Implementation Playbooks4 min read

AI Literacy Training for Business: Teach Three Levels of Authority

AI literacy training for business becomes actionable when every workflow states whether AI may advise, prepare or act—and what human control each level needs.

Bokili EditorialPublished September 15, 2026· Verified September 15, 2026

Illustration of AI moving through advise, prepare and act authority levels with human approval and monitoring

AI literacy training for business often teaches what generative AI is, how to prompt it and which data not to share. That baseline matters, but it leaves a practical question unanswered: what is the AI allowed to do in this workflow? A tool may advise a person, prepare an action for approval or act within a defined boundary. Each level needs different skills, checks and escalation rules.

The safest useful rule is to define authority per workflow, not per product. The same assistant could advise on a supplier email, prepare a draft reply and trigger a low-risk internal notification. Calling the tool “approved” does not explain which of those actions is permitted.

What AI literacy training for business must add

The European Commission’s AI literacy Q&A tells organisations to consider people’s role, existing knowledge, the context of use and the risks involved; simply asking staff to read system instructions may be ineffective. NIST’s Govern and Map playbooks likewise call for clear human roles, delegated authorities, training and documented oversight. These principles become teachable when employees can name the authority level of a real task.

Three levels of AI authority

Advise

The AI analyses, suggests or explains. A person decides what to accept and performs the action. Teach source checking, uncertainty and how to challenge a recommendation.

Prepare

The AI drafts, fills or stages an action, but execution waits for human review and approval. Teach comparison with source material, edit ownership and the approval record.

Act

The AI executes a narrowly defined action inside explicit limits. Teach monitoring, exception handling, audit evidence, escalation and the stop condition. Higher-consequence decisions still need meaningful human control.

Authority changes the lesson

	Authority level	Training emphasis
Advise	Human decides and acts	Question the recommendation; verify material claims; record the reason for the decision
Prepare	AI stages; human approves	Compare draft with sources; inspect sensitive fields; approve or return with a reason
Act	AI executes inside limits	Know the boundary; monitor outcomes; handle exceptions; use the stop and escalation route
Across all levels	A person remains accountable	Know the owner, evidence requirement, prohibited inputs and review cadence

Worked example: an expense-claim exception

Consider a fictional finance workflow that detects a possible duplicate expense. At Advise level, the AI flags the two entries and explains the match; an analyst decides whether to investigate. At Prepare level, it assembles the relevant receipts and drafts a review note, but the analyst checks the evidence and submits the case. At Act level, it may route a claim to a named exception queue and notify the submitter under a tested rule. It does not approve, reject or pay the claim unless that separate authority has been explicitly designed and governed.

A generic prompt lesson would miss the most important differences. The adviser needs judgement practice. The preparer needs an approval checklist. The acting workflow needs limits, monitoring and a stop route. The task is similar; the authority is not.

Reading is a start. Practice makes it stick.

Start learning

Build an authority card for every workflow

Authority card

Task and business outcome
Authority level: Advise, Prepare or Act
Action the AI may take—and actions it may not take
Human owner and required approval point
Permitted inputs and evidence the person must inspect
Consequence if the output is wrong
Exception and escalation route
Monitoring signal, stop condition and review date

Do not copy one authority card across a whole tool. A chatbot used for internal brainstorming and the same chatbot connected to a customer workflow have different consequences. Review the card when the model, data source, integration, policy or business process changes.

Teach the level with a ten-minute practice

Ten-minute authority sort

Pick one AI-assisted task your team already performs or plans to pilot.
Write the final action in one verb: recommend, draft, route, send, approve, pay or another concrete action.
Classify the current workflow as Advise, Prepare or Act.
Name the human decision or approval point and the evidence that person sees.
Write one prohibited action, one escalation trigger and one stop condition.
Ask a colleague to classify the same workflow. Resolve any disagreement before access expands.

The disagreement is useful evidence. If two people assign different levels, the workflow boundary is not clear enough to train or govern. Rewrite the action and approval point until both can explain it in the same plain language.

Link literacy to practice, data and stopping

This authority model complements, rather than repeats, existing Bokili guidance. Practise manager choices with the four-decision AI literacy lesson. Define what information may enter a system with a prompt data boundary. Give every operational workflow a stop rule. Those pages cover decisions, inputs and failure response; the authority card connects them to what the AI may do.

Four manager decisions to practise — https://bokili.com/en/learn/ai-literacy-training-business-manager-decisions
Build a prompt data boundary — https://bokili.com/en/learn/prompt-data-boundary
Give every AI workflow a stop rule — https://bokili.com/en/learn/ai-workflow-stop-rule
Bokili for HR and L&D — https://bokili.com/en/for-hr

Good AI literacy is not a single course completed once. It is the ability to recognise the task, the consequence and the authority in front of you—and to apply the right check before work moves. Start with one authority card, rehearse one exception and make the boundary visible where the work happens.