One checklist for three kinds of tool
Conditional logic reads the first task and adds the oversharing, coding assistant and AI literacy tasks only where they apply. Template versioning ties each rollout to the version it ran.
AI Tool Rollout & Governance Checklist Template
Staff were using AI tools long before IT chose one. Then the approved assistant arrives and, in its first week, answers a question from a file nobody realised was shared with the whole company.
This free AI tool rollout checklist is for IT managers, security leads and MSPs bringing a generative AI tool into an organisation: an office-suite assistant, a chat assistant or a coding assistant. It runs from the approved-tools list and shadow AI, through vendor data terms, an oversharing check, labels and DLP, SSO, licences and logging, to the acceptable use policy, training and a pilot that ends in a named go-live decision.
An AI Rollout Is a Software Rollout With Three Extra Questions
Most of an AI rollout is ordinary implementation work: single sign-on, licences, training, a pilot and support. Three questions are new. Where do prompts, files and answers go once staff type them in? What can the assistant read on each person’s behalf? And which AI tools are staff already using without asking?
The second question catches most organisations out. Microsoft’s documentation says Microsoft 365 Copilot only surfaces organisational data that each user has at least view permission for, and Google says Gemini in Workspace follows existing Workspace permissions. Neither breaks your permissions, but both make them visible. A salary spreadsheet shared with everyone years ago was hard to stumble on. An assistant asked about pay finds it in seconds, which is why the oversharing check comes before licences.
Software rollout
Configure, train, go live
Asks: does it work, and do people use it?
Owned by: the application and business owners.
Template: the New Software Implementation Checklist.
AI tool rollout
Data, permissions and behaviour
Asks: what goes in, what can it see, and how do staff check what comes out?
Owned by: IT and security, with the business owner.
Template: this checklist.
An organisation rolling out a bought assistant is usually a deployer under the EU AI Act, not a provider, and drafting or summarising is not a high-risk use unless the tool is pointed at something Annex III lists, such as screening job applicants. The duty that reaches almost every deployer is AI literacy under Article 4, handled in Phase 5.
What the AI Tool Rollout Checklist Covers
Seven phases take one AI tool from approval to everyday use. Three answers on the first task decide whether the coding assistant, oversharing and AI literacy tasks appear, and the go-live approver signs off before licences go beyond the pilot. Run it once per tool; Phase 7 books a quarterly governance review.
Phase 1
Phase 1: Scope, Owners & Shadow AI
The first task records the three scope answers, the business owner and the go-live approver. They decide what appears in Phases 3 to 6.
- Record the tool, its type and the people who own the rollout — product and plan, office-suite, chat or coding assistant, the business owner and the go-live approver
- Write down what the tool is for and what it is not for — drafting, summarising and code completion in; decisions about people and unreviewed customer-facing content out
- Find out which AI tools staff already use — sign-in logs, expense claims, browser extension inventories and a short anonymous survey
- Decide the stance on unapproved AI tools — block, allow for public data only, or tolerate until the approved tool is live; say which in writing
- Publish the approved AI tools list — each tool with its plan, approved uses and the highest data classification allowed in it
- Add the tool to the AI system inventory — so the AI Act and ISO 42001 work sees it, with your role recorded, usually deployer
Phase 2
Phase 2: Vendor Terms & Data Use
The full vendor security review runs before this phase. These tasks check only the AI-specific terms, on the plan you are buying.
- Confirm the SaaS security review is complete — certifications, SSO, subprocessors and the data processing agreement, with its conditions carried in here
- Confirm prompts, files and outputs are not used to train the vendor’s models — in the contract for your plan; free and consumer plans often differ
- Record how long prompts and outputs are kept — retention, any human review for abuse monitoring, and how history is deleted
- Record where prompts are processed — regions, any data boundary commitment, and model providers or overflow capacity that sit outside it
- Check who owns the output and what the vendor covers — ownership of generated content and any copyright indemnity, with its conditions
Phase 3
Phase 3: Data & Permissions Readiness
The three oversharing tasks appear only when the tool reads company files, mail or chat. An assistant shows each person what they could already open, so fix permissions before it does.
- Set which classification levels may go into the tool — restricted data usually stays out until the controls below are proven
- Run an oversharing report across sites, drives and teams — links open to the whole organisation, very large groups on sensitive sites, sites with no owner
- Fix or fence the riskiest sites before any licence is assigned — tighten permissions, or hide sites from organisation-wide search and the assistant while owners review
- Ask owners of HR, finance, legal and board sites to confirm who has access — the content an assistant must never summarise for the wrong person
- Apply sensitivity labels to the most sensitive content — labels that encrypt limit what the assistant can use to the rights each person holds
- Set data loss prevention rules for AI use — restricted data types blocked from prompts and from pasting into unapproved AI sites
Phase 4
Phase 4: Identity, Licences & Configuration
The coding assistant task appears only when the tool type is Coding assistant.
- Connect single sign-on and claim the company email domain — staff on personal accounts with a work address sit outside your contract
- Assign licences through a group, starting with the pilot — access follows group membership, and leavers lose it with the group
- Decide each admin setting on purpose — web search, connectors, agents, file upload and chat sharing, each on or off with a reason
- Set the coding assistant’s scope — repositories in scope, the public-code match filter where offered, and secret scanning on those repositories
- Turn on audit logging for AI interactions and set retention — who used the tool, when and which files it read; name the few people who may see prompts
Phase 5
Phase 5: Policy, Training & AI Literacy
The AI literacy record appears only when the EU AI Act is in scope. The training belongs in every rollout.
- Publish the acceptable use policy — approved tools, what data may go in, checking output before use, and what is banned
- Train people on what the tool gets wrong — invented facts stated confidently, out-of-date answers and instructions hidden in untrusted content
- Give higher-stakes roles their own guidance — HR, legal, finance and developers each need examples from their own work
- Record the AI literacy measures for each role — what was delivered and to whom, as evidence of the measures Article 4 asks for
- Give staff one route to report a bad output or a data leak — a service desk category or a form, watched by someone named
Phase 6
Phase 6: Pilot & Go-Live Decision
The audit log check appears only when the tool reads company files, mail or chat. The go-live decision halts the checklist until the approver named in Phase 1 records it.
- Choose a pilot group across roles — a few dozen people is a common size, including a sceptic and someone from a sensitive team
- Take a baseline before the pilot starts — time spent on the tasks the tool is meant to help with
- Log every pilot issue with a decision — wrong answers, missing features and confusing settings, each fixed, accepted or a reason to stop
- Check the audit log for sensitive files reaching the wrong people — answers drawn from HR or finance sites for users outside those teams
- Go-live approver records the decision — Approved or Not approved, with the reason; licences beyond the pilot wait for it
Phase 7
Phase 7: Scale, Measure & Review
- Expand in waves, one department at a time — with the service desk briefed before each wave and the policy sent with the licence
- Reclaim licences nobody uses at 30 and 90 days — seats assigned is not adoption, and AI licences are rarely cheap
- Measure against the baseline at 90 days — time saved and output quality, compared with the pilot figures
- Re-check the vendor terms when the plan, models or subprocessors change — Phase 2 described the tool you bought, not the one it becomes
- Book the quarterly governance review — the approved list, shadow AI findings, incidents and policy changes, on their own schedule
Three Kinds of AI Tool, Three Different Risks
Each kind of tool fails in its own way, which is why the first task asks for the type.
| Question | Office-suite assistant | Chat assistant | Coding assistant |
|---|---|---|---|
| What it can see | Mail, files, chats and meetings the user can open | What staff paste or upload, plus any connectors you enable | Code in the editor and the repositories in scope |
| Main risk | Oversharing made visible | Sensitive data in prompts; personal accounts | Secrets in code; suggestions copied from licensed code |
| First control | Permissions clean-up and labels | SSO, domain claim and DLP | Repository scope and secret scanning |
| Watch in the pilot | Answers built from files the user should not see | Restricted files uploaded | Suggestions merged without review |
Article 4 of the EU AI Act has applied since 2 February 2025. The Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since 27 July 2026, rewrote it: providers and deployers must take measures to support the AI literacy of staff and others using AI systems on their behalf, taking account of their knowledge, experience and the context of use, rather than ensure a sufficient level of it. The duty is softer but has not gone; a record of what each role received is the simplest evidence. This is not legal advice.
Check the logging before the pilot, not after. In Microsoft 365, Purview Audit (Standard) records Copilot interactions automatically when auditing is on, including the files used to answer, and Restricted Content Discovery can temporarily hide chosen SharePoint sites from organisation-wide search and Copilot without changing permissions. Other suites have their own equivalents.
Why Run AI Rollouts in CheckFlow?
Go-live waits for a named decision
The go-live approver is picked in a members field and the decision task is assigned to them. The checklist halts until they record Approved or Not approved, and the activity trail keeps who decided and when.
Evidence the governance work can reuse
Oversharing reports, vendor terms and training records attach to the tasks that produced them. The approved tools list can live in a data set each run reads, and a quarterly recurring schedule opens the governance review.
This checklist rolls out one tool. The EU AI Act Compliance Checklist classifies every AI system you build, buy or use, and the ISO 42001 AI Management System Checklist keeps governance audited year after year. Before Phase 2, the SaaS Procurement Security Review Checklist covers certifications, SSO and the data processing agreement, and the Data Classification Review Checklist makes sure the levels Phase 3 relies on are applied.
If your assistant connects to your process tool, the CheckFlow MCP server lets it start checklists and complete tasks inside processes you have already approved. Its key is bound to a named member, so it reaches no further than that person, and the activity trail records every change it makes. Put it on the approved list like any other tool.
Frequently Asked Questions
What should an AI acceptable use policy include?
The approved tools and plans, the data classifications allowed in each, and the uses that are out of bounds, such as decisions about people without human review. Add the duty to check output before relying on it, when to say a document was drafted with AI, how to report a problem, and that unapproved tools must not be used with company data.
How do you fix oversharing before rolling out Microsoft 365 Copilot?
Copilot only shows each user content they already have permission to see, so the work is fixing permissions nobody knew were wrong. Run an oversharing report, start with links open to the whole organisation and sensitive sites shared with very large groups, and have site owners confirm access. Restricted Content Discovery can keep chosen sites out of Copilot and organisation-wide search while that review runs.
Do business AI assistants use our data to train their models?
Usually not on business plans, but check the contract for the plan you buy. Microsoft states that prompts, responses and data accessed through Microsoft Graph are not used to train the foundation models behind Copilot, and Google says Workspace does not use customer data to train models without permission. Free and consumer plans often have different terms, which is one reason to keep company data out of them.
What does Article 4 of the EU AI Act require when we roll out an AI tool?
Providers and deployers must take measures to support AI literacy among staff and others who operate or use AI systems on their behalf, suited to their knowledge, experience and the context of use. It has applied since 2 February 2025; the Digital Omnibus softened the wording in 2026 without removing the duty. Role-based training with a record of who received it is the practical answer. Confirm your position with your advisers.
How should IT deal with shadow AI?
Find it first, without blame: sign-in logs, expense claims, browser extensions and an anonymous survey show what people use and why. Then offer an approved tool that does the job, publish which data may go where, and block or restrict the rest. Banning everything without an alternative tends to push use onto personal devices, where IT sees nothing.
Is CheckFlow free for this template?
14-day free trial, no card required. The Business plan is $10 per user per month after the trial. Full details at checkflow.io/pricing.