I have been banging on about context for a while. First, in relation to enterprise AI adoption. More recently, in relation to strategy and what I call context rot: the gradual loss of the institutional knowledge, history, relationships, assumptions, and judgment that explain why an organisation works the way it does.
Data ontology
Now Databricks has added another piece to the puzzle: data ontology. Its article, Data Ontology defined: The context layer your AI agents are missing, makes a deceptively simple point. Companies have spent decades organising data on the assumption that a knowledgeable human would eventually interpret it. AI agents break that assumption.
Which revenue?
Ask five people in the same company what “revenue” means, and you may get five different answers. Gross revenue. Net revenue. Recognised revenue. Revenue for a particular division. Revenue, according to the definition Finance uses for one specific report. Humans navigate this ambiguity almost unconsciously. Put an AI agent into that environment, and all that invisible knowledge suddenly becomes a problem. The agent can access the data, but access does not mean understanding. Give it more data, and you may simply give it more opportunities to be confidently wrong.
Context is becoming infrastructure
Hence ontology. A schema describes how information is structured. An ontology describes what that information means inside the organisation: definitions, relationships, authoritative sources, calculations, expertise, and rules. For years, organisations treated context as soft stuff. Data was infrastructure. Context wasn’t. Agentic AI changes that. If machines are going to make decisions, initiate workflows, and take actions rather than simply answer questions, organisational context has to become machine-readable, governed, and accessible. Your organisation needs a digital representation not just of what it knows, but of what things mean. Context becomes part of the technology stack.
But ontology does not solve the whole problem
There is a trap here. Once we discover that context matters, the temptation is to believe we can capture all of it. We can’t. An ontology can tell an AI what revenue means today. It can identify the authoritative source, document the calculation, and establish who is allowed to see it. That is enormously valuable. But strategy operates one level higher. AI can process extraordinary quantities of information, but strategic judgement depends on history, constraints, relationships, tacit knowledge, and the ability to recognise when the existing frame itself needs to change. The danger is not just missing context. It is frozen context. The ontology of yesterday instead of the ontology of tomorrow.
Context rot becomes an AI problem
This also makes context rot far more serious. When experienced people leave, when definitions diverge between departments, when decisions are made without recording the reasoning behind them, or when institutional knowledge stays trapped in people’s heads, organisations are not simply losing knowledge. They are degrading the intelligence layer that their future AI systems will depend on.
The context tax
Databricks calls the current inefficiency a “context tax”: analysts repeatedly rediscovering definitions, reconciling contradictory reports, and explaining business logic that should already be available. AI exposes that problem because agents cannot quietly fill the gaps with years of organisational experience. What looked like a knowledge-management problem is becoming an AI infrastructure problem.
Context mapping
- How much of the context required to run this organisation actually exists outside people’s heads?
- Do your teams agree on the meaning of critical concepts?
- Do you know which information is authoritative?
- Are important decisions accompanied by the reasoning behind them?
- Can your systems distinguish formal policy from habitual practice?
- Can you identify expertise, relationships, and dependencies?
- Can that context change as quickly as your organisation does?
Understanding the terrain
The next phase of enterprise AI is not just about making machines more intelligent. It is about making sure they understand the environment in which that intelligence is applied. The model may be artificial. The context is yours.
