How to run AI training on your team's real work

The way to run AI training on your team's real work is to arrive with a working draft of one of their own tasks, built beforehand against their actual documents, and spend the hour letting the person who owns that task correct it. The corrections become written instructions, and the owner leaves with something they use the next morning.

Most of the effort sits before the session. The rest of this article gives the mechanics in order: how to choose the task, what to gather, how the hour runs, how to capture what was learned, and how to tell whether it worked.

Why generic AI training does not change how people work

A session on prompt writing with invented examples teaches people about a tool. It leaves them to work out on their own, during a busy week, how any of it applies to the renewal summary or the lease abstract on their desk. Most people do not get to that, and the reason is time.

The Conference Board's 2026 skilling research, which surveyed nearly 1,300 workers worldwide, found that only 48.0% of workers agree their organization provides sufficient time during work hours for AI skills development, and only 47.6% agree they have sufficient tools, access, and resources. The same release advises leaders to focus on developing applied capabilities that improve business outcomes, and to go beyond AI literacy alone.

Training on a real task answers both problems at once. The hour is work time, the output is work product, and the tool is already set up for the job by the time the person sits down.

How do you choose the first task to train on?

Pick a task that passes four tests.

1. It recurs. Weekly is ideal. A task that comes up twice a year gives nobody a chance to practice.
2. It is made from documents. The inputs are files the firm already holds, such as leases, policies, statements, proposals, or call notes.
3. One named person owns it. That person does it today and can say what a good version looks like.
4. It has a known right answer. You can put a draft next to a finished version and see where they differ.

A lease abstract, a policy comparison, a recurring client report, and the first draft of a proposal all pass. A strategy memo fails the fourth test, and a task shared loosely among six people fails the third.

Ask the owner directly which part of their week they would hand off if the result were dependable. The answer is often a better first task than anything a planning meeting produces. For more on choosing the task and building the first version, see how to build an AI workflow.

What should you prepare before the session?

Plan to spend more time preparing than the session itself takes. Gather these before you build anything:

• One past example with its source documents. Choose a case that is finished, where the correct output already exists and went out the door. Pick a typical one for the first session and save the awkward ones for later.
• The finished output for that example. This is your answer key. Do not show it to the AI while you build the draft.
• The firm's own template. The draft should come out in the firm's format, column order, and naming, so the owner is judging substance and has nothing to translate.
• Fifteen minutes with the owner. Ask what they check first when a junior person hands them this work, and what the common mistakes are. Write the answers down as plain instructions.

Then build the draft. Put the instructions, the template, and the source documents into the firm's AI workspace and run the task. Compare the result with the answer key yourself and fix the obvious problems. You want to walk in with a draft that is mostly right and wrong in ways only the owner would catch, because those are the corrections worth an hour of their time.

Which documents should you use, and where should they live?

Use real documents, and use them only inside the AI workspace the firm has approved for client work. That means the firm's own account, on the plan and settings someone at the firm has reviewed. Personal accounts and free tools stay out of the session, including for people who join remotely.

Before the first session, confirm with whoever handles compliance at your firm that the documents you chose may be used this way under your client agreements. Our guide to secure AI workflows for confidential client data covers what to check, and it is a matter to settle with your own counsel or compliance lead.

Keep everything for the task in one shared location: the instructions, the template, the example, and the answer key. If the materials live in the trainer's own files, the training ends when the trainer is busy.

The one hour working session, step by step

Minutes	What happens
0 to 5	State the task, name the owner, and show the source documents. Say plainly that the draft is expected to be wrong in places and that finding those places is the point of the hour.
5 to 15	Run the task live from the source documents so everyone sees the steps. Put the draft on screen beside the finished version.
15 to 35	The owner reads the draft aloud against the finished version and marks every difference. The trainer writes each correction down in the owner's words. Nobody fixes anything yet.
35 to 50	Turn the corrections into instructions, rerun the task, and compare again. Repeat once if time allows.
50 to 60	The owner runs the task themselves on a second, live piece of work. Agree what they will run it on this week and who reviews the output.

Two rules keep the hour honest. The owner does the correcting and the trainer does the typing, because the owner is the one who knows the work. And the last ten minutes are protected, because a session where the owner never runs the task alone has produced a demonstration.

Who should be in the room?

Keep it to three to five people: the owner, the person who reviews the owner's work today, one or two colleagues who do the same task, and the trainer. The reviewer matters because they will be asked to trust the output later, and they trust it more readily when they watched it get corrected.

BCG's 2025 survey of more than 10,600 leaders, managers, and frontline employees in 11 countries and regions found that regular usage is sharply higher for employees who receive at least five hours of training and have access to in-person training and coaching. A small group around one task is a format where that kind of coaching happens as a matter of course.

How do you turn corrections into instructions the whole team can use?

Write each correction as a plain rule, in the order the owner raised it, with the reason attached. "Report the rent commencement date and the lease commencement date separately, because the two differ whenever there is a free rent period" is a usable instruction. "Be more careful with dates" gives the next person nothing to act on.

Save the rules in the shared instructions for that task, alongside the template and the example. The next person who runs the task gets the owner's standard without having attended the session. OpenAI's guidance for internal AI champions makes a related point. It contrasts a prompt that helps one person complete one task with a repeatable process a team can follow, supported by examples that are specific, practical, and connected to real work.

There is some evidence that capturing an experienced person's standard helps newer staff most. A study of 5,179 customer support agents found that access to a generative AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal impact on experienced and highly skilled workers. The authors report suggestive evidence that the tool spread the practices of more able workers to newer ones. That study covered customer support, so treat it as support for the design principle and not as a forecast for your firm.

Where does a person review the output?

Review stays where it sits today. If a senior person signs off on the abstract or the report now, they sign off on the AI-assisted version too, and the instructions should say so by name.

Make the review easier by having the draft show where each figure came from, such as the page and section of the source document. A reviewer who can check a number against its source in seconds will keep checking. We cover the design choices in how to design the human review step.

What happens between sessions?

The owner runs the task on live work and keeps a short list of what the draft got wrong. The trainer updates the instructions from that list and prepares the next task the same way: one past example, one answer key, one working draft.

The following session opens with ten minutes on last week's task, to confirm it is still in use and to fold in the new corrections, and then moves to the new one. This is the same one hour a week pattern we use in our own engagements, where most of the effort happens between sessions. A firm can run it with its own people if someone has the hours to prepare, and you can get in touch if you would like help with that part.

How do you know the training worked?

Check three things two weeks later, all of which you can see without a survey.

1. The owner is running the task on live work without the trainer present.
2. A second person has produced an acceptable draft using only the saved instructions.
3. The reviewer's corrections on new drafts are getting shorter, and each new one has been added to the instructions.

If the first is true and the second is not, the instructions are still in the owner's head and need another pass. Counting logins or attendance tells you much less than these three checks, which we explain in how to measure AI adoption on a team.

Common questions

How long should an AI training session be?

One hour is enough when the trainer arrives with a working draft, because the time goes to correcting real output and none of it goes to setup. A longer session usually means the preparation was skipped and the group is building from a blank page.

How many people should be in an AI training session?

Three to five works well: the person who owns the task, the person who reviews their work, one or two people who do the same task, and the trainer. A larger group turns the session into a demonstration, because only one person can correct the draft at a time.

Is it safe to use real client documents in AI training?

Use real client documents only inside the AI workspace your firm has approved for client work, under the plan and settings your firm has reviewed. Check the vendor's data terms and your client agreements with your own counsel or compliance lead before the first session, and keep personal accounts out of it.

What is a good first task to train a team on?

Choose a task that recurs at least weekly, is produced from documents the firm already holds, belongs to one named person, and has a known right answer you can check a draft against. A lease abstract, a renewal summary, or a recurring client report all fit.

How often should a team hold AI working sessions?

Weekly, one task at a time, is the rhythm we use. A week gives the owner enough live runs to find what the draft still gets wrong, and it gives the trainer time to prepare the next task properly.