Bringing People Along in Practice

The first lesson was about knowing where a team stands. This one is about the people side: why individuals hesitate, how to talk honestly about work and roles, and how to keep the tool useful once the launch energy fades.

Champions and peer learning

People learn most readily from a colleague who does similar work and can show a real example. A champion is that colleague: someone respected in the team, with protected time, who demonstrates how they use the tool on live tasks and is willing to share what went wrong. Champions work badly when they are chosen for seniority, treated as unpaid support desks or given targets to convert their peers. They work well when their line manager has agreed the time and when they pair with a colleague from a different role, so that examples are not dismissed as irrelevant.

What resistance usually signals

When people ignore a tool, leaders are tempted to read it as fear of change. In practice, resistance usually signals something specific: the tool adds steps that someone has to repair later, errors in the output have not been acknowledged, or nobody has said how roles will be affected. Treat resistance as information. Collect specific examples, check whether the complaint is accurate, and change what is wrong. Some concern will remain, and that is normal. But a leader who skips the investigation and goes straight to reassurance or deadlines loses the one source of honest feedback on the workflow.

Being honest about jobs

Staff will ask whether the tool puts jobs at risk, and a vague answer is heard as a yes. Do not promise that nobody will lose work if you cannot guarantee it, and do not predict reductions you have not decided. Separate what is decided from what is not, say which is which, and tell people when and how they will hear about changes that touch their roles. Involve them before decisions are made, not after. An honest, uncomfortable answer protects trust better than a comforting one that has to be taken back.

Automation bias and safety to say it was wrong

As people grow used to a tool, they check less. This is automation bias, the tendency to over-rely on a system's output. The EU AI Act lists awareness of it among the abilities that people overseeing a high-risk system should have (Art. 14(4)(b)). In any workplace, the remedy is practical: train reviewers on the tool's known limits, have them form a view before seeing the recommendation on difficult cases, ask them to record what they checked, and sample reviews to see whether challenge still happens. Do not rely on a reminder to be careful.

None of this works if people fear blame for reporting errors. Make it safe to say the tool got it wrong: thank the person, log the example, and show what changed. Be cautious when error reports fall to near zero after leadership praises the tool. That may mean fewer errors, or it may mean people have stopped speaking up. Test with sampled outputs before deciding which.

Feedback loops that change something

A suggestion box that produces comments but no change teaches staff that feedback is pointless. A working loop has an owner, a rule for choosing what to act on, a change to the workflow, and a report back that says what changed and when. Prioritise by effect on error risk and rework, and confirm with real examples before changing the process.

Measure honestly, and avoid mandating the unproven

Count repeat use on real tasks and the change in outcomes, not logins or training attendance. If a tool is mandated before its value is shown, usage figures rise while genuine use stays shallow, and problems are hidden behind compliance. Prefer a bounded group with room to opt out until the evidence supports wider use.

Finally, when use has become stable, consistent and well understood, it should become normal work. Keep light checks, a named owner and an exception route, and aim monitoring at exceptions and changes rather than carrying pilot-style reporting forever.

Worked example: Pennant Freight

Pennant Freight, an invented haulage firm, used an AI tool to summarise customs documents. Reviewers began approving summaries in seconds, and an audit found two flawed summaries had passed unchallenged. Error reports had also dried up since the managing director praised the tool at an all-staff meeting. Operations lead Idris did not issue a reminder. He trained reviewers on the types of document where the tool struggles, asked them to note what they had checked, introduced sampled second reviews, and invited reviewers to a session where missed errors were discussed without blame. Within a quarter the sampled reviews showed the checks were being done properly. Idris then moved routine customs summaries to normal-work monitoring and kept a closer watch on unusual document types.

As of 24 September 2026.