Adaptive AI Architecture

A capable model doesn’t settle which decisions it may take. Someone must decide what counts as evidence and who answers for the result.

As the work changes, the design may need to change too. A system that fits one task may be unsuitable for another.

The system around a model determines whether an institution can rely on it.

An institution needs to be able to inspect, challenge and withdraw delegated authority. Research must test whether the whole system supports those decisions, rather than measuring the model alone.

For executivesBefore delegating a decision to an AI system, a board should be able to say what counts as evidence, who can overrule it and how the authority is withdrawn.

Delegating an assessment changes the design of the institution.

Consider a bank using AI to assess a proposed lending-policy change. The quality of the recommendation depends on more than the model: what counts as evidence, how conflicting interpretations are handled, where human judgement sits and which authority can approve a change.

Evaluate alternative architectures against the same institutional decision and explicit acceptance criteria. The question is whether the whole arrangement supports reliable judgement at an acceptable cost, with authority and accountability preserved.