Guide
Rolling out AI in your company: four weeks with one team
One team, two recurring tasks, four weeks. At the end you'll know, measured on your own cases, whether AI saves your team time on those tasks.
Many AI rollouts start with accounts for everyone and a training session. Three months later, management asks what it achieved. If nobody counted anything beforehand, the only answers are impressions.
This plan reverses the order. One team tests two tasks it does every week anyway, for four weeks. What counts as success is settled before the start. The plan suits teams of about three to fifteen people.
In short
- Measure the tasks without AI first, or you’ll have nothing to compare against.
- Count review time. A fast draft that takes ages to correct saves nothing.
- Write down the decision rule before you see the first results.
Week 0: choose the team and tasks
A good fit is a team that writes many similar texts and has someone who can judge the results. Choose two tasks that come up at least ten times a month. For example:
- drafting replies to customer emails, as in our customer email prompt template
- translating quotes or customer letters, as in our guide to translating documents with AI
- answering questions about an internal policy, as in using AI with your own documents
Then measure ten cases of each task without AI. How long does it take to get a result that’s ready to send, and how often does someone have to rework it? The UK government’s evaluation guidance for AI projects describes this as a baseline, collected before rollout starts, and uses it to estimate impact.
Week 1: rules and introduction
Before the team starts, it needs two things: a short AI policy that says what data can go into the tool, and an introduction of about an hour. Work on the real pilot tasks there, not general examples. Note who attended. In the EU, the AI Act requires measures on staff AI literacy. According to the European Commission, you don’t need a certificate, and you can keep an internal record of training.
Weeks 2 and 3: work and log
The team uses AI for every case of the two tasks. One line per case is enough:
| Date | Task | Minutes to ready-to-send | Reworked? | Error in draft |
|---|---|---|---|---|
| 17 Nov | Customer email | 6 | yes, date corrected | wrong date |
| 17 Nov | Translation | 9 | no | none |
These lines show the format, not measured values. The “Error in draft” column matters most, because it shows whether the errors are harmless (a comma, the tone) or dangerous (an invented commitment, a wrong price). If dangerous errors appear, change the prompt or the documents straight away and note the change in the table.
Week 4: decide
Set the rule back in week 0. For example:
- Continue if handling time including review is below the before measurement in at least half of the cases, and no dangerous error reached a customer.
- Adjust if time drops but dangerous errors are common in the drafts. Improve the instructions or documents and test for two more weeks.
- Stop if review takes so long that no time is saved.
The thresholds are suggestions. Pick values that suit your company. What matters is that they’re fixed in advance.
What you have afterwards
After four weeks you have a comparison based on your own cases, tested instructions for two tasks, and a policy that has been through everyday use once. That’s enough to bring in the next team. If you’d like help, onAI offers workshops for choosing use cases and training for teams.
Write down today the two tasks your team does most often. That’s week 0 half done.
