Every employee should know what AI can help with, where its answers can fail, what information they may share, how to check the result and when to ask for help. For everyday generative AI use, that knowledge should be visible in a small piece of work: choosing a suitable task, using permitted inputs and deciding whether the output is good enough to use.
A useful AI literacy baseline gives employees a shared starting point. A sales coordinator and a software engineer need different specialist skills, but both need to recognize an unsupported answer and understand the limits of their permission to act.
For managers and learning leads, the practical question is what someone should be able to explain and demonstrate before working independently. The people and change guide covers the wider support that makes that learning possible.
Know enough to choose and question
Start with a plain explanation of the tool people will actually use. A generative AI system produces text, images or other content in response to instructions and context. A fluent answer does not establish that its facts are correct, that it has seen the latest company information or that its recommendation fits the task.
NIST's generative AI risk profile describes confabulation, often called hallucination, as confidently presented false or erroneous content. It also notes that generated explanations and citations can be wrong. This matters when an employee tries to check a doubtful answer by asking the same tool to explain itself. NIST AI 600-1, section 2.2.
Teach people to ask three questions before starting:
- Can this tool help with the task? Drafting a product description from an approved specification sheet is a bounded use. Deciding an exception to a customer's contract requires authority and context that a drafting tool does not supply.
- Can I provide the necessary information? Knowing a fact does not mean being permitted to put it into any tool or account.
- Can I judge the result? If the employee cannot check the important claims, identify a qualified reviewer or choose another way to do the work.
England's AI foundation skills for work benchmark similarly extends beyond giving instructions. Its foundation includes routine tasks, adapting results, recognizing risks and checking information. It is a useful reference for the breadth of learning, rather than a substitute for the company's own working rules.
Make the baseline observable
A lesson titled “responsible AI” can cover many things. Translate the knowledge into decisions employees can practise. The following baseline is a starting point for ordinary drafting and information tasks; specialist work needs additional checks.
| What an employee should know | What they should be able to demonstrate |
|---|---|
Suitable uses and limits | Choose a bounded task and explain why the result can be checked. Identify a case where they should stop or seek approval. |
Permitted tools and information | Find the approved tool and account, then check whether the intended input is allowed. Ask when the rule is unclear. |
Useful instructions | State the task, audience, permitted source material and required output. Tell the tool what to leave unresolved. |
Accuracy and completeness | Compare important claims with source material. Catch missing conditions, invented facts and omissions. |
Judgment about people and context | Notice unsupported assumptions about a person or group. Check whether language and recommendations fit the situation. |
Responsibility and escalation | Explain who reviews, who may act and where to report a problem. Distinguish generating a draft from sending it or changing a record. |
These are observable behaviors, not a universal scoring system. A person who can edit a customer message has not thereby demonstrated that they can validate a spreadsheet analysis or approve an automated workflow.
A Reddit user working in technical sales described a useful complication. They suspected a colleague's messages were AI-generated, but struggled to explain the problem beyond occasional inaccuracies. After the discussion, they identified more concrete concerns: generic content, unnecessary length and errors. The suspected AI use was not established.
The lesson for a baseline is to make quality explicit. Employees should be able to explain what the recipient needs and show that the finished work meets it. Judging whether a message “sounds like AI” leaves both the standard and the correction unclear.
Give employees the answers only the company can supply
An employee can learn how to check a claim and still be unable to find out whether a customer document is permitted in the tool. That is an unresolved company decision, not evidence of poor AI literacy.
In a September 2026 account of work with a large European financial brokerage group, researcher Alexander Benlian described uneven use and uncertainty about appropriate applications. His account argues that training needs to connect with access, shared practices and governance. It does not establish that one training program solves those conditions.
Give employees a short, maintained reference they can use during work. It should answer:
- Which tool and account may I use? Include where to find the current approved list and how to request access.
- Which information may I enter? Explain the relevant data categories with examples from this team's work, including what to do when classification is unclear.
- What may the tool do? Separate drafting or suggesting from sending, publishing, executing code or changing records. Specify the required approval for each permitted action.
- Who checks and owns the result? Name the role responsible for review, any required disclosure and the person authorized to make the final decision.
- Where do I get help or report a mistake? Provide an actual contact route for uncertain outputs, inappropriate content or information shared in error.
Keep that reference beside the practice material and put an owner and review date on it. Employees should not have to infer company policy from a chatbot's reassurance or a generic course.
Practise with a source that has a missing fact
Suppose Jamie Lee, a sales coordinator at an office furniture supplier, is learning to draft a short product reply. A customer wants to know whether a chair has adjustable arms, comes in grey and can arrive next Friday.
For the practice exercise, Jamie has an approved product sheet confirming adjustable arms and grey upholstery. It gives the chair's dimensions but says nothing about stock or delivery dates. The customer question is supplied as permitted practice text, without customer identifiers.
Have Jamie complete the task and explain the choices:
- Check the boundary. Find the approved account and confirm that the practice sheet and question are permitted inputs.
- Give a bounded instruction. Ask for a brief customer reply using only the sheet for product facts. Require any unanswered question to remain unresolved.
- Compare the draft with the source. Check the arms, colour and any added specifications. If the draft promises Friday delivery, remove that unsupported promise.
- Resolve the next step outside the draft. Identify the stock or delivery contact who can confirm availability. Keep the message as a draft until the required answer and review are complete.
The important result is Jamie's decision about the missing delivery fact. A polished reply that invents a date fails the task. A reply that accurately explains the chair and says delivery still needs confirmation preserves what is known.
Ask Jamie to show where each claim came from. Then try another permitted example with a different gap, such as a product sheet that omits the warranty terms. This helps distinguish learning the checking habit from remembering one exercise. It is still a limited observation of those tasks, not proof of readiness for every AI use.
Extend the baseline as the work changes
After the shared foundation, use examples from the employee's own responsibilities. A service adviser may need practice checking policy exceptions; an analyst may need to inspect calculations and source data; someone configuring an automation needs to understand its permissions and effects before it runs. The role-specific learning path guide shows how to give employees, managers, builders and reviewers different practice tasks within one workflow.
The Skills England employer guide to AI upskilling recommends practical, context-sensitive learning that combines responsible use with technical and nontechnical skills. Treat the baseline as the beginning of that work. A common foundation need not mean an identical lesson for every job.
Make space for practice through an explicit agreement about learning time and delivery work. Revisit the baseline when approved tools, permissions, data rules or responsibilities change, and when a mistake exposes a gap the examples did not cover.
Before commissioning another course, complete the local reference and watch one appropriate task. The missing answer may be a skill to teach, a source to make available or a decision the company still owes its employees.
Questions about AI literacy for employees
Does every employee need to learn coding?
Coding is not a prerequisite for the everyday drafting and checking tasks in this baseline. Employees do need enough understanding to choose a suitable use, follow the information rules and evaluate the result. A role that builds software or configures automation needs additional training and permissions for that work. Start with the observable baseline, then add the skills the actual role requires.
Is completing an AI course enough?
Course completion shows that someone completed the course. Before independent use at work, check whether they can apply the learning with the company's approved tools, permitted inputs and review requirements. Use a relevant exercise, such as the product reply with a missing delivery fact, and record where help was needed. A successful exercise supports that specific observation, not unrestricted permission to use AI.
What should an employee do when they cannot check an answer?
Keep the answer out of decisions or messages that rely on the uncertain claim. Identify what cannot be verified, find an authoritative source or ask a qualified reviewer through the company's help route. If that support is unavailable, use an established process or leave the fact unresolved. Asking the model for greater confidence does not resolve the evidence gap. The local reference should make the next contact clear.
Should everyone receive the same AI training?
Everyone can share foundation topics, including limitations, permitted information, checking and escalation. The examples, accessibility support and depth should fit the person's work and starting point. For a furniture sales coordinator, checking a product reply is relevant practice; it does not teach an analyst to validate calculations. Use the shared baseline to identify common expectations, then choose role-specific exercises and review them as the work changes.
