The workplace AI rollout now needs a training trail

The office AI tool often arrives in the language of convenience. It promises quicker summaries, neater notes, first drafts, customer-service prompts, search across internal files or a less painful way to handle forms. The temptation is to treat it like any other software rollout: switch it on, send a link, add a short policy page and wait for the productivity story to prove itself.

Europe’s AI Act makes that look too thin. One of its quieter obligations, Article 4, requires providers and deployers of AI systems to take measures, to their best extent, to ensure a sufficient level of AI literacy among staff and other people using AI systems on their behalf. It is not the most dramatic part of the law. It does not have the headline force of banned practices or high-risk classifications. But it may be one of the first parts ordinary workplaces feel.

The useful word is sufficient. The Act does not say every employee has to become a machine-learning specialist. It defines AI literacy as skills, knowledge and understanding that help providers, deployers and affected people make informed deployment of AI systems, while gaining awareness of opportunities, risks and possible harms. The level depends on technical knowledge, experience, education, training, the context in which a system is used, and the people or groups on whom the system is used.

That makes AI literacy less like a one-off webinar and more like an operational trail. A sales team using a text assistant needs a different conversation from a hospital administrator, a hiring manager, a public-service caseworker or a developer wiring a model into a product. The same tool can carry different risks depending on whether it drafts marketing copy, ranks applicants, summarises a complaint or touches personal data. The point is not to scare people away from the software. It is to stop the rollout pretending that context does not matter.

The European Commission’s AI literacy pages underline that Article 4 entered into application on 2 February 2025. They also make clear that enforcement sits with national market surveillance authorities, while the AI Office supports implementation with Member States and the AI Board. In plain terms, the training story is not only a human-resources nicety. It is becoming part of how an organisation explains that people using AI have been prepared for the task in front of them.

The Commission’s living repository of AI literacy practices shows why this will not be solved by a single template. The repository gathers examples from companies and the public sector, but it carries an important warning: copying a listed practice does not automatically create a presumption of compliance, and publication does not mean the Commission endorses or evaluates it. That caveat is useful. It discourages a box-ticking market in borrowed slide decks and points back to the real question: what does this particular group need to know before this particular AI system is used in this particular setting?

For workers, the change may be more practical than legalistic. A good AI rollout explains when the system is helpful, where it is weak, what data should not be entered, when a human check is needed, how errors can be reported, and whether the output affects someone outside the organisation. It also admits uncertainty. A tool that sounds fluent can still invent details, flatten nuance, reflect bias in training data or make a decision look more objective than it is.

For vendors, the literacy trail matters because buyers will increasingly ask for more than a demo. Documentation, training materials, risk explanations, update notes and clear limits may become part of the product package. The AI Act already separates different roles in the AI chain, including providers and deployers. The people buying or switching on a system will want to know what support helps their own staff use it responsibly.

The new AI Act Service Desk and Single Information Platform point in the same direction. The Commission describes them as tools to help stakeholders navigate obligations, with an AI Act Explorer, a compliance checker and a route for questions. That does not turn a public website into legal certainty. It does show that AI governance is moving from abstract debate to practical infrastructure: explain the system, map the risk, train the people, keep the record.

There is a risk of overclaiming what training can fix. AI literacy will not remove weak data, poor procurement, vague accountability or pressure from managers who prize speed over judgement. It cannot turn a harmful system into a safe one by adding a course at the end. But it can change the first question around workplace AI from “can we use this?” to “who understands what this is doing, where it belongs, and when to pause?”

That is a healthier starting point. The workplace AI rollout is no longer just a button in a productivity suite. It is a test of whether organisations can make new software legible to the people expected to use it, and to the people affected by its results.