AI Training Programme for a Company: A Four-Week Staff Rollout Plan

Run it as four weekly sessions of 45 to 75 minutes, each ending with a task done on real work: week one rules and safe first tasks, week two each person's own jobs, week three shared prompts and checking, week four review and decide who gets a licence. Most of the value comes from the tasks between sessions, not the sessions.

The common failure is a single big presentation followed by silence. People nod, go back to their inbox, and three weeks later two enthusiasts use the tool and everyone else has forgotten the login. This plan spreads the same effort over a month, keeps each session small, and finishes with a check on what changed, so you know whether to spend more on tools or on training.

Follow me on Instagram@sagnikteaches

The worked example is a small manufacturer with 40 staff: eight in the office (orders, purchasing, accounts, HR), twenty on the shop floor, six in sales and customer service, and six managers. The same plan fits a courier firm, a wholesale distributor or a translation agency; only the practice tasks change. If you want the hours and licence costs first, the time and money an AI staff training takes has the sums.

Connect on LinkedInSagnik Bhattacharya

Before week one: the four decisions the programme depends on

Spend a couple of hours on these before anyone is invited. Skipping them is why many programmes stall in week two, when somebody asks "can I put this customer's spreadsheet in?" and nobody can answer.

Subscribe on YouTube@sagnikteaches

Pick one tool for the whole programme

Training four people on ChatGPT, three on Copilot and two on Gemini means every question gets three answers. Pick the tool most of the team can already reach. If you run Microsoft 365, Copilot Chat comes at no extra cost with eligible business plans, and the paid Microsoft 365 Copilot Business add-on is $21 per user per month on annual billing ($25.20 billed monthly; Microsoft was also running an $18 first-year offer until 31 December 2026). If you run Google Workspace, Gemini is built in, but how much depends on the plan: Google's pricing page showed Starter with Gemini in Gmail and the Gemini app, and Standard adding it in Docs, Sheets, Meet and more. ChatGPT Business and Claude Team are each $20 a seat a month billed annually or $25 monthly, both from two seats. I checked those figures on the vendors' pages on 4 November 2026; confirm them before you commit. For the programme itself, one tool is enough, and staff can compare others later.

Name a programme owner and a data decider

The owner books sessions, collects task results and chases. The data decider is the single person who answers "can I paste this in?" within a working day. In a 40-person manufacturer that is usually the operations manager and the finance lead respectively. Put both names on the rules sheet.

Choose the jobs to practise, not the job titles

List the ten most repetitive writing, summarising and checking jobs across the firm. For the manufacturer the list was: reply to customer order queries, summarise supplier emails, draft purchase-order chasers, tidy meeting notes, write job adverts, turn safety incidents into a report outline, rewrite work instructions in plain language, draft the monthly sales summary, translate customer messages, and check a spreadsheet for odd entries. Every practice task in the four weeks comes from this list. A quick way to build the list: ask each team lead for the three jobs they would most like to hand over, keep only those that recur every week, and cut anything that needs data the rules sheet will forbid.

Set the success measures now

Pick three numbers you can collect without a survey firm: the share of staff who used the tool for a work task in the last seven days, the hours saved on the three most-used tasks (self-reported, rough), and the number of rules incidents (data pasted that should not have been). You will compare them in week four.

Week one: rules first, then three safe tasks

Session length: 75 minutes. Everyone attends, shop floor included, because the rules matter to anyone who has a phone and a customer's details in front of them.

The agenda (75 minutes)

The staff rules sheet

One page, plain words, signed or acknowledged by each person. A version that has worked in small firms:

AI USE AT [COMPANY]: THE ONE-PAGE RULES

1. Use only [tool name], through your work login. No personal accounts for work.
2. Never paste: customer or employee personal details, bank or card details,
   passwords, supplier price agreements, anything marked confidential,
   health or disciplinary information.
3. When you want to use real customer text, replace names, addresses and
   account numbers with placeholders first ([CUSTOMER], [ORDER NO]).
4. AI drafts. You decide. Read every word before it leaves the building, and
   check every number, date and name against the source.
5. If the AI states a fact, a price or a legal point, find it in an original
   document before you rely on it.
6. If you are not sure whether something can be pasted, ask [name] first.
   A wrong guess costs more than a ten-minute wait.
7. Tell your manager when AI helped with something a customer will see.

Questions: [data decider's name], [how to reach them]

Notice what is missing: no ban on trying things out, and no ten pages of policy. If you need the full policy later, checking an AI tool's privacy policy and terms shows what to read behind the tool you chose. Business plans from ChatGPT, Claude, Microsoft 365 Copilot and Gemini in Workspace say they do not train on your business content by default, whereas Claude's own pricing page lists model training on its individual plans as opt-out, meaning it stays on until the user switches it off. That gap is the reason rule one forbids personal accounts.

The three starter tasks

These are deliberately low-risk and quick to check, so people see a win in their first hour.

1. Rewrite a rough email. Paste a messy internal draft (no customer details) and ask for a clearer version in your usual tone.
2. Summarise a long document. Use a public document, such as a supplier's published brochure, and ask for five bullet points and one question to ask the supplier.
3. Make a checklist. Ask for a ten-line checklist for a task the person does weekly, then edit it for accuracy.

A sample prompt and an illustrative reply from the second task, so the facilitator knows what good looks like:

Prompt: Here is a supplier's brochure text. Give me five bullet points a
purchasing clerk should know, then one question I should ask before ordering.
Do not add information that is not in the text.

Reply (illustrative):
- Minimum order is 200 units per colour.
- Lead time is quoted as 4 to 6 weeks.
- Prices exclude packaging.
- A 3% discount applies to prepaid orders.
- Returns need approval within 14 days.
Question: Does the 4 to 6 week lead time start at order or at payment?

What the facilitator should point out: the reply sounds confident but is only as good as the text pasted. The facilitator asks "which of these bullets can you find on the page?" and the group checks. That habit is the heart of week three.

The week-one task

Each person uses the tool for one real work task from the firm's list, writes two lines in a shared sheet (what I asked, what I changed), and posts it by Thursday. It takes about fifteen minutes. Expect around 60 per cent to do it; the rest will do it after a nudge from their manager, and that is fine.

Week two: each person's own jobs

Session length: 75 minutes, in role groups of six to ten. Office staff, sales and customer service, shop-floor supervisors and managers each get a room or a table. The shop floor does not need the same session as the office; give supervisors a shorter practical on work instructions and incident notes.

Run it as a build session, not a talk

Open with five minutes on what people tried in week one. Then each group picks two jobs from the list and builds a prompt that does the job properly. The facilitator walks round, and the rule is that the group tests the prompt on a realistic sample before it is accepted.

Here is a before and after from the manufacturer's customer-service group, working on order-query replies:

BEFORE (what someone typed first):
write a reply to this customer who is angry about a late order

AFTER (what the group agreed):
You are a customer-service assistant for a small manufacturer. Write a reply
to the message below. Tone: calm, direct, no corporate phrases. Structure:
one-line apology, what happened using only the facts I give, the new
delivery date, one thing we are doing. Under 120 words. If I have not given
you a fact, leave [FACT NEEDED] rather than guessing.

Facts: order [ORDER NO] left the factory on [DATE]; carrier delay;
new delivery date [DATE].
Customer message: [PASTE WITH NAME REMOVED]

The second prompt is longer, but the reply needs one edit rather than five. The placeholder rule ("leave [FACT NEEDED]") stops the AI inventing a delivery date, which is the commonest failure in this kind of reply.

What each group builds

• Office and purchasing: a supplier-chaser template and a summary prompt for long supplier emails, with a rule that the summary always lists dates and amounts separately so they can be checked.
• Accounts: a prompt to explain a spreadsheet's variances in plain English, using figures typed in by the user, never uploaded customer ledgers until the data decider has cleared the tool for it.
• Sales and customer service: reply templates for the five commonest enquiries, saved in a shared document.
• Supervisors: a prompt that turns a list of bullet-point steps into a numbered work instruction in plain words, which the supervisor then walks through on the floor to check it matches reality.
• Managers: a prompt for a job advert and one for turning a messy meeting note into actions with owners and dates.

At the end each group posts its two best prompts to a shared folder called "Prompts that work". Name each with the job it does, not the tool: "Reply to late-order complaint", not "ChatGPT prompt 4".

The week-two task and the licence decision

Each person uses one of their group's prompts on at least three real jobs during the week and logs time taken with and without the tool (a rough guess is fine). Those who finish earn a place in the licence decision in week four; the guide on who needs a paid AI licence sets out the test I would apply.

Week three: shared prompts, checking and what goes wrong

Session length: 60 minutes, whole team in two halves if you have more than 25 people. This is the session that makes the programme different from a demo, because it trains people to distrust output sensibly.

Spot the error: an exercise that takes 20 minutes

The facilitator prepares four AI outputs ahead of time, each from a real job on the list and each containing one planted fault. Groups of three get ten minutes to find the faults. For the manufacturer:

The point is not to scare people. It is to let them see, in a safe setting, that fluent text can hide a wrong number. If your team works with spreadsheets, link them to why AI is bad at maths as the follow-up reading.

The checking routine

Teach one routine that fits on a sticky note: numbers, names, promises, then tone. Numbers (prices, dates, quantities) against the source; names (people, products, places) spelt as in the file; promises (refunds, deadlines, guarantees) that the firm can actually keep; tone last, because it is the easiest to fix.

Review the shared prompt folder

Spend fifteen minutes looking at what people have saved. Delete duplicates, rename vague ones, and promote the three best to "Firm standard". A common find: someone's prompt contains a customer's real name in the example. Treat it as a teaching moment, not a telling-off, remove it, and note it in the incident count.

Where it goes wrong in practice

• Personal accounts creep back. Someone's work login is slow, so they use their phone's app. Fix the login problem; do not just repeat the rule.
• The enthusiast does everyone's tasks. One person becomes the team's prompt writer and nobody else learns. Rotate who demonstrates in sessions.
• A manager forwards AI text unread. Make "I used AI to draft this" an acceptable phrase, so people feel free to say it, and also free to say "I checked it".
• The prompt folder rots. Give it an owner and a monthly five-minute tidy.

The week-three task

Each person finds one task in their own week where the AI got something wrong or nearly wrong, and writes two lines about how they caught it. These become the examples for the next induction.

Week four: measure, decide and hand over

Session length: 45 minutes. This is where the owner decides whether to spend money.

Run the three measures

Collect the numbers you set before week one. An illustrative result for the 40-person manufacturer might read:

These are invented numbers to show the shape; your own will differ. Treat self-reported hours with scepticism, and halve them if you need a prudent figure for a budget.

A quick sum for the licence decision

Say 13 people used it weekly and 8 of them saved more than an hour each. At $25 per user per month on a monthly business plan, 8 seats cost $200 a month. If each of those eight saves even one hour a week (about four hours a month) and an hour of their time is worth $20 to the firm, that is $640 of time against $200 of cost. If only three of the eight really save that hour, the benefit is $240 against $200, which is marginal. The decision is to buy for the eight, review again at month three, and cancel any seat nobody opened for 30 days.

Decide what continues

1. Who gets a seat. Those who completed weeks one to three and showed regular use.
2. Who owns the prompt folder. One named person, ten minutes a month.
3. What the induction says. Add a 20-minute "AI at [Company]" slot to every new starter's first week, using the rules sheet and week-three examples.
4. When you check again. Put a 15-minute review in the diary at three months, repeating the three measures.

The check that it stuck

Do this one at six weeks, not in the session. Pick five people at random and ask each to show you, on their screen, the last time they used the tool for a real task. Not whether they use it: the actual last time. If three can show you something from the past week, the habit has taken. If none can, repeat week two for that group with a manager sitting in.

Adapting the plan to other firms

The sequence of rules, own jobs, checking and decision carries across. What changes is the practice material.

• A courier firm works from dispatch notes, delivery exception emails and driver-handbook updates. Week-two prompts focus on summarising exception reports and drafting "your parcel is delayed" messages. Drivers get a short briefing on what not to put in a chat tool, such as recipients' addresses.
• A wholesale distributor trains the buying and customer-account teams on supplier chasers and price-list change notices. The checking exercise uses a price file with one planted wrong unit price.
• A translation agency has the strictest data question, because client documents are confidential. Week one must settle which tools clients allow, and week two practises style-guide prompts and terminology lists on public text only.

If your industry has a regulator or contractual data limits, the rules sheet needs a line about them. For a more detailed example of a single-session version, see the team training session plan for solicitors' firms and the four-week tool rollout for a law firm, which show how a professional firm handles client data.

Running the four weeks on a small time budget

The full plan asks for about 3 hours of session time per person plus roughly 1 hour of tasks across the month. For a 20-person office team that is 80 person-hours, around two working weeks of one person, spread over four weeks. Programme ownership adds about 2 hours a week. If that feels heavy, shorten the sessions to 45 minutes and move week one's pair practice into a take-home; do not drop the rules sheet or the checking exercise.

The cost of tools during the programme can stay near zero. Run the sessions on the business-plan tool you already hold, or on a two-seat trial for the facilitator and the data decider, and buy further seats only after week four. The plan's low spend on software and its high spend on attention is deliberate.

A one-line message for each week

Short reminders keep the programme alive between sessions. Sample wording for the programme owner to send to the whole team:

Week 1: Thanks for coming. Task by Thursday: do one real job with [tool],
note what you asked and what you changed in the sheet. 15 minutes.
Week 2: Use your group's prompt on three real jobs this week and log the
time. Prompts are in the "Prompts that work" folder.
Week 3: Find one time the AI got it wrong or nearly wrong. Two lines in
the sheet, please. Tell us how you caught it.
Week 4: Last session. Bring your time log; we decide who gets a seat.

Handing the programme to the people who run it day to day

By week four the owner of the programme should be able to step back. Give the data decider a short list of standing answers (a "yes, no, ask" list of data types) so the answers are consistent, and keep it in the same folder as the rules sheet. Review it every quarter, because tools change their data terms and new tools appear. If you see staff drifting back to personal accounts, ask what the work account lacks before you tighten the rule.

If you would rather map the first month with someone who has done it for other firms, that is the sort of work my AI implementation consultation covers. The plan above stands alone: a manager with a calendar, a one-page rules sheet and four real afternoons can run it.

Questions owners ask before starting the four weeks

Do I need an outside trainer for this plan?

No. The plan is written so a manager who has used ChatGPT, Claude, Gemini or Copilot for a few weeks can run every session. An outside voice helps in week one if the team is sceptical, but the sessions are practice-heavy, and practice needs your own documents, not a trainer's slides.

How long should each session be?

Seventy-five minutes in weeks one and two, sixty in week three, and forty-five in week four. Keep them in working hours and in the diary a fortnight ahead. Sessions at the end of a Friday, or lunchtimes people give up, quietly tell staff the training does not matter.

What if some staff refuse to use AI?

Ask what they are worried about, because it is usually data, quality or job security, and each has a straight answer in the rules sheet. Do not force daily use. Require only that everyone completes the week tasks once; adoption grows from colleagues showing real time saved.

Should everyone get a paid licence before week one?

Only give seats to people who finish week two. Until then use a shared business plan or the free tier with the data rules applied. Paying for twenty seats that nobody opens after month one is the commonest waste in rollouts like this.

Further reads

• How to Build a Staff Training Matrix With AI — Track who has finished which week of training.
• Is It Safe to Put Customer Data Into ChatGPT? — Settle the data question behind rule two.
• AI Training for Construction Companies: Workshop or Self-Taught? — Workshop or self-taught: a trades view.
• AI Security Risks for Small Businesses and How to Close Them — The wider security risks your rules sheet sits within.
• How to Train Front-of-House Staff to Work Alongside AI — Training staff who work customer-facing, not at a desk.
• AI Training for Marketing and Comms Teams: A UK Buyer's Checklist — A scoring checklist for small UK marketing and comms teams buying AI training: what to ask, what to refuse and how to check it worked.
• AI Training for Pharmacists and Counter Staff: A Two-Session Plan — Two short sessions an owner can run: safe uses first, then hands-on tasks, with a do-not-paste list and a staff sign-off sheet.
• AI Training for School Staff: Session Plan, Tasks and Handout — A ready-to-run one-hour AI session for school staff, with timings, a live demo, data red lines, practice prompts and a handout.
• AI in Employee Training: Turn Your SOPs Into Short Modules and Quizzes — A practical SOP-to-lesson workflow, including quiz checks, manager approval and a small-team completion sheet.
• How to Start a Training Company: Accreditation, Insurance and AI — First-year costs, how to pick an accreditation route, which insurance to price, and the proposal, booking and certificate admin AI can handle.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT Business pricing, Claude pricing, Google Workspace pricing, Microsoft 365 Copilot Business pricing and Microsoft's Copilot privacy documentation (all read 4 November 2026).