A 30 day rollout runs in four weeks. In week one you choose a business plan and agree a one-page policy. In week two you train people on three real tasks from their own work. In week three you share the prompts that worked and name a champion. In week four you check what stuck and stop what did not. The tool is the easy part. The habits, and the rules around them, are the work.
We are rolling out Claude with training for a construction firm in Hertfordshire, and this is the shape I would use for any team of 10 to 50 people. It works for ChatGPT as well, so pick one and run the plan with that.
Week 1: choose a business plan and agree the policy
Begin with the account, because it decides what happens to your data. Both OpenAI and Anthropic sell plans for organisations, and at the time of writing both are on public pricing pages. ChatGPT Business lists central administration, single sign-on (SAML) and multi-factor authentication. Claude Team lists single sign-on, central billing and administration, shared projects and no model training on your content by default. OpenAI says it does not train on business customers’ data by default either. Check both pages for current prices and seat rules before you commit, because they change.
Personal accounts sit outside those admin controls, and their data terms differ. Our guide to where your data goes in ChatGPT and Claude sets out the difference.
Then do four things:
- Pick one tool for the 30 days. Running two at once splits the team’s attention and the prompts you share later.
- Set up the workspace. Invite people with their company email. If you already use Google Workspace or Microsoft 365 as your sign-in, turn on single sign-on. Choose two admins.
- Hold a ten minute amnesty. Ask which AI tools people already use for work and what for. Make it clear nobody is in trouble. You will learn which tasks matter most, and you will find the personal accounts you need to close.
- Agree the one-page policy. Our one-page AI policy template covers approved tools, what data never goes in, checking outputs, who to ask and customer-facing use. Fill in the names, have the owner sign it, and send it before anyone gets access. If you also want a view of the wider picture, our AI governance guide goes further.
Week 2: train the team on three real tasks from their own work
A generic demo is easy to nod along to and hard to apply on Monday. So ask each person, a few days beforehand, to bring three tasks from their last month that were repetitive or slow, with a real document or message for each. Strip anything that falls under the policy’s red category first.
Run one session of about 90 minutes, in groups of five to eight. Spend the time like this:
- Twenty minutes on how to ask. Show a vague request and a good one. A good one says who the output is for, what you want, what material to use, what format, and what to avoid. That is all prompting is.
- Forty minutes doing their own tasks. Each person works through one of their three with the tool while someone walks the room. People learn faster on a task they understand.
- Twenty minutes on checking. Have everyone ask for a summary of a document they know well, then find what it got wrong. Expect someone to find something. This is the moment the policy line “you are responsible for what you send” becomes real.
- Ten minutes on what is off limits. Walk through the green, amber and red data rule once, aloud.
Some examples of tasks that suit this. An office manager might use it to draft chaser emails to suppliers, turn site notes into a weekly report, and write a spreadsheet formula. An estimator might draft a quote reply from an enquiry and a price list, summarise a long specification, and compare two supplier quotes in a table, then check every number. A director might ask for a plain summary of a monthly report. We cover more of these in AI for small businesses, including where it needs a human check.
The British Chambers of Commerce’s 2026 AI research notes that the risk of AI mistakes can be mitigated by training staff on how and when to use AI, and on not relying on outputs as gospel. The checking exercise is that training in miniature.
Week 3: shared prompts, shared projects and a named champion
By now each person has found one or two prompts that work. Week three is about not losing them.
Collect the prompts. Ask everyone to send the best prompt from week two to one place. A shared document is fine. Keep the ones that are specific to your business, such as your quote reply, your site report layout and your supplier chaser, and drop the generic ones.
Use shared projects. Claude Team lists project sharing, and ChatGPT Business supports shared projects too. A project holds instructions and reference files, so you can store your price list, standard terms and tone of voice notes once, and everyone starts from the same context instead of pasting it each time. Keep anything in the policy’s red category out of it, and decide who may edit a project and who may only use it.
Name a champion. One person, by name, for the next three months. Pick someone curious and respected, not necessarily the most technical person in the building. Their job is small and specific: keep the shared prompts tidy, answer questions, run a 30 minute office hour each week, and tell the policy owner about anything odd. Give them protected time, around two hours a week. A champion with no time is a title.
Week 4: check what stuck and decide what to stop
On day 25 or so, talk to every person for ten minutes. Ask four questions:
- Which task do you now use it for every week?
- Where did it save time?
- Where did it create rework, or make you uneasy?
- What would you like to try next?
If your admin console shows usage, look at it; ChatGPT Business lists usage analytics on its pricing page. A conversation tells you more than a dashboard, though, because a heavy user may be producing rework and a light user may have found one very good task.
Then sort every use into one of three piles. Keep: it saves time and the output is checked; write it into the shared prompts. Fix: it is promising but people are unsure how; give it a short follow-up session. Stop: it creates more checking than it saves, or the risk is too high. Stopping is a result, not a failure. A team with four reliable uses is better off than a team with fifteen uncertain ones.
Finish by reading the policy again. Did anything come up that it did not cover? Update it, change the version number and set the next review date.
What comes after day 30
There are two sensible next steps. One is a second round: new tasks, new people, the same four weeks. The other is connecting the tool to your own systems, so a first draft of a quote or a coded supplier invoice arrives without anyone pasting anything. That is engineering work, and it needs clear rules about what the system may do on its own. Our AI consultancy page covers how we work.
If you would like help with the first 30 days, book a 30 minute call. We can run the training session with your own tasks, help you set up the workspace, and draft the policy with you.