Executives buy AI agents and frontline staff review their work, but the middle layer decides whether the deployment succeeds. Team managers, operations managers and risk managers staff the held queue, adjust limits, judge escalations and set the culture of review. They are usually the last to be trained. Reverse that, and most of the common failures never happen.
What do middle managers actually decide?
Look at the decisions a governed agent deployment needs every week and ask who makes them.
- Who reviews today. The manager builds the rota and decides whether the held queue has enough certified people on it.
- How limits move. When the queue grows, does the team ask for the agent's limits to be loosened, or add reviewers? The manager decides which request to make.
- What gets escalated. A reviewer is unsure about a held item. Does it go up, and to whom?
- What counts as good review. Whether the team takes time over held items or approves them fast depends largely on what the manager rewards.
- When to pull back. If the record shows something going wrong, does someone narrow what the agent may do, or wait to see?
Executives rarely see these decisions. Frontline reviewers cannot make most of them. They sit with managers.
Which decisions should never leave people at all?
Managers also hold the line on decisions that should not move to an agent at any depth. Two examples make the point.
Whether to reimburse a fraud victim turns on facts, policy, fairness and sometimes the customer's circumstances. An agent can assemble the case; the decision belongs with a person. The reasoning is set out in whether an agent can decide to reimburse a fraud victim.
Whether to report a rule breach to a regulator involves judgement about materiality, timing and the firm's obligations. An agent can draft the facts; a person decides. See whether an agent can decide to report a rule breach to a regulator.
Frontline staff may not have the authority to make these decisions. Executives are too far from the facts. Middle managers are where they land, and they need to recognise them as decisions that stay human.
Why are managers so often left out of training?
Because training plans follow the obvious users. The frontline will review held items, so they get trained. Builders will design agents, so they get trained. Executives need to understand the investment, so they get a briefing. Managers are assumed to pick it up.
They do not, at least not quickly enough. A manager who has not been trained sees a growing held queue and thinks the agent is underperforming. They see reviewers approving quickly and think the team is efficient. They see a spike in refusals and log a fault. Each reaction is reasonable from the old way of working and wrong in the new one.
How does this look in a regulated market?
Risk managers in retail banking show the pattern clearly. They sit between the board's risk appetite and the teams running agents. In Ireland, they work under the Central Bank of Ireland and EU rules including the AI Act and DORA, and must show that oversight of agents in areas like credit and payments is real; AI agent training for risk managers in retail banking in Ireland covers the detail. In Japan, they answer to the Financial Services Agency and must handle personal data under APPI, with a different supervisory style; see AI agent training for risk managers in retail banking in Japan. In both, the risk manager's daily judgement decides whether the controls on paper operate in practice.
What is the trade-off?
Training managers first delays frontline training, and the frontline is where the volume is. Managers are also expensive to take off the floor, and some will resist training they see as beneath their experience.
There is a real risk, too, that managers trained deeply on the system start doing reviewers' work themselves, recreating the bottleneck the agent was meant to remove.
The answer is to train managers on managing, not reviewing. They need to read the workspace, understand depth and verdicts, run the rota, judge escalations and sample decisions. They need to know when to split work between junior and senior reviewers, as set out in how to split review work between junior and senior staff. They do not need to approve every held item. Trained this way, managers make the frontline's work easier rather than replacing it.
What should you do differently?
Put managers at the front of the training plan. Teach them to read the workspace as a system, not a dashboard. Make clear which decisions stay human and that managers are the ones who hold that line. Reward careful review in how teams are measured. And make sure every manager understands that the answer to a long held queue is usually people and process, not looser limits. The guide to AI agent skills by role sets out what managers need alongside the other roles.
Where the Academy fits
Pryme Intelligence Academy's Operator track is built for the people who run a live workspace every day, including team leads and operations managers. Its modules cover reading the workspace, when something goes wrong, usage and billing, and running the rhythm. Module tests are set in the learner's own industry and role, so a risk manager's questions differ from a support lead's.
Questions people ask
Should managers be certified before their teams?
Yes, ideally. A certified manager can coach reviewers through their own practicum and spot problems early. A manager learning alongside the team cannot.
What if a manager disagrees with an agent's refusal?
They should find out which bound decided it. If the business's own limit is too tight, they can propose changing it through the proper route. They should never work around a refusal by hand.
How do you measure whether managers are making agents work?
Look at held-item waiting times, the quality of sampled approvals, how escalations are handled and whether limit changes are justified on record. Those reflect management decisions more than agent performance.