The person who wrote a process knows why each step and threshold exists and where the exceptions hide. That knowledge should shape both the agent's brief and the reviewers' training, delivered by the same person. When different people teach the agent and the reviewers, the two drift apart, and the agent starts holding actions its reviewers do not understand.
What does the process author know that others do not?
The reasons. A procedure document says that changes to supplier bank details must be confirmed by phone to a number already on file. The person who wrote that rule remembers the fraud that prompted it, the variant that nearly got through afterwards, and the two suppliers who always change details legitimately at year end.
Rules without reasons get misapplied. A reviewer who knows only the rule will confirm by phone and approve, even when the number on file was itself changed last week. A reviewer who knows the reason will spot the gap. The same is true of the agent: a brief that states the rule produces an agent that follows it literally, while a brief shaped by someone who knows the reason includes the bounds and sources that catch the variant.
The exceptions matter as much as the reasons. Every mature process has cases the document does not cover, handled by people who know them. Those cases are exactly where an agent's drafts tend to be weakest and where reviewers most need preparation.
Why should the same person teach both?
Because the agent and the reviewers have to agree. Under a governed operating system the business does not change model weights; it qualifies the agent by changing its brief, sources, bounds and certification. Those choices decide what the agent holds and why. The reviewers then decide the held cases.
If a builder briefs the agent from the procedure document and a learning team trains reviewers from the same document, both inherit its gaps, and differently. The agent may hold supplier changes on a bound the reviewers were never taught about; the reviewers may release cases the agent's brief was written to stop. The record fills with overturned holds and puzzled notes. When the process author shapes both, the agent's verdicts and the reviewers' understanding come from one source and stay consistent.
This also changes how many people the business needs to train for each agent. How many people need to be trained to run one AI agent depends partly on how much of the expertise sits in one head; a process author who teaches spreads it.
Do they need to be technical?
No. Briefing an agent is writing the job in plain language: what it does, which sources it may use, what it must not do, when to hand back. Agent Studio assembles the brief from that description. Choosing bounds is a business judgement about limits and red lines. Reading an Attest result means looking at which of the seven gates an agent cleared on the business's own cases and asking why it failed the others.
The process author is better placed for all of this than an engineer, because the questions are about the process, not the software. Wealth firms see this clearly: the paraplanner who built the suitability process is the right person to brief the agent that drafts suitability reports and to teach the people who review them, which is part of why wealth management will not scale AI agents without trained reviewers who learned from them.
What is the trade-off?
Several, and they are real. Process authors are usually senior and busy, and teaching takes them away from the work. Some are poor teachers: expert in the process, impatient with people learning it. Some have workarounds they no longer notice, which they will pass on to the agent and the reviewers as if they were the process. And some will defend the process against improvement, because they built it.
Each has a remedy. Budget the author's time explicitly in the deployment plan rather than assuming it. Pair them with a newer reviewer who asks naive questions; naive questions surface workarounds. Use the evidence, not opinion, to challenge the process: the agent's Attest results and the record of held, overturned and refused actions show where the process itself is weak. Aviation has used a similar model for a long time, with experienced line pilots as trainers and checkers, which what aviation can teach businesses about training for AI agents explores.
How should the teaching be delivered?
Short and repeated, because the first session is not what people remember. Reviewers tend to lose the exceptions and the reasons first and keep the screens; what people forget soonest after AI agent training covers the pattern. So the author should return to those points in calibration sessions, using fresh cases from the queue.
Informal formats help. A process author walking the team through three held cases over lunch reaches people who will not attend a formal course, as long as the session is built around real cases and a clear point; how to run a lunch-and-learn on AI agents that people remember shows how.
What should you do differently?
Name the process author in every agent's deployment plan as the person who shapes the brief and teaches the reviewers. Give them time for it. Have them write the brief and the reviewer guide together, so the reasons appear in both. Pair them with a newer colleague, and review the record with them monthly so the process improves as the agent does. The order in which to train each group is in the guide to training teams for AI agents.
Questions people ask
What if the person who wrote the process has left?
Find whoever now handles the exceptions; that person holds the working knowledge. Reconstruct the reasons from past incidents and the record, and write them into the brief and the reviewer guide before the agent goes live.
Should the process author also approve the agent's held actions?
Not as the main reviewer. They shaped the agent's brief, so reviewing its output is partly marking their own work. Keep them as the escalation point and the teacher.
How much time does this take from the process author?
Most at the start: shaping the brief, the shadow period and the first reviewer sessions. After that, a regular calibration session and a monthly look at the record are usually enough.