:38]** still count as a customer?

**[0:39]** I've watched four people on one leadership team

**[0:42]** give four different answers.

**[0:45]** Then, what's an expansion?

**[0:48]** Are we talking about an increase in committed contract,

**[0:51]** more usage, and what does it mean to be likely to expand?

**[0:56]** AI doesn't know, so it does what it always does.

**[1:00]** It makes up the middle and infers as best as it can.

### 01:03 — Everyone rents the same Formula One car

**[1:03]** The Frontier Labs have built a Formula One engine.

**[1:07]** It's extraordinary, it gets better every single quarter,

**[1:10]** and you and your closest competitor

**[1:12]** are renting the exact same one.

**[1:14]** Nobody in your market has a horsepower problem.

**[1:18]** What you don't have is the map, the corner ahead,

**[1:22]** the line that won the last 10 races.

**[1:24]** Nobody sells you that, and without it,

**[1:26]** the car just reaches the same corner faster.

### 01:28 — Einstein on your payroll: Acme Corp, Acme Inc., Acme Limited

**[1:28]** Or to choose a different metaphor,

**[1:30]** put Einstein on the payroll,

**[1:32]** smartest hire you'll ever make, and he starts Monday.

**[1:36]** Ask him which deals are at risk.

**[1:38]** He won't be able to tell you.

**[1:39]** Not because he isn't clever,

**[1:41]** not because he isn't intelligent,

**[1:43]** because he's looking at Acme Corp. in the CRM,

**[1:45]** Acme Inc. in the marketing automation platform,

**[1:49]** Acme Limited in billing,

**[1:50]** with no idea if that's one company or four.

**[1:53]** Intelligence is not the bottleneck, access is.

### 01:58 — Why bother: getting more yield out of the GTM machine

**[1:58]** So why go through the trouble?

**[2:00]** Because all of it serves one thing,

**[2:02]** getting more yield out of the go-to-market machine

**[2:05]** you've already paid for.

**[2:06]** Most companies pour millions, if not tens of millions,

**[2:09]** into an organization whose whole job is producing revenue

**[2:13]** and have never measured what it yields.

### 02:16 — One point of conversion is $1M on a $20M plan

**[2:16]** Do the math on a $20 million plan.

**[2:19]** One point of conversion, 20% to 21%,

**[2:23]** is worth a million dollars.

**[2:25]** One day off the sales cycle allows you to cycle

**[2:28]** in another quarter million within that year.

**[2:31]** No new headcount, no price change, not one extra lead.

**[2:36]** Same exact machine, higher yield,

**[2:39]** and the two levers continue to compound together.

### 02:43 — The model keeps climbing. Your usable context is flat.

**[2:43]** And I think this is the part that tends to get missed.

**[2:47]** What the Frontier model can do climbs every single quarter.

**[2:50]** What you can actually use has been flat for two years

**[2:54]** because it's capped by your context, not by your model.

**[2:58]** Every single point of the gap is yours to close.

### 03:02 — Layer one: the semantic layer (the GTM brain)

**[3:02]** Two pieces to closing it.

**[3:03]** And neither one needs a data science team

**[3:06]** or a fine-tuned model to accomplish.

**[3:08]** First is the semantic layer.

**[3:10]** We call ours the GTM brain.

**[3:12]** It's your business logic written down in plain English

**[3:15]** in one place your AI can read.

**[3:17]** What win rate means here, how we count bookings,

**[3:20]** what qualified means.

**[3:22]** Ours lives in GitHub.

**[3:23]** Every model can connect cleanly,

**[3:25]** but even a Google share drive will do.

### 03:27 — The one-owner rule

**[3:27]** One rule matters more than the tooling.

**[3:29]** Every metric gets exactly one owner, a human, not a team.

**[3:35]** A definition owned by RevOps is owned by nobody.

**[3:40]** Inside a quarter, you've got two versions of win rate

**[3:43]** and an argument about which is right.

**[3:46]** Who's the owner?

**[3:47]** Whoever has to defend the number in the board meeting.

### 03:50 — Layer two: the context graph

**[3:50]** Second is the context graph, which is a fancy way of saying

**[3:54]** connect the dots between your data.

**[3:57]** Contracts to the right opportunities,

**[3:59]** to the right accounts, with the right people on them.

**[4:01]** For almost everybody watching,

**[4:03]** your CRM is actually your first context graph.

**[4:08]** It's already there.

**[4:09]** So if you go clean it up, you have a great V zero.

### 04:13 — Line-item fan-out and false engagement

**[4:13]** Two things can happen when you don't.

**[4:16]** Products fan out across line items.

**[4:18]** So a $10,000 deal shows up as $30,000 forecast.

**[4:22]** Nobody catches it.

**[4:23]** Or an account has 150 touch points and looks fantastic

**[4:27]** until you look closer and they're unanswered support tickets.

### 04:31 — Never let the model author your business logic

**[4:31]** One hard rule, never let the model

**[4:33]** author your business logic.

**[4:35]** Ask it what a good renewal looks like

**[4:37]** and it'll grab standard industry practice and ship it

**[4:40]** because that's the fastest path from A to B.

**[4:43]** Unless you're running the most vanilla lemonade stand

**[4:46]** SaaS company alive, that isn't your business.

**[4:49]** You write it down, let AI turn the Google Doc into YAML,

**[4:53]** that part it's great at.

### 04:55 — The health score that used to need a data science team

**[4:55]** What really surprised me is that once the data model exists,

**[4:58]** the agents are the easy part.

**[5:02]** When I ran customer success at a company called Emailage,

**[5:04]** I had over 40 customer success managers around the world

**[5:09]** and hundreds of customers.

**[5:11]** And the one thing every head of CS wants,

**[5:14]** a health score you can actually trust

**[5:16]** was effectively impossible.

**[5:18]** You would have needed at the time

**[5:20]** a complete data science team to even attempt answering them.

**[5:24]** With the two layers underneath, it's days of work.

**[5:27]** Start with the agents instead

**[5:29]** and you'll reverse engineer a private little data model

**[5:31]** for every single agent you build.

### 05:34 — A 10-week project, and three things to do Monday

**[5:34]** Realistically, if you wanna implement this in your company,

**[5:36]** it's a 10 week project and half of it isn't even engineering.

**[5:40]** It's getting your leaders to agree on what words mean.

**[5:43]** That's the hard part, as it always has been,

**[5:46]** but especially now.

**[5:47]** Three things you can do on Monday that requires zero budget.

**[5:51]** Write down the 20 definitions,

**[5:52]** every number you're reporting depends on,

**[5:55]** put one name against each,

**[5:57]** write down where each number comes from,

**[5:59]** their source of truth.

**[6:01]** For everything I'm talking about today,

**[6:02]** the deck, the worksheet, tens of agents,

**[6:05]** it's all at leanscale.team/mcp.

**[6:09]** Your model is fine, go build the map.


---

_LeanScale Knowledge Hub. Free to quote and cite with attribution to The LeanScale Podcast (https://www.leanscale.team)._