Preparing a team to work with AI: from scattered tools to an agreed way of working

In almost every company somebody already uses AI. Usually nothing has been agreed about it: one person drafts emails, another tidies data, a third uses nothing because nobody has said whether it is permitted. That stage is normal, but it is also where a company gets the least benefit and carries the most risk at the same time.

In short

• The first task is not training but agreeing which information may be used. Without it people either use nothing or use everything.
• Train by role. One shared session for everyone gives each person the least.
• The key skill is not writing prompts but checking output: recognising an answer that sounds convincing and is wrong.
• What is worth standardising company-wide: information limits, when a person must read the output, and where AI is not used at all.
• If a task is identical every time, that is automation work rather than training.

Do not start with the tool

The first question is not which tool to buy. It is what information an employee may enter. That is not hard to answer, but in most companies nobody has written it down, so people decide for themselves and each decides differently. A written answer can be short: what is allowed, what is not, and what to do when it is unclear.

That single document does more than any session. It stops being a warning and becomes permission: people start using what is allowed, because they no longer have to guess.

What needs learning, and what does not

Training frequently opens with how models work. In practice that is barely needed. Somebody working with text or documents needs three things: supply their own material rather than ask for a generic answer; say what result is required; and check what came back.

The third is the hardest and the most often skipped. A wrong answer looks exactly as tidy as a correct one — that is a property of the technology, not a defect somebody will fix. So training time is better spent finding errors in answers already received than on writing prompts. That skill is not acquired from theory.

What to standardise company-wide

• Information limits: which data may be used, and where it must not be.
• When a person must read the output before it goes out. Usually that means anything reaching a customer.
• Where AI is not used at all: calculations, totals, legal thresholds. A result can look correct without having been computed.
• Who answers questions after the training. Without that, a team returns to its old habits within a fortnight.
• How to share what works. One person frequently finds an approach that helps everyone.

Leave the rest at role level. Company-wide rules describing exactly how to phrase requests go stale faster than they can be published.

When training is the wrong answer

There are cases where training a team is not the right move. If a task is done identically every time and its rule can be written down, automate it: a rule gives the same result every time, needs no checking and does not forget. Teaching people to do the same thing by hand with a new tool leaves work in place that could have been removed.

The other case: if a system rather than a person has to produce the result, that is software work. Training delivers a skill, not a working solution. These three are worth separating from the outset — training teaches the team, automation removes repetitive work, and building a solution produces the result without a person.

How to tell whether it helped

A general impression tells you nothing. Before the training, agree a few specific tasks people should be able to complete unaided afterwards — using your own material, not a demonstration set. Check again a fortnight later: some skills hold and some do not, and that shows what is worth repeating.

Illustrative scenario, not a description of a client project

Imagine a company running one shared session for all staff. A week later the marketing team uses it daily, accounting does not use it at all, and customer service has sent several emails containing facts nobody verified. The training was not bad — it was simply one session for everyone, so each group got the wrong part of it. This is an example of how we judge scope, not a promise about an outcome.

If you only do one thing, write down the agreement about information. It costs the least and changes the most: after it people stop guessing, and training gains clear limits inside which it can be useful.