The cheapest control in an AI agent deployment is a trained person

A trained person is the cheapest control you can add to an AI agent deployment. Limits, rules and certification catch what can be specified in advance; a trained reviewer catches what cannot, such as context, intent and the case that looks routine but is not. Training costs less than another layer of engineering, and it improves every control around it.

Why does a trained person cost less than another technical control?

Because the person is already there. In most deployments someone already reviews held actions: a credit controller, a support agent, a payments analyst. The question is not whether to employ them but whether they know what they are doing. Training them is a one-off cost plus a short refresh each year.

Compare the alternatives. Writing a new rule for every pattern the agent misses means specifying the pattern, testing it, certifying the change and maintaining it. Each rule handles one case. A reviewer who understands how the agent decides, what provenance means and when to hold back handles the whole class of cases, including ones nobody predicted. The marginal cost of catching the next new problem is close to zero.

What does a trained reviewer catch that limits do not?

Limits are blunt by design. A refund limit of 100 holds every refund above 100 and lets every one below it pass. It cannot tell that this customer has claimed non-delivery four times this month, or that the "manager approval" pasted into the ticket is fake. A trained reviewer can.

Specifically, trained reviewers catch:

• External claims dressed as facts. Under the Pryme Context Protocol, an emailed invoice or a pasted approval is external, and an external claim cannot authorise a payment. A trained reviewer reads the label.
• Context the agent was never given. A phone call, a complaint in another channel, a relationship history.
• Drift. A run of drafts that are slightly wrong in the same way because a source went stale.
• Pressure. A customer pushing for an override of a refusal, where the right answer is to hold firm.

Where does the argument stop?

At three limits, and they matter. First, people are inconsistent. Two trained reviewers will still decide some cases differently, which is why calibration is part of training rather than an extra. Second, people get tired and busy. A reviewer at the end of a long shift with a full queue is a weaker control than the same reviewer at the start. Third, people can be pressured, by customers, by targets or by managers.

So a trained person does not replace platform controls; they sit inside them. Every action meets five bounds at once: role scope, certificate, sources, the organisation's own limits and the platform guardrails. Some actions are withheld at the platform floor for every agent at every depth, regardless of who approves. The trained person handles what remains after those bounds have done their work. That division is what makes the person cheap: they are not asked to catch everything, only the judgement calls.

Why does training make other controls work better?

Because most controls depend on someone using them properly. A held action is only a control if the person who decides it reads it. A refusal is only useful if the reviewer does not override it under pressure. A certificate lapse is only safe if the team knows that certificates lapse after 90 days, that a run that clears no depth drops the agent to depth 1 the same day, and what to do about the sudden flood of drafts.

Untrained people turn good controls into formalities. Trained people turn them into decisions. Every verdict and every human decision is filed in an append-only record, and an auditor reading it can tell the difference quickly: approval times, override reasons, edits to drafts. The articles on compliance officers as the real control and collections agents as the real control show this from two very different seats.

What does training a reviewer actually involve?

Less than most people expect, but more than a memo. Reviewers need the concepts (depth, verdicts, provenance, the record), practice on realistic cases from their own work, an assessment that tests judgement rather than recall, and regular calibration. For a credit control team, for example, the method is set out in how to train the credit control team to review AI agent decisions, and a drill to test it under pressure in how to run an AI agent incident drill for the credit control team.

Format matters less than practice. Whether the concept layer is online or in person is a scheduling choice, covered in whether AI agent training should be online or in person. What cannot be skipped is doing real review decisions and getting feedback.

What should you do differently?

Before funding another technical control, ask whether a trained reviewer would catch the problem. If yes, train first. Budget training as part of the deployment, not as a later extra. Measure reviewers on correctness, sampled after the fact, rather than speed. And make sure your platform controls handle what people should never be asked to catch alone. The approach to training whole teams is in training teams for AI agents.

Where the Academy fits

Pryme Intelligence Academy's lessons, labs and tests are free, and every lab runs in a free Build workspace. For teams on paid Pryme Intelligence plans, seats are included along with certificates for the people on them, and an admin view shows progress, results and expiry dates, so the cost of training the control is mostly people's time.