Ontology is all you need

Coding agents work. Not demo-work. Actually work. Engineers hand off tickets, get back working code, merge it, move on. Support agents are close behind, resolving real tickets without a human in the loop.

Everything else: demos that dazzle, pilots that fizzle, production that requires a human to redo the work. Same models, same capabilities. So what do coding and support have that nothing else does?

Most people say training data. GitHub is massive, support knowledge bases are structured, models got good at these things because they saw a lot of them. There's truth in that. But it's not the real answer.

The real answer is context. A codebase is a self-contained world. Point an agent at a repo and every fact it needs is already there. Types tell it what things are. Tests tell it what things should do. Imports tell it how things connect. Git history tells it why things changed. Without anyone intending it, programmers built a complete ontology: a machine-readable map of a small universe, its entities, their relationships, and the rules governing them.

But it gets better. A codebase also comes with a labeled training set. A Jira ticket is a prompt: here's what I want, here's the acceptance criteria. A merged pull request is the approved output: here's the correct response, reviewed by a human. Thousands of these pairs, generated organically over years, sitting right there in the issue tracker and git log.

The same labeled pairs exist in every profession. An analyst's inbox is full of "MD asked for a model, here's the version that went to the client." A lawyer's email history contains "client asked a question, partner approved this answer." The training data is everywhere. It's just not anywhere a machine can reach it. Yet.

Context is what makes agents smart regardless of the model underneath. Give an agent complete context and a mediocre model will outperform a frontier model running blind. This is what the benchmarks miss. They test capability in a vacuum. Production tests capability with whatever context you can scrounge together.

In the enterprise, context is a disaster.

We spent the last fifteen years buying SaaS. Hundreds of tools, each best-in-class at one thing. Gong for calls. HubSpot for campaigns. Salesforce for pipeline. Zendesk for tickets. Amplitude for usage. Jira for feature requests. Stripe for payments. Google Analytics for acquisition. Notion for strategy docs.

Then there's the stuff that doesn't fit neatly into any tool. The Slack thread where your team debated whether to give a customer a discount. The email the CEO sent personally after a rough QBR. The offhand comment on a Zoom that changed the product roadmap. All of it is customer data. Almost none of it is reachable.

So what would it take to give an enterprise agent the same complete picture a codebase gives a coding agent?

Every piece of data in a business connects back to one of two things: a customer or an employee. That's it. Pipeline, revenue, support tickets, campaign performance, churn analysis: customer. Payroll, provisioning, compliance, headcount: employee. Every SaaS tool in the stack is a different view of data that ultimately joins on one of these two keys.

At a high level, a business does two things: get new customers and serve existing ones. There is no third thing. Finance exists to measure it. Product exists to enable it. Marketing exists to drive it. Strategy exists to direct it. Every function, every workflow, every decision traces back to the customer.

Now here's the problem. Everyone assumed Salesforce was the customer ontology. The system of record. The source of truth.

It isn't. 80% of customer data lives outside the CRM. Salesforce has deal stages and whatever a rep logs after a call. The actual customer relationship, what they said on the discovery call, what frustrated them in onboarding, which feature they begged for, why they almost churned, lives in Gong, Zendesk, Slack, email, product analytics. Salesforce is a scoreboard. It tells you the score. It can't tell you how the game was played or what to do next.

A scoreboard is not an ontology.

But the raw material for a real customer ontology already exists. The calls are recorded. The emails are logged. The support tickets are timestamped. The product usage is streaming. The data is in the cloud, scattered across dozens of systems, waiting to be unified.

Whoever gathers all of the data about acquiring and serving customers into a single, machine-readable graph builds the codebase equivalent for the rest of the business. That's the platform that automates white-collar work. Not because of a better model. Because of a better ontology.

Not the model layer. Not the agent layer. The meaning layer.

Some people call this a revenue ontology. We just call it the future of enterprise software.