A company brain should remember, cite and forget.

The legion is a useful metaphor: large groups need ways to move information, divide responsibility and coordinate action. Companies inherited layers and reporting lines because human attention is limited and local knowledge does not travel by itself.

That does not make hierarchy a mistake. A manager may coach people, resolve conflict, protect standards and make decisions under uncertainty. But some layers also spend a great deal of time collecting updates, translating plans and forwarding context. AI can reduce that routing work. It does not remove the need for authority, care or judgement.

A copilot usually helps one person produce or retrieve something faster. Useful, but narrow. It leaves the organisation's memory scattered across inboxes, documents, ticket systems and people. The next person still has to find the same evidence and reconstruct why a decision was made.

A more consequential use of AI is shared operating memory: a governed system that can retrieve what the organisation knows, show where it came from and carry an approved decision into later work. That possibility grows as intelligence becomes a thing you rent, but the rented model is only one component. The controls around the memory determine whether it deserves trust.

Knowledge lives in agreements, policies, tickets, notes and experienced colleagues. A company brain does not ingest everything and become omniscient. It creates a governed path from selected evidence to a usable answer.

Each retained item needs provenance: its source, author, date and version. It needs a named owner, a review date, an access rule and a retention or deletion rule. A summary must point back to the evidence it compresses. Without those controls, a queryable corpus is not organisational memory. It is an unlabelled pile that can produce confident answers from stale or forbidden material.

The operating unit can become a controlled loop. A new event arrives, the system retrieves permitted evidence, produces a cited answer and checks it against policy. It acts only within a defined permission boundary. Otherwise it escalates to the person who owns the outcome.

Review triggers matter as much as retrieval. Required evidence is missing, policies disagree, an evaluation fails or the action crosses a predefined financial threshold or affects employment, safety or a legal commitment. Any of those conditions should stop automation and ask for human judgement. Repeated failures should trigger workflow review, not become acceptable merely because the same output recurs.

A useful company brain does not remove people from the loop. It removes repeated reconstruction while keeping a named person at the authority boundary.

Coordination tasks can shrink when status, rationale and evidence are available without another meeting. That may change management roles, but coaching, prioritisation, negotiation and responsibility do not become database fields.

Every consequential workflow needs a responsible human who can approve, refuse, correct and retire it. Builders maintain the mechanism; leaders decide what the organisation is willing to do. This is the redraw in when shipping became table stakes: automation increases the reach of each decision.

Consider a hypothetical subscription company handling a cancellation request. Its public policy says customers have 30 days. The signed customer agreement says 60. A recent internal message suggests making an exception, but does not say who approved it.

A weak system blends those sources and returns one confident deadline. A governed memory shows all three, with dates and owners, recognises the conflict and treats the signed agreement as controlling for this account. Because money and a contractual commitment are involved, it sends the proposed answer to the named support owner. That person approves the response, records why, and assigns someone to reconcile the public policy with the contract template.

The approved rationale becomes a versioned record with a review trigger. Access stays limited to roles that need the account. Raw material is deleted on schedule, including indexed and cached copies, while required records follow the applicable policy. Memory includes forgetting on purpose.

Human authority is not a temporary patch for tasks the model cannot yet perform. People decide the purpose of the system, whose interests count, which evidence is legitimate and which actions may never be delegated. A model can prepare a recommendation for a hiring, safety or contractual decision. It should not become the accountable party merely because it can produce fluent reasons.

I want decisions to survive scrutiny and corrections to improve the next attempt. Fatherhood sharpens the privacy test for me: cheap storage is not a right to collect. Stewardship includes deciding what the organisation has no need or authority to remember.