Prompting courses teach people how to get better text out of a chatbot. Agents that act need something else: people who understand how far the agent may go, why it held or refused an action, where its facts came from and what the record will show. A team fluent in prompts but untrained in oversight is unprepared for the moment an agent proposes moving money.
What do prompting courses actually teach?
How to ask. Write clear instructions, give context, provide examples, ask for a format, iterate. These are useful skills for anyone drafting with a general assistant, and many people benefit from them.
But the model they assume is a conversation. A person asks, the system answers, the person decides what to do with the answer. Nothing happens in the world until the person acts. In that model, the person is always the actor and the AI is always the adviser. Prompting skill improves the advice.
How is an agent that acts different?
The agent is the actor, within bounds. A governed agent can draft at depth 2, execute routine work at depth 3, or move money and take hard-to-undo actions at depth 4, depending on what it has been certified for. Every action meets five bounds at once: role scope, certificate, sources, the organisation's own limits and the platform guardrails. The lowest wins. The verdict is allowed, held for a named person, refused as outside the agent's bounds, or withheld at the platform floor for every agent.
None of that is controlled by prompting. A reviewer in a claims team does not type instructions to the agent; they receive its held settlement, check its sources and decide. A payments analyst does not refine a prompt; they read the provenance of a changed beneficiary and approve or refuse. The skill is judging, not asking, and it is tested every time a held action reaches them.
What do people need instead?
The skills that matter for agents that act are different in kind:
| Prompting course teaches | Agent oversight needs |
|---|---|
| How to phrase a request | How to read a verdict and which bound decided it |
| How to give the model context | How to read provenance: sealed, signed, declared, external |
| How to iterate on output | How to decide a held action, and when to refuse |
| How to spot a poor answer | How to spot a confident draft built on weak facts |
| Personal productivity | Accountability recorded in an append-only record |
| One person, one tool | A team, a queue, limits, certificates that lapse |
The right-hand column is what keeps money, customers and the organisation safe. It is also what supervisors and auditors will ask about.
Why do organisations keep buying prompting courses?
Because they are easy to buy and easy to complete. A half-day session, a certificate of attendance, a completion rate to report. Leaders feel the organisation has been trained in AI. Staff feel more confident.
That confidence is the risk. A reviewer who has learned to coax good answers from a chatbot may trust a well-written agent draft more, not less. They may try to "prompt" their way around a refusal rather than understand why it happened. They have no idea that certificates lapse after 90 days, or that a run that clears no depth drops the agent to depth 1 the same day, so the first time drafts flood the queue they assume something is broken.
There is a governance risk too. A board told that "all staff have completed AI training" may believe the organisation is ready for agents that move money. If that training was about prompting, the board has been given false comfort, and the gap will be found by an auditor, a supervisor or a customer rather than by the organisation itself.
What is the trade-off?
Oversight training takes longer and needs the real product. Concepts like depth and provenance can be taught in a few hours; judging held actions well needs practice on realistic cases and feedback on decisions. That costs more time than a prompting workshop and cannot be done with a generic chatbot.
Prompting skill is not worthless. People who build agents, design their briefs or draft content will use it. But it belongs in a builder's toolkit, not as the foundation of training for everyone who works alongside agents that act. For most staff, an afternoon spent learning to read verdicts is worth more than a day spent learning to write prompts.
What should you do differently?
Separate the two needs. If staff use general assistants for drafting, a short prompting session is fine. For anyone who will review, approve or oversee agents that act, build training around the decision chain, depth, verdicts, provenance and the record, with practice on the organisation's own cases. Then ask learners afterwards what to improve, using something like the training feedback form template for AI agent courses, and check whether they decide better, not whether they enjoyed it.
Pryme Intelligence Academy is built on this distinction: its Operator track teaches human review, reading the workspace and when something goes wrong on the real product, rather than prompting. Practical guides show what oversight training looks like for specific teams, such as training the advice team to review AI agent decisions and training the casework team. Sector pieces on healthcare and logistics show why reviewers are the constraint on scale. The overall approach is in training teams for AI agents.
Questions people ask
Is prompting irrelevant once agents are governed?
No. Builders who design agents' briefs use similar skills. But reviewers and operators do not control agents by prompting, so it should not be their main training.
Can a prompting course be adapted for agent oversight?
Only by replacing most of it. The concepts that matter, such as depth, verdicts, provenance and the record, are not prompting topics. Start from oversight and add prompting where it helps.
How do you test whether oversight training worked?
Sample real review decisions after training and check how many were right. Watch override rates on refusals and approval times. These show judgement; course completion does not.