Teach AI literacy through real work decisions

A practical learning programme built around source checking, permissions and recognising when a suggestion should be rejected.

A workplace AI course can teach ten prompting tricks and still leave staff unsure what information they may upload or how to check a confident answer. Practical literacy begins with the decisions a person makes during their own work. The course should help them recognise what the tool knows, what it does not establish and who remains responsible.

Support research reports learning-related benefits in a particular deployment. HBS highlights task-dependent capability. METR found an early developer slowdown and later measurement difficulties. These findings justify testing learning and judgment locally, without promising that a short course will reproduce a research effect. [Support study](https://arxiv.org/abs/2304.11771), [HBS](https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/), [METR](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/), [update](https://metr.org/blog/2026-02-24-uplift-update/).

A fictional services company builds its course around three ordinary assignments: summarising a public document, drafting a reply from an approved procedure and preparing a meeting agenda from fictional notes. Each assignment includes a decision about evidence, data permission and review. Staff practise the whole workflow rather than treating prompt wording as an isolated skill.

The first exercise supplies a public article and asks for a summary. The instructor then introduces an AI draft containing a plausible claim absent from the article. Participants locate the unsupported sentence and rewrite it. They learn that a source link is useful only when it actually supports the statement. An attractive citation does not establish that the source was read correctly.

The second exercise uses a procedure with a date and a later replacement. The participant checks which version applies before drafting. One fictional customer request is ambiguous, so the correct result is a clarifying question. The instructor discusses why the assistant should not fill the gap simply to produce a complete response. The skill is recognising missing context, not coercing a more confident answer.

The third exercise asks participants to identify information that should stay out of the prompt. Fictional notes contain an unnecessary personal detail, a sample password and a confidential commercial assumption. The examples are invented; no real secret is used in training. Staff practise minimising the input and choosing the approved tool or a manual process. They also learn where to ask when permission is unclear.

A fourth short exercise examines action authority. A model drafts an email, but the course makes clear that a draft has not been sent. A suggested calendar entry has not been accepted by participants. A proposed case update is not a stored record. The learner checks the appropriate system receipt before describing an action as complete. This distinction prevents everyday workflow mistakes as well as exaggerated claims about AI.

Include a failure journey. The tool becomes unavailable after a participant has reviewed half the draft. The interface preserves safe work and explains how to finish manually or retry. The learner should not copy a raw service error into a customer reply or repeatedly trigger an action with an uncertain outcome. Plain recovery and state checking are part of competence, not advanced specialist topics.

Evaluation uses a new task with a deliberate ambiguity and an unsupported suggestion. The participant explains what they would verify, what they would remove and when they would ask a colleague. Assess the reasoning and final result. A quiz about terminology can be supplementary, but it cannot establish that the person knows how to handle a real draft.

Adapt examples to roles. Finance staff may practise receipt evidence; marketing staff may practise unverified product claims; support staff may practise a procedure exception. People need enough common vocabulary to collaborate, while the difficult cases should resemble their actual responsibilities. Experienced employees contribute examples and review material so the course reflects the work rather than generic demonstrations.

Reserve time to update the programme. When an approved tool changes, revise relevant exercises and tell staff what the change affects. Keep a simple channel for reporting confusing suggestions. The reports can inform both training and product improvement, provided necessary diagnostics are restricted and private content is not collected casually.

Costs include protected learning time, instructor preparation, approved software and maintaining the examples. No current provider price or workplace trial is verified here. Research effects remain specific to their settings and tools; the proposed programme is an original educational design. It makes no prediction about job loss or guaranteed productivity.

A person finishing the course should be able to say why a suggestion fits, where its evidence comes from and what still needs a human decision. That practical confidence makes workforce change more manageable. It also gives the organisation a clearer view of where AI helps its people and where the work still needs better information or a different process.