Journal / Kirill Naumov
AI in Workplace Learning: Practise the Decision, Not Just the Lesson
How to use AI in corporate training to practise real decisions, coach employees safely and check whether learning transfers to work.
By Kirill Naumov ·
A completed course is not the same thing as a capable colleague
A learning platform can tell you that someone opened a module, watched a video and passed a quiz. It cannot, by itself, tell you whether that person can make a sound decision when a customer is waiting, the policy is unclear and the manager is busy.
That distinction matters when companies introduce AI at work. The tempting first move is to produce more learning content: a chatbot, a library of explainers, perhaps a course with “AI-powered” printed on the cover in a font large enough to be seen from procurement. The more useful question is smaller: what should a person be able to do differently after the learning?
The answer might be to check a generated product claim against an approved source, ask a better discovery question, coach an employee through an objection, or recognise when a request must be escalated. Start there. The technology is a method; the capability is the point.
Put AI inside a practice loop
AI can help make practice more available and varied. An employee can rehearse a conversation, receive feedback against a rubric and try again with a different customer situation. A manager can use an approved assistant to turn a policy into short, role-specific practice questions. A learner can ask for a plain-language explanation of a concept before applying it to a case.
Those uses are helpful only when the learning loop has four parts: a real work decision; a bounded scenario with no real customer details or confidential data; feedback that explains what was noticed, missed and worth checking; and a later transfer check at work by a manager or coach.
The UK Government’s employer guide to AI upskilling recommends connecting learning to real tasks, providing relevant practice and making responsible-use boundaries clear. The OECD notes that AI adoption increases demand both for specialists and for workers with a general understanding of AI. Neither source says that a chatbot automatically makes training effective. That part still depends on the task, the content and the feedback.
Use role-play for judgement, not for theatrical confidence
AI role-play is easy to make entertaining. It is harder—and more useful—to make it fair. A simulated customer can be impatient, confused or sceptical, but the exercise should test a skill the organisation actually values. Otherwise, learners end up practising how to satisfy the simulator.
For example, a sales-practice exercise could ask an employee to understand a buyer’s needs before recommending an option. The scoring guide might ask whether the employee asked a relevant follow-up question, used approved product information, checked a claim rather than treating fluent AI output as evidence, noticed uncertainty and respected the customer’s choice without manufacturing urgency.
Have a subject-matter expert review the scenario and rubric. Test the exercise with a small group. If the AI rewards the wrong behaviour, fix the exercise before you call it a learning intervention. A confident score generated by a poorly designed rubric is still a confident score generated by a poorly designed rubric.
Keep a human accountable for the learning environment
An AI tutor may suggest an explanation, generate a scenario or offer practice feedback. It should not quietly become the owner of company policy, performance ratings or decisions about an employee’s future. The NIST AI Risk Management Framework emphasises clear roles, appropriate training and human responsibility for managing AI risks.
For a learning pilot, make the division of work visible. The business owns the standard. A qualified person checks the examples and answers. The learner is told when AI is being used and how to challenge a result. A manager remains responsible for coaching and escalation. The tool may help deliver practice; it does not get to invent the rules of the job.
Also decide what the system must not receive. Use fictional or approved data, apply the organisation’s privacy and security rules, and make it possible to report a bad answer. If a learner cannot tell whether a response is grounded in policy or merely sounds plausible, the design needs work.
Measure capability before celebrating completion
Choose a baseline before the pilot. Then look at a few different signals rather than one convenient dashboard number: can employees complete the target task to the agreed standard; do they verify information and recognise when to escalate; is the customer-facing explanation clearer; does the task take less effort without creating extra checking elsewhere; and do managers see the behaviour in real work?
The right measures depend on the task. A short course-completion rate may help with administration, but it is not proof of transfer. Compare similar situations, review examples with employees and managers, and be honest about what a small pilot can—and cannot—show. If the process changes halfway through, record that too. Otherwise, the evaluation becomes an impressive-looking comparison of two different things.
A practical first month
Week 1: choose one capability. Talk with employees and managers to find a recurring task where confidence, quality or consistency needs improvement. Week 2: write the standard, safe data boundaries, approved sources and the point at which a person should take over. Week 3: run two or three realistic practice scenarios with a small group. Week 4: sample the real task, speak with managers and learners, then decide whether to improve, expand or stop.
The aim is not to make every employee an AI specialist. It is to help people build sound judgement for the work they already do—and to give managers a practical way to support that change. The test is not whether the course is clever. It is whether the work gets better.
Explore the English AI for Retail workshop → · Discuss an AI workshop for your leadership team →
Further reading: UK Government employer guide to AI upskilling · OECD on the AI skills gap · NIST AI Risk Management Framework.