Rollout checklist
| Area | Decision to make | Evidence or rule |
|---|---|---|
| Use cases | Two or three tasks with a measurable output | Pick tasks with a clear before-and-after metric |
| Plan type | Business or team plan with admin controls | Claude Team: "No model training on your content by default" |
| Data rules | What may and may not be pasted in | Free tiers may use data to improve products |
| Access | SSO, seat types, who approves new tools | Central billing and SSO on team plans |
| Training | Short course before access, refreshed yearly | EU AI Act Article 4, applicable since 2 February 2025 |
| Risk review | Which outputs need human sign-off | NIST AI RMF: govern, map, measure, manage |
| Measurement | Baseline, then monthly comparison | Time saved, error rate, adoption |
The European Commission states that AI Act AI literacy obligations have applied since 2 February 2025, so EU deployers must support AI literacy among staff who use AI systems.
Plan and data terms to check
Business plans exist mainly for control and data terms, not extra features. Two documented examples, as listed 17 October 2026:
- Anthropic Claude Team: for teams of 2 to 150, USD 20 per standard seat per month billed annually or USD 25 monthly, and USD 100 or USD 125 for a premium seat with five times the usage. The Claude pricing page lists SSO, central billing and "No model training on your content by default." Enterprise is USD 20 per seat per month plus usage at API rates, with a 20-seat minimum.
- Google Gemini API: the Gemini API pricing page marks free tier usage as "Used to improve our products: Yes" and paid tier usage as "No." The same split applies to many tools: free access often means your inputs help train the product.
Our verdict: give staff a paid business plan before they reach for personal free accounts, because once shadow use starts, data has already left. Our AI tool privacy checklist covers retention, training and deletion terms in detail.
A 90-day rollout in eight steps
- Name an owner and a sponsor. One person runs the rollout day to day; one leader removes blockers and signs off on policy.
- Pick two or three use cases. Good first tasks have a clear output and a reviewer: drafting customer replies, summarizing meeting notes, first-pass research, code review comments. Our guide on choosing AI tools helps match tasks to tools.
- Measure a baseline. Record time per task, error or rework rate and volume for two weeks before anyone gets access.
- Write a one-page data policy. List what is allowed (public material, internal drafts), what needs approval (client data) and what is banned (passwords, health records, unreleased financials).
- Run the pilot (days 1 to 30). Give 5 to 15 volunteers access, a shared prompt library and a weekly 30-minute check-in. Log every failure type.
- Train before expanding (days 31 to 60). A short session on what the tool is good and bad at, the data policy, and how to check outputs. The NIST AI RMF Generative AI Profile lists 12 risks, including confabulation, data privacy and information security; pick the three that fit your use cases and train on those.
- Set review rules by risk. Anything sent to customers, regulators or courts gets human sign-off. Internal drafts can go lighter. The NIST AI Risk Management Framework, released 26 January 2023 for voluntary use, organizes this into four functions: govern, map, measure and manage.
- Launch and measure (days 61 to 90). Compare against the baseline monthly. Keep, change or drop each use case based on numbers, not enthusiasm.
What results to expect
The best field evidence shows real but uneven gains. Brynjolfsson, Li and Raymond's NBER working paper studied 5,179 customer support agents and found access to an AI assistant raised issues resolved per hour by 14% on average. Novice and low-skilled workers improved 34%, while experienced, highly skilled workers saw minimal change. The authors also report better customer sentiment and employee retention.
An NBER study of 5,179 support agents found AI assistance raised issues resolved per hour by 14% on average and 34% for novice workers.
Two lessons follow. First, the biggest wins often come from onboarding and junior staff, so include them in the pilot. Second, plan for experts to use AI differently, for drafting or search, rather than expecting the same speed-up.
Worked example: a 40-person marketing agency
- Use cases: first drafts of client reports, meeting summaries and ad copy variants.
- Baseline: a client report takes 3.5 hours on average; 20% need a second revision round.
- Plan: Claude Team at USD 20 per standard seat per month billed annually, 40 seats, so USD 800 a month or USD 9,600 a year at list price, with five premium seats added for heavy users.
- Policy: client names and figures allowed inside the business workspace only; no client data in personal accounts or unapproved tools.
- Training: a 45-minute session plus a prompt library of ten tested templates. Our prompt writing guide is a ready source for those templates.
- Result to check at day 90: if report time falls to 2.5 hours across 60 reports a month, that is 60 hours saved monthly; at the same time, the revision rate must not rise above 20%.
Common mistakes
- Starting with a tool, not a task. Licenses go unused when nobody knows what to do with them.
- Allowing personal free accounts. Data terms on free tiers often allow training on inputs.
- No baseline. Without "before" numbers, you cannot show value or spot harm.
- One training session forever. Tools and models change often; refresh training and the prompt library each quarter.
- Same review rule for everything. Customer-facing outputs need sign-off; brainstorming does not.
- Ignoring EU duties. If you deploy AI in the EU, the AI literacy obligation already applies, per the European Commission's AI Act page.
How we researched this
We read the NIST AI RMF page and Generative AI Profile, the European Commission's AI Act page, Article 4 of the AI Act, Anthropic's Claude pricing page, Google's Gemini API pricing page and the NBER working paper on 17 October 2026. Seat prices are list prices on that date. To compare paid and free tiers further, see free vs paid AI tools, and for the people who design these workflows, the AI workflow design learning path.
Frequently asked questions
How long should an AI pilot run? About 30 days with 5 to 15 people is enough to find the main failure types. Extend it if volume is too low to compare with the baseline.
Do employees need AI training by law? In the EU, yes for deployers of AI systems: Article 4 of the AI Act requires measures to support AI literacy, applicable since 2 February 2025. Elsewhere it is good practice rather than a legal duty in most places.
Is it safe to let staff use free AI accounts for work? Usually not for company data. Free tiers such as the Gemini API free tier state that usage may be used to improve products, while paid business plans typically exclude training.
How much does an AI team plan cost? Claude Team lists USD 20 per standard seat per month billed annually or USD 25 monthly, as listed 17 October 2026. Prices for other vendors vary by seat type and billing period.
What should an AI usage policy include? Allowed and banned data types, approved tools, review rules by risk level, and who to contact about problems. Keep it to one page so people read it.
How do we measure whether AI is working? Compare time per task, error or rework rate and volume against a pre-rollout baseline. Track adoption separately, since unused seats cost money.
Which framework should a small company follow? The NIST AI RMF is voluntary and free, and its four functions scale down well. Use the Generative AI Profile to pick the risks relevant to your use cases.