Listen to what operations staff ask when an AI agent arrives: most questions are about blame, not redundancy. If I approve its draft and it is wrong, is that on me? If it acts under my name, who answers? Resistance follows when those questions go unanswered, and clear bounds and records answer most of them.
What does resistance actually look like?
Rarely open refusal. More often, quiet behaviours that protect the person and undermine the agent.
- Redoing the work. Reviewers ignore the draft and rebuild each case themselves, so nothing they approve can be pinned on the agent's reasoning.
- Refusing by default. Anything uncertain is rejected, because a refusal is safer for the reviewer than an approval that might go wrong.
- Escalating everything. Decisions are pushed upwards so someone more senior carries the risk.
- Keeping parallel records. Staff keep their own notes of what the agent did and what they did, in case they need to defend themselves later.
- Slow adoption. The agent is used for the easiest cases only, and its scope never grows.
Each of these is a rational response to unclear accountability. Each makes the agent look worse than it is, and each costs the time the agent was meant to save.
Why is accountability the real fear?
Because people in regulated operations are used to answering for decisions. They know complaints, audits and incident reviews will look for a name. Before the agent, the name was theirs and so was the decision; the two matched. After the agent, the decision may come from the agent while the name on the record is still theirs.
That mismatch is the core of the fear. Nobody wants to be accountable for a decision they did not make, could not check properly, or were pressured to approve. If the design of the agent leaves that possibility open, resistance is the sensible response.
Job worries are real too, and should not be dismissed. Our piece on whether AI agents will replace jobs in operations addresses them directly. But in many teams, the accountability question comes first, because it affects people from the first day, whereas job changes come later and less predictably.
What answers the accountability question?
Design choices that make it clear, in advance and in the record, who decided what.
| Staff question | What answers it |
|---|---|
| Will the agent act under my name? | The agent has its own identity; its actions are recorded as its own |
| If I approve a draft and it is wrong, is that my fault? | The record shows what the agent proposed, what sources it used, why it was held, and what I changed |
| Can I refuse without being blamed? | Refusing is part of the reviewer's job; refusals are not counted against reviewers |
| What if I do not have time to check properly? | Review time is sized and protected; queues are staffed |
| Who answers when the agent acts on its own? | A named capability owner who set its bounds and accepted its evidence |
| What exactly is the agent allowed to do? | Its capabilities and depth are written down and shared with the team |
Each row turns a vague fear into a specific commitment. Our explainer on what AI accountability means covers the concept more broadly.
Why does the record matter so much to staff?
Because it is their defence. A reviewer who approved a draft can show, from the record, exactly what the agent proposed, what they were shown, how long they had, and what they changed. If the agent's draft relied on a source that later proved wrong, the record shows the source, not just the approval. If the reviewer refused and was overruled, the record shows that too.
Without such a record, staff are left to defend themselves from memory against a log that may show only their name and a timestamp. With it, accountability can be placed where it belongs: on the reviewer for what they decided, on the capability owner for the bounds, on the source owner for the facts.
How should the rollout be communicated?
Lead with accountability, not efficiency. Staff hear "the agent will save time" as "some of you will go" and "you will approve its work" as "you will take the blame". Explain instead what the agent may and may not do, how its actions are recorded separately from theirs, what reviewers will see before they decide, and that refusing is expected.
Be specific. General reassurance does nothing; showing staff the review screen, the record and the written bounds does. Our guide on communicating an AI agent rollout to staff covers the sequence, and a staff communication plan template gives a structure to fill in.
How does staffing the review queue affect resistance?
Directly. A reviewer asked to approve items with no time to check them knows they are being set up to carry risk they cannot manage. Sizing the held queue, protecting review time and giving reviewers authority to refuse are the most convincing signals that accountability is being taken seriously. Our guide on staffing the queue of actions AI agents hold covers the arithmetic.
What is the trade-off?
Answering the accountability question properly costs effort: separate identities, a detailed record, protected review time, written bounds and a culture where refusal is safe. It can also slow early adoption, because staff who understand their role clearly may still refuse a lot at first.
Leaving it unanswered is more expensive. The agent's value disappears into redone work, defensive refusals and stalled scope, and leaders conclude the technology does not work. Before the next rollout, list the questions your staff will ask about blame, and make sure each has a specific, visible answer in the design. The deployment approach that builds this in is set out in our guide to deploying an AI agent: train, shadow, certify, widen.
Questions people ask
Should staff have a say in an agent's bounds?
Experienced staff should be consulted, especially on which cases the agent should and should not take. They know where the process is fragile. Final decisions sit with the capability owner, but involving staff early reduces resistance and improves the bounds.
What if a reviewer refuses far more than others?
Look at what they refuse and why before treating it as a problem. They may be catching real issues others miss. If not, calibration sessions on shared cases help bring reviewers to a consistent standard without making refusal feel risky.
Do unions or works councils need to be involved?
Where they exist, early involvement is usually wise and in some places required. Accountability and monitoring questions are often central to their concerns, so the design choices above give a concrete basis for discussion.