In short: What has shipped as an "enterprise brain" is the context layer. It is the easy two-fifths of the idea, and it is the part being paid for. The missing three-fifths share one root cause: nothing in the stack writes an outcome back to a decision.
Key points
- The full promise has five parts: knowledge, memory, meaning, governance and judgment.
- Data owners built knowledge, software incumbents built governance, and consultancies built a narrow form of meaning.
- Judgment can only be evidenced later, when an outcome arrives and is connected back to the decision.
- Outcomes live in other systems, arrive slowly and are awkward because they name who decided.
- The cheapest fix is one decision class with expected outcome, check date and owner on every record.
Take the full promise apart and it has five components. Knowledge: the documents, data and definitions the organisation holds. Memory: what happened, across sessions and people. Meaning: a semantic layer that makes terms consistent across agents. Governance: permissions, identities and audit. Judgment: how the organisation decides, and whether it decided well.
In 2026 the market built the first and the fourth with real money behind them, built the third in a few places, and largely skipped the second and fifth.
Data owners built knowledge. The major financial-data incumbents adopted the same agent connection protocol in quick succession, and the ones that disclose numbers are monetising distribution of trusted data into other people's AI. One reported that more than 10% of its growth came from AI products that were close to zero a year earlier. Another reported more than 500 clients on its agent-ready API, up 70% in a quarter.
Software incumbents built governance. Microsoft reported nearly 40 million agents registered in its control plane within two months of launch. ServiceNow, Workday, Salesforce and Google each want to be the registry other vendors' agents must check in with. The fight is about who holds agent identity.
Consultancies built a narrow form of meaning: curated retrieval, maintained by large teams of human data stewards, with an expert-finder that routes the user to a person.
None of these is a criticism. Context is genuinely hard. Entity resolution across hundreds of millions of companies, entitlements that follow every query, decades of history, definitions that actually agree: that is a moat, and it took decades to dig.
But it is a moat around the inputs to a decision, not around the decision.
Why the last two pieces keep getting skipped
Memory and judgment share a property the other three lack. They depend on time running forward after the system has acted. Knowledge can be loaded today. Governance can be configured today. Judgment can only be evidenced later, when the outcome arrives, and only if something connects that outcome back to the call that preceded it.
Three things make that connection rare.
Outcomes live in other systems. The credit decision is in one platform; the default shows up in another, eighteen months later, owned by a different team.
Outcomes are slow. A pricing call is graded in two quarters, a partnership in two years. Most programme roadmaps are planned in sprints.
Outcomes are awkward. A record that links a decision to a bad result names the person who made it. Systems that would surface that link meet quiet resistance.
So the industry ships the parts that can be demonstrated in a pilot, and labels the whole thing a brain.
A quick test for your own programme
Open your AI roadmap and classify every item into one of the five components. Be strict: a connector is knowledge, a permissions model is governance, a vector index is memory only if it spans time, and nothing is judgment unless it records a prediction and later compares it with a result.
A typical roadmap, run through this test honestly, comes out heavy on knowledge and governance, light on meaning, and empty on judgment. That is a fine portfolio for a context layer. It is the wrong portfolio for a programme whose sponsor was promised compounding.
The cheapest correction is not a new platform. It is one decision class with outcome write-back: a single field on each record that names what should happen if the decision was right, a date to check, and someone accountable for checking. Start there and the other components acquire a purpose.
In practice
- Classify your roadmap against the five components and show the split to your sponsor.
- Pick one decision class with outcomes inside two quarters. Add expected outcome, check date and owner to every record.
- Rename what you are building if it is a context layer. Promising a brain and delivering a catalogue costs credibility you will need later.
What would prove this wrong. A horizontal product that, without per-client outcome plumbing, measurably improves decisions across unrelated decision classes. If one ships with independent evidence, the judgment layer is a product problem, not an organisational one.
Common questions
What is an enterprise brain?
The article breaks the promise into five components: knowledge, memory, meaning, governance and judgment. What has actually shipped under the name is mostly the context layer.
Why don't enterprise AI systems learn from their decisions?
Nothing in the stack writes outcomes back to decisions. Outcomes live in other systems, arrive slowly, and are awkward because they name the person who decided.
How can I tell if my AI roadmap is building judgment or just context?
Classify every item into the five components. Nothing counts as judgment unless it records a prediction and later compares it with a result.
What is the first step to add judgment to an enterprise AI programme?
Pick one decision class with outcomes inside two quarters and add an expected outcome, a check date and an accountable owner to every record.