At the end of a training session, a participant may be able to reproduce a demonstration yet still feel lost when faced with their first real case. The information is less organised, the request less precise and the result harder to judge. This is when learning needs to help them.
AI training at work should therefore be prepared around the work teams will need to carry out. Which tasks will they be able to handle? Which decisions will they need to retain? How will they know whether a result is good enough to use?
These questions produce a more useful programme than a list of features to explore.
Describe an observable skill
“Knowing how to use AI” covers very different activities. Preparing a product decision, developing with agents and organising their use within a team require some shared skills, but also expertise specific to each role.
For a product team, an objective might be to work with a set of user feedback while distinguishing observations from interpretations. For an engineering team, it might be to carry out a limited change and check its consequences. For a manager, it might be to examine a use proposed by their team and clarify the responsibilities it requires.
These objectives are examples of what training can address, not promises of mastery within a few hours. Their value is that they make something observable: can the person do the work, explain their choices and identify what they still need to check?
Let participants encounter the difficulty
An exercise becomes more instructive when it reveals a problem participants can recognise. For example, we can start with a set of user feedback and ask them to summarise it using AI. The first result may be clear, yet erase a disagreement or give too much weight to an isolated comment.
Returning to the documents then allows us to discuss what makes a good summary. Providing context, asking for source references and distinguishing an observed signal from a hypothesis become useful. The method addresses a difficulty experienced during the exercise.
This progression extends a principle from our approach to teaching product work: give participants the opportunity to encounter the problem before introducing the practice that helps solve it. It requires carefully prepared cases, realistic enough to expose difficulties without making the exercise hard to follow.
Learn to examine the result
The quality of a request matters. The ability to assess the response matters just as much. Training needs to allow time to compare a result with its sources, look for a counterexample and explain why a proposal should be accepted or corrected.
This includes knowing when to stop. If the available information does not support a conclusion, a well-written answer must not stand in for evidence. If an action involves data or a system the user is not allowed to access, a good prompt does not resolve the permission issue.
Anthropic’s study of learning with AI assistance, published in January 2026, found lower immediate understanding on average in the assisted group. Its context was limited to learning a programming library. It nevertheless highlights a useful distinction: completing a task with assistance and acquiring a skill are two outcomes that need to be examined separately.
Adapt the learning programme to people and their environment
Participants who are new to AI do not have the same needs as those who already use it every day. Bringing them together for an introduction can be useful. Asking them all to do the same advanced exercises is less so.
The available tools, data rules and business constraints also need to be understood before the session. Working on participants’ own tasks is valuable when access and confidentiality conditions allow it. In other cases, prepared examples can preserve the difficulties of the work without exposing sensitive data.
The organisation also needs to make room for applying what people learn. A team trained to check results carefully, then assessed solely on output volume, receives two conflicting messages. Managers have a role in setting quality criteria, allowing time and making it possible to ask for help.
Check what remains after the exercise
An assessment can introduce a new case similar to everyday work. The participant needs to produce a result, explain the information used and show their checks. They may also be asked to examine a deliberately imperfect result.
The level of assistance depends on the skill being assessed. When learning to work with AI, the tool has a place. To assess understanding of a decision, an explanation in the participant’s own words or an unassisted question may reveal more.
A supported learning programme then provides opportunities to revisit practical application. What have participants used again? Where did the work get stuck? Does the difficulty lie in skills, data, the tool or the way the team works? The answers guide further learning.
Our Train your teams to work with AI offer follows this progression: start with situations from the work, practise, examine the results and revisit the task. Training can then connect with the work teams actually need to move forward.