Legal capability.
Organisational purpose.
All insights

Data sovereignty & agent use

Keeping people in control of AI agents

Define what an agent may access and do, keep a record of its actions, and give people practical ways to intervene.

Carry the boundaries with the knowledge.

An agent preparing a group briefing may encounter information that belongs to one entity or is restricted to a particular purpose. The briefing itself can reveal that information even if the original document remains protected. Control must therefore extend to derived knowledge, search results and activity records.

Agree permitted purposes, sources, recipients and actions. Understand the provider’s contractual and technical limits, including retention and export. Define what an authorised user or agent can continue to use when a provider or working arrangement changes.

Observe activity and test intervention.

A plausible explanation of an agent’s work is different from a record of the sources it accessed and actions it performed. The responsible people need enough evidence to investigate a material departure and identify which outputs or decisions may be affected.

Test a failure case: an agent attempts an action outside its remit, loses a source or relies on stale information. Establish what is detected, who is alerted and how activity can be paused or access revoked. A monitoring gap should remain visible. A quiet interface alone does not establish that the work is sound.

Make attention effective.

People should shape the inquiry as well as review its results. They set priorities, challenge assumptions, recognise significance and redirect the work. Agents can investigate within their remit and bring an unexpected issue forward with its supporting context.

Give a meaningful exception an owner, a decision or next action, and a practical escalation route. Match intervention to consequence and uncertainty. Measure the effort spent investigating, correcting and coordinating the work alongside its usefulness. Routine authorised steps should not require avoidable approval clicks.