AI literacy is a duty because the people who approve, correct or overrule an AI agent are part of the control. If they cannot read a held action, tell a refusal from a fault, or see which bound decided a verdict, the control is empty. Treat literacy as a condition of acting, not an optional course in the benefits catalogue.
Why does literacy count as a duty rather than a benefit?
A perk is something you offer to make people happier or more employable. You can take it or leave it, and the business carries on either way. Literacy for people who work with agents is different, because the business does not carry on safely without it.
Consider what a governed agent actually does. It drafts a refund and waits for a named person. It refuses to pay an invoice whose bank details arrived by email. It is certified to draft but not to send. Each of those outcomes hands a decision to a person, and the quality of that decision depends on whether the person understands why the agent stopped where it did. A reviewer who does not understand depth will approve things the agent was never trusted to do alone. A reviewer who does not understand provenance will accept an external claim as if the system of record had issued it.
That is why literacy belongs in the same category as knowing the payment limits or the complaints procedure. It is part of doing the job, and the organisation owes it to its customers to make sure the job is done properly.
What does the duty actually cover for people who work with agents?
General awareness of AI is not the duty. Knowing that models can make things up, or that prompts shape answers, is useful background; the piece on what prompt literacy means covers that layer. The duty is narrower and more practical. It is the set of things a person must understand to make good decisions about a specific agent's work:
- Depth. What the agent may do on each capability: read, draft, execute routine work, move money, or coordinate other agents.
- The decision chain. Every action meets role scope, certificate, sources, the customer's own limits and the platform guardrails, and the lowest bound wins.
- The four verdicts. Allowed, held, refused and withheld, and what each one asks of the person.
- Provenance. The difference between a sealed fact from the ledger and a declared one from a customer.
- The record. Every verdict is filed, so every approval a person gives is filed too, with their name on it.
Someone who understands those five things can supervise an agent. Someone who does not is guessing, and guessing at scale is what a governed system exists to prevent.
Is the EU AI Act the reason, or only one reason?
The EU AI Act makes the point explicit. Article 4 expects providers and deployers to take measures to ensure a sufficient level of AI literacy among staff dealing with AI systems on their behalf, taking into account their knowledge, experience and the context of use. If you want the detail, read whether the EU AI Act requires AI literacy training and how the same duty plays out in a specific sector, such as AI literacy for agents in media.
But the law is not the root of the duty. In the UK, the FCA's Consumer Duty asks firms to deliver good outcomes for retail customers, and a reviewer who waves through a poor agent decision produces a poor outcome regardless of any AI-specific rule. The article on Consumer Duty and training staff who supervise AI agents works through that link. Outside regulated sectors, the duty still exists: it is owed to the customer whose refund was wrongly declined and to the colleague who has to unpick the mistake.
Framing literacy as a legal checkbox invites the minimum. Framing it as a duty of care invites the right question: could this person explain, today, why the agent did what it did?
What is the trade-off?
Treating literacy as a duty costs something. People spend hours learning instead of working queues. Some will fail a test and need to retake it. Managers have to track who holds what and chase the ones whose standing has lapsed. A team that wanted to switch agents on next month may need another few weeks.
There is also a cultural cost. Calling something a duty implies consequences for not meeting it, and that can feel heavy-handed to people who were managing perfectly well before agents arrived. Some will hear it as a statement that they are not trusted.
The alternative costs more, though it costs later. Without tested literacy, the first serious mistake gets traced back through the record to an approval given by someone who did not understand what they approved. At that point the business cannot show it took reasonable steps, and the person who clicked approve carries a blame that properly belongs to whoever decided training was optional. The trade is a few weeks of friction now against a credibility problem later. Most organisations, once they see it written down, choose the friction.
What should you do differently?
Start by naming who the duty applies to. It is not everyone equally. Everyone in the business benefits from a short grounding in why agents must be governed. The people who approve held actions, correct drafts or set limits need much more, and they need it tested.
Then change three things:
- Stop counting attendance. A completion record shows that someone opened a course. It does not show they can read a verdict. Test understanding in the person's own industry and role.
- Tie approval rights to current standing. If someone's assessment has lapsed, their right to approve held actions should pause until they renew. This is a practice a business can choose to adopt, and it is the clearest signal that literacy is a duty.
- Keep the evidence where it can be checked. A training record should be a fact you can verify at source, not a spreadsheet cell someone typed.
None of this needs to be punitive. It needs to be honest about what the job now involves. The guide to AI literacy and certification for people who work with agents sets out the levels in more detail.
Where the Academy fits
Pryme Intelligence Academy separates the two layers. The free Foundations track gives anyone in the business a grounding in why agents must be governed and how the decision chain works. Reviewers and team leads take the Operator track, which ends in a written exam and a two-week practicum of real review decisions signed off by a Pryme assessor. The resulting certificate is valid for 12 months and can be checked at source through VerifyDoc.ai.
Questions people ask
Does every employee need the same level of AI literacy?
No. Everyone benefits from a short grounding, but the duty grows with the decisions a person makes. Someone who approves payments an agent has prepared needs tested, role-specific understanding; someone who only reads agent summaries needs much less.
Is a one-off training session enough to meet the duty?
Rarely. Agents change, their depth changes as certificates are earned or lapse, and the product itself changes. A duty that is met once and never revisited drifts out of date, which is why renewal matters as much as the first assessment.
Who is responsible if a reviewer was never trained?
The record will show who approved the action, but the failure usually sits with whoever let an untrained person hold approval rights. That is why the decision to require literacy belongs with management, not with the individual reviewer.