Designing the Handshake Between People and AI.

01The rubber-stamp problem

Oversight is easy to require and hard to make real. Give a reviewer a queue of a hundred proposed actions, each one a wall of text, and one of two things happens: the queue backs up until the business switches the approval step off, or the reviewer starts approving without reading. Either way, the control written in the risk register no longer exists in practice.

Good review design starts from the reviewer's minute, not the agent's output. What does this person need in front of them to decide in seconds — the proposed action, the two facts that justify it, what changes if they approve, and what the agent was unsure about? Everything else is one click away rather than in the way.

02Handing work back without losing the thread

When an agent reaches the edge of what it is allowed or able to do, it stops and hands over. Too often the person receiving the work starts again from nothing: the context is buried in logs, the reasoning is gone, and the customer has to explain themselves a second time.

We design an escalation as a case file, not an alert — what was asked, what the agent already did, what it found, where it got stuck and what it recommends, laid out so a person can pick the work up mid-stride. The same thinking runs in reverse: when a person hands a task to an agent, they should see what it intends to do before it does it, and be able to stop it once it starts.

The measure of an agent is not what it does alone. It is how well it hands the work back.

03Designing for the people who are not in the room

AI systems have audiences their builders rarely meet: the customer reading a drafted reply, the auditor following a decision trail a year later, the new joiner who never saw the old process, the colleague using a screen reader or working in a second language. Each of them experiences the AI through an interface somebody chose for them.

Human-centred design brings them in early — into research, prototype testing and the accessibility and language requirements — instead of discovering them through a complaint. It is also what separates responsible AI on paper from responsible AI in use.

• 01Every automated action is explained in words the affected person understands.
• 02Every decision trail can be read by someone who was not there.
• 03Accessibility and multilingual use are designed in, not retrofitted.
• 04Everyone the system affects has a route to a human.

04People and AI, on the same team

The organisations that get the most from AI are not the ones with the cleverest models. They are the ones where people understand what the AI is doing, feel in control of it, and find their own work genuinely better for it. That feeling is not an accident of culture. It is the sum of hundreds of small, deliberate decisions about how intelligence is presented, questioned, corrected and trusted.

Those decisions are the work our experience designers do beside our engineers on every AI programme — and it is the part of the work we are proudest of, because it is the part people feel every day.