An agent with documents and tools can still misunderstand the business it is operating inside.
A pile of PDFs, wiki pages, tickets, and dashboards is not a business model. It is raw material. If the agent does not know which customer definition counts, which source wins, how fresh a fact must be, or what it is allowed to remember, it can produce fluent nonsense with operational consequences.
The failure often looks reasonable. The agent summarizes a renewal using an old account note. It drafts a customer reply with an internal risk label. It recommends an escalation based on a product event that stopped being canonical months ago. Nothing explodes. The work is just wrong in a way that feels plausible.
The AI context layer is the governed business context agents can safely use for a task. It sits close to the semantic layer and systems of record, but it has its own job: package current, permissioned, source-aware business context for agents before they read, decide, write, recommend, or act.
That means context should include objects, relationships, states, source priority, freshness rules, permissions, identity, memory limits, and escalation paths. More documents are not the answer if the agent cannot tell which ones are current or authoritative.
This lane is separate from the control plane and harness engineering. Runtime controls matter. Tool orchestration matters. But neither solves the problem of business meaning.
An agent needs a model of the company small enough to use and strict enough to prevent improvisation. Otherwise it is just autocomplete with access.
This is part 1 of 10 in The AI Context Layer.
