How to Start AI Training at Work: A 30-Day Pilot

By James Finley. Published Oct 24, 2026.

A common way to start AI training is to buy licenses, send a link and hope for the best. A few months later, a few people use the tool every day, others have tried it once or twice, and nobody can say whether it helped.

There's a better first step, and it's smaller. Pick one task, train one small group, and measure what changes over about 30 days. Then decide, with real numbers, whether to expand it, fix it or stop. The rule behind it: start small, measure, then decide.

Step 1: Pick one task, not "AI for everyone"

A good first task passes five checks:

Frequent. It happens many times a week, so small savings add up and people get practice.
Safe data. It works with only the information your approved tool is allowed to hold.
A person checks first. Someone reads the AI's work before it reaches a customer or a decision.
Not a people or money decision. Hiring, pay, discipline, credit and discounts stay with people.
Countable. You can time it or count it before and after.

Replies to common customer questions often pass all five. "Use AI to decide which late-paying customers go on credit hold" fails the fourth check, so it isn't a first task, however tempting.

Step 2: Use work accounts, not personal ones

Running the pilot on personal accounts puts company information into tools nobody at your company controls. Use a business plan, approved by whoever decides which tools your company uses.

The major providers say their business plans don't train their models on your company's data by default. OpenAI says this for ChatGPT Business and Enterprise. Microsoft says it for Copilot prompts and responses at work (the product formerly called Microsoft 365 Copilot). Anthropic says it for its commercial Claude products. Google says Workspace doesn't use your data to train its AI models without your permission. Ask whoever manages your accounts to confirm the setting for your plan, because plans and terms change. Also ask whether the feedback buttons (thumbs up or down) send chats to the provider.

Ask one more question before you start: can anyone at your company read people's chats, for example through an export or an admin log? People will ask you. Have the real answer.

Step 3: Clock it before you change it

Before anyone uses AI on the task, time how long it takes today. That's your baseline.

Define the task so everyone times the same thing, including looking things up and checking.
Time 6 to 10 real examples per person, spread over several days of a normal week.
If you can, also time someone who won't be in the pilot. They're your comparison.
Don't ask people to guess hours. A guess isn't a measurement.

Tell people what you're timing and why. Share group averages only, and never use timing sheets in anyone's review.

Step 4: Write a one-page pilot plan

Keep it to one page:

Who: 5 to 8 people who do the same task. Include at least one enthusiast and one skeptic.
What: one task (two at most).
How long: about 30 days, with a 15-minute check-in each week.
Finish line: decided before day 1, not after. Use three measures: time per task including checking, quality (how many drafts needed a big fix), and people (would they keep using it).
Stop rules: for example, "an AI draft reaches a customer without a person reading it" or "confidential information goes into a tool not approved for it." Either one pauses the pilot right away.
Roles: who leads, who approves, who sets up accounts, and who to tell when something goes wrong.

Step 5: Train on the real task, in short sessions

Long lectures are easy to forget. What tends to stick is short, hands-on practice on the task people will do:

Three 30-minute sessions: one before day 1, one in week 1, one in week 2.
Practice on made-up examples first, then real ones.
Teach the check, not just the prompt. Every session ends with comparing a draft against the source and fixing it.
Give people a prompt card: one copy-ready prompt for the task, with blanks to fill in, kept where they work.

Pick a champion: a coworker who learns early and helps others. A champion answers questions and shares good prompts. A champion doesn't police who's using the tool.

Step 6: Be honest about worries

People will ask: "Is this how you replace me?" "Are you watching what I type?" Answer only with what has been decided. If the owner has decided no one's job changes during the pilot, say so in those words. If nothing is decided about later, say that, and say when people will hear.

Skip the hype. "10 times faster" makes skeptics more worried, not less.

Step 7: Measure fairly, then decide

At the end of the pilot, compare the same task, timed the same way:

Include checking and fixing time. Timing only the drafting makes the gain look bigger than it is.
Subtract what would have happened anyway. If the person outside the pilot also got faster, part of the change isn't the AI.
Treat the tool's own "time saved" figure as an estimate, not a result.
Don't stretch a small test into a yearly number. Say "about," and say how many people and weeks it covers.

Then label each task:

Go: it met the finish line. Expand it to the next group, one group at a time, with training before access.
Adjust: worth doing, but one named thing must change first, like more examples or a better check step. Change it and test again.
Stop: it missed the finish line after a fair test, or it failed one of the five checks or hit a stop rule. Write down why, remove the access, and keep what you learned.

What to do this week

List three tasks your team repeats every week, and run each through the five checks.
Find out who at your company approves AI tools and who sets up accounts.
Time one task for a few days, the same way each time.

If you'd like a guided version, our Leading an AI Rollout course walks through every step with a fictional company and builds a Rollout Kit you can reuse: a pilot request, timing sheet, pilot plan, announcement and results report. If your company doesn't have an AI policy yet, start with Write Your Company's AI Policy. Both have a free preview lesson, with no account needed. Training a whole team? See Teams plans.