Agent recommends. Agent acts.

Here’s a question we ask before any AI agent runs unattended: if it gets something wrong at 2am, who notices, and what can they do about it?

Letting an agent act, not just advise, is a decision about authority. It deserves the same care as handing authority to a person.

Before going live, we want 4 answers. Who approved this? Within what boundaries? On whose behalf? What happens at the edges?

We'd start with one well-defined task and real evidence from typical cases. Then we'd test the exit: can the team pause it, trace it and put things right, even when an action can’t be reversed?

Naming a human owner isn’t enough. They need the information, time and authority to intervene. A button marked “approve” isn’t oversight.

Human-first AI is an organisational question before it’s a technical one. People set the purpose, choose the risk and keep responsibility for what they delegate. More activity is easy to measure. Better outcomes for clients are what matter. Earn autonomy with evidence, clear permissions and a recovery process you’ve genuinely tested.