By Catarina Mestre
·
7 min
·
28 October 2026
AI literacy means understanding enough about the AI tools you use to make informed decisions at work. A small agency should connect that learning to real tasks, permitted data and human review. The EU duty requires measures supporting that learning; it does not require a guaranteed level for each person.
Start with the jobs your team already does. Build a shared foundation, then add the knowledge each role needs before its work reaches a client.
What does Article 4 require now?
Article 4 requires AI providers and deployers to take measures supporting the development of AI literacy. A deployer is an organisation using an AI system under its authority. Buying an assistant for agency work commonly places the organisation in that role.
The July 2026 amendment to the AI Act retains that duty. It explicitly says organisations do not have to guarantee any specific level of an individual's literacy. Measures must take account of people's existing knowledge and experience, the use context and those affected.
The European Commission's Article 4 questions and answers confirm that ordinary uses, such as drafting advertising text, are relevant. It recommends an approach matched to the systems and their risks. A certificate is not required, and an internal record of learning can help. Article 4 also prescribes no AI officer or particular governance structure.
The duty has applied since 2 February 2025. This article checks the amended law and guidance as at 27 October 2026. It offers general information and practical suggestions, not legal advice or a verdict on your agency's compliance.
Who should your agency's plan cover?
Include the people using AI systems on the agency's behalf, including relevant freelancers and contractors. Map what they do before choosing what they need to learn.
An account manager drafting a routine update and a strategist interpreting research face different mistakes. A designer generating concepts needs a clear route for rights questions. The operations lead needs to know who can authorise tools and handle an incident.
For a UK agency, the territorial conditions matter. Article 2 of the AI Act can cover organisations outside the EU when their AI system's output is used there. Serving an EU client is a reason to assess scope, rather than a complete legal test by itself.
Check unusual uses with a qualified adviser. Recruitment, decisions affecting people and systems built or marketed under your agency's name can raise obligations beyond everyday content assistance.
Match learning to the work
These are agency examples to adapt. They are not a prescribed syllabus or a legal compliance test.
| Role or task | Learning to prioritise | Practical demonstration |
|---|---|---|
| Everyone using AI | Limits, permitted data and escalation | Spot an unsupported claim and identify who can help |
| Account and project leads | Client terms, approvals and handover | Explain how an AI-assisted update reaches release |
| Copywriters and designers | Sources, brand judgement and asset rights | Check a draft or concept against the agreed brief |
| Analysts and strategists | Original data, calculations and uncertainty | Trace a report claim to its source and recalculate it |
| Operations and owners | Tool permissions, incidents and responsibility | Show how access is approved and a problem is handled |
| Freelancers acting for the agency | Relevant agency rules and client constraints | Confirm the permitted tool and reviewer before starting |
What should everyone learn first?
Give the team a common explanation of what their tools can get wrong and what to do when that happens. Use one familiar job rather than a catalogue of AI terms.
For a team exercise, use fictional campaign data showing twelve enquiries. Give the team an AI-written update claiming twelve sales. Ask whether the update can be sent.
The team should trace the number to the source, correct the claim and name the reviewer. Then introduce a client rule forbidding personal data in that tool. Ask what must change before the task can continue.
Cover the team's actual working rules: which tools are approved, what information may go into them, who reviews the result and when work must stop. Explain that an answer's fluent wording does not establish accuracy.
Then practise the handover. The next person needs the brief, source material, known uncertainties and the review decision. A draft labelled “checked” is unhelpful if nobody knows what the check covered.
Use the agency AI review gate to agree the evidence required before release. Staff learning should make that process easier to run and easier to question.
How can a busy agency teach this without a large programme?
Start with a short session around one job, followed by a supervised attempt. This is a practical starting format, not a legally sufficient duration.
Ask people to explain their decisions while doing the task. An analyst can show where a number came from. A copywriter can show why a passage fails the client's brief. A project lead can name the person who approves release.
Use what they struggle with to choose the next session. If several people paste information without checking permission, teach that decision before adding another prompting technique.
For wider examples, the European Commission's living repository of AI literacy practices, checked on 27 October 2026, describes varied learning approaches. Copying a practice does not create a presumption of compliance. Select a format that fits your team and test whether people can use it in their own work.
Keep a small record of the measures taken
Add one entry for each use case. Record what happened, rather than ticking a generic completed box.
| Record | Example entry | When to revisit |
|---|---|---|
| Tool and task | Approved assistant drafting campaign summaries | Tool, plan or task changes |
| People covered | Analyst, account lead and relevant freelancer | Someone joins or changes role |
| Known risks | Unsupported claims and incorrect comparisons | A new failure appears |
| Rules used | Approved data only; source and number checks | Client terms or agency policy change |
| Learning completed | Team exercise and supervised report review | A gap remains or a risk changes |
| Evidence and owner | Worked example, learning date and named lead | Owner changes or evidence is missing |
| Next action | Practise checking percentage changes | At the agreed review point |
How do you know the learning reached the work?
Check whether people can make the relevant decisions on a realistic example. Attendance records show participation. A completed task also shows where the team needs help.
Ask a few ordinary questions: which inputs can go into this tool, which claim needs a source, and who may release the result? Keep these as human questions around the work. There is no need to generate a test through another AI tool.
Choose the exercise for the role. A designer might identify a rights question and escalate it correctly, without being expected to settle a legal dispute. An account lead might recognise that a client restriction changes the approved workflow.
Record gaps without turning the session into a performance ranking. A person who admits uncertainty gives the agency something useful to address. Assign the next learning action and an owner.
When should you revisit the plan?
Revisit the relevant learning when the work, tools or risks change. A new client restriction, a different model or a recurring error can justify a focused update.
The owner should also choose a review point that suits how often the team uses the system. A monthly check may suit an active workflow. This is an editorial suggestion, not a statutory schedule.
Keep the distinction between staff learning and public disclosure clear. The agency transparency article covers Article 50. Knowing how to review a draft does not by itself settle whether a particular output needs disclosure.
The rules for high-risk systems include further training and human-oversight requirements, with separate application dates. Obtain qualified advice before using a general agency plan to cover decisions about employment or other consequential uses.
What should your agency do next?
Choose one AI-assisted task already happening this week. Name the people involved, the permitted inputs and the person who owns release. Use the learning table to find one gap worth closing.
Run a short exercise on a fictional example, then record the measure taken and the next action. Build the plan around what people can explain and do in that task.
ViraOps AI Foundations helps the team understand AI limits, data handling and responsible use through short practical reads. Use it for the shared basics, then add your agency's tools, client terms and review rules.
Bring the team up to speed on the decisions behind everyday AI use. AI Foundations provides short practical reads to support your own working rules.
