Customer World Models: How to Build the AI Layer Anthropic Can't Ship

Anthropic can ship your features as a plugin. They can replicate your workflows, your compliance checks, your reporting dashboards.

What they can’t replicate is your understanding of why 5,000 customers use those features differently.

That understanding — structured, per-user, compounding — is what we call the customer world model. And it’s the only AI layer the frontier labs can’t ship.

What is a customer world model?

It’s how a company understands and organizes its model of who the customer is and why they do what they do.

Not just analytics. Not just persona cards. A continuously updated, per-user understanding that connects behavior to motivation to revenue.

Traditional product analytics tell you what happened. Session counts, DAU/MAU, conversion funnels — all aggregate, all backward-looking. They can tell you that 31% of sessions generated zero revenue. They can’t tell you why, or predict which users are about to churn.

A customer world model tells you why. And it gets smarter every interaction.

The self-model primitive

At the core is a self-model: a structured representation of an individual user that tracks beliefs, preferences, context, and behavior across every interaction.

Session 1: the model knows nothing. Generic experience.

Session 10: the model knows the user prefers evening interactions, responds well to direct communication, and has tried three different product features.

Session 50: the model predicts which actions this user will take, what content will resonate, and when they’re at risk of churning — based on 50 sessions of accumulated understanding.

The model gets smarter every session. The map stays current with the territory — automatically.

This is fundamentally different from recommendation engines that optimize for engagement metrics, or personalization systems that rely on stated preferences from onboarding surveys.

Self-models track revealed preferences — what users actually do, not what they say they’ll do. And they compound.

B2C proof: 60% revenue growth

We built the customer world model for The Relationship Psychics — a session-based app with 150+ daily readings, $30K/month in Meta ad spend, and revenue through Stripe.

The problem: the founder checked Stripe every morning and couldn’t answer why Tuesday’s revenue was 40% lower than Monday’s. Marketing data lived in Meta. Product data lived in the app. Revenue data lived in Stripe. None of them talked to each other.

We connected the three layers — acquisition, engagement, and revenue — through per-user self-models. For the first time, the founder could see the full lifecycle of every user, session by session, dollar by dollar.

What the data revealed:

• 31% of sessions generated zero revenue despite costing ad spend to acquire
• One user segment was worth 4x the average but was getting the generic experience
• The month 3 churn cliff was a fixable UX issue, not a fundamental product problem

We fixed what the data revealed. $50K+ in annual revenue leaks recovered. 60% revenue growth in 60 days. Same traffic, same product.

The customer world model didn’t just fix problems once — it became the founder’s operating system. Every interaction makes it smarter.

B2B application: the same substrate, longer loops

In B2C, the value chain between user behavior and revenue is tight and fast-cycling. You see results in days.

In B2B enterprise software, the data patterns take shape across a longer timeline. But the substrate is the same:

Product adoption patterns — how 5,000+ customers adopt HR software differently. Which features predict expansion? Which workflows correlate with churn? The patterns are there — they’re just not structured or compounding.

At Workday, I built a framework to answer exactly this question. It encoded how enterprise customers adopt software — and that understanding helped Workday grow from $2B to $4B. The framework was valuable because it captured domain-specific learning that couldn’t be replicated by a competitor overnight.

Customer success signals — which behaviors predict renewal vs. churn? Not aggregate metrics — per-customer understanding that compounds across every support ticket, every product session, every QBR.

Expansion triggers — what pattern of usage predicts when a customer is ready for a new SKU? Self-models can capture this across thousands of accounts simultaneously, learning patterns that no CSM could hold in their head.

How to start

Most companies don’t have a customer world model. They have fragments:

• Marketing has campaign performance data
• Product has session analytics
• Sales has deal notes and CRM records
• Customer success has health scores and NPS
• Finance has revenue and churn numbers

Nobody has the full picture. The customer world model connects these fragments into a compounding, per-user understanding.

Sprint Zero diagnoses the gaps. In 4 weeks, we map what you know about each user, what you’re missing, and where the learning is decaying.

For B2C products, the Lifecycle Lens connects marketing → product → revenue immediately. For B2B, we instrument the longer-loop patterns that predict adoption, expansion, and churn.

Start with Sprint Zero →

---

This post is part of the SaaSpocalypse Survival Guide series.