AI Governance
Why Long-Running AI Work Needs Governance
As AI becomes an active layer in sustained programme and knowledge work, continuity, authority, and accountability can no longer depend on conversational memory alone.
- Author
- Towseef Ahmed
- Published
- Draft — not yet published
- Reading time
- 6 min
This piece is a draft and is not yet published. It is excluded from search indexing.
Most discussion of AI governance still assumes a short interaction: a question is asked, an answer is produced, a person decides what to do with it. That model is already out of date in the environments where AI is doing the most consequential work.
In sustained programme and knowledge work, AI does not appear once. It appears continuously — across months, across tools, across teams, and across changes of personnel. Analysis produced in one context is carried into another. A recommendation made under one set of assumptions is acted upon after those assumptions have quietly changed.
The failure is separation, not error
The risk in long-running AI work is rarely a single wrong answer. It is the gradual separation of things that were once connected: the analysis from the context that produced it, the decision from the evidence behind it, the approval from the scope it actually covered, and the execution from the authority that legitimised it.
Each of those links can break silently. Nothing announces that a decision is now resting on stale context, or that an automated step has drifted beyond what anyone approved.
Continuity cannot depend on memory
Conversational memory is not a governance mechanism. It is not addressable, not verifiable, and not accountable to anyone. When work spans quarters rather than minutes, continuity has to be a property of the record, not of the session.
That means decisions need to be traceable to identifiable context, held by an identifiable person, at an identifiable point in time — and that automated execution needs a boundary it cannot widen on its own.
Authority has to stay explicit
The principle worth defending is straightforward: AI may analyse, draft, recommend and execute within approved limits. It should not create its own authority, approve its own recommendation, expand its own execution boundary, or silently replace a human decision.
Stated that way it sounds obvious. In practice it requires deliberate design, because the drift happens without anyone choosing it.
Where this leads
Governance of this kind is not an obstacle to AI-enabled work. It is what allows organisations to rely on that work over a long enough period for it to matter — and to explain, afterwards, how a decision was actually reached.
That question is the subject of ongoing framework development, and it is where much of my current work is focused.