Short answer
A safe AI rollout gives a team AI tools in stages, with the rules decided before anyone starts: which business accounts are approved, which uses are allowed, what data stays out, who pilots first, and what would make you pause or roll back. For a New Zealand business the plan fits on one page and is reviewed at each gate, not filed and forgotten.
- Decide approved uses, red lines and data rules before handing out accounts.
- Pilot with a small mixed group, then widen only after a review gate.
- Write stop conditions in advance, as MBIE’s quickstart guidance suggests.
- Use business accounts, not personal ones, for any work material.
- If roles or duties change, NZ employment law expects consultation first.
- Goal
- Draft a one-page plan for introducing AI tools to your team in stages, with the rules and stop conditions written before anyone starts.
- Who it is for
- Owners, operations managers and team leads in New Zealand businesses of roughly 5 to 100 people.
- Time
- 30 minutes.
- You need
- ChatGPT, Claude, Copilot or Gemini, plus 10 minutes of thinking about which tasks and people you would start with.
- You will produce
- A one-page rollout plan with phases, gates and stop conditions, ready to share with your team.
What makes an AI rollout safe?
A safe AI rollout is one where the hard decisions are made before the accounts go out: what the tools are for, what they must never touch, who goes first, and what would make you pause. It is staged, so a problem shows up in a pilot of six people rather than across the whole business.
MBIE’s quickstart for businesses using AI tools puts the two core ideas plainly: “Start small and experiment”, and “Think about what you will do if something goes wrong or you want to stop using the AI tool.” Most rollouts we see skip the second one.
What goes in a one-page rollout plan?
Nine short sections cover it. Each answer should be a sentence or two, specific enough that a new staff member could follow it.
| Section | The decision you are making | Illustrative answer |
|---|---|---|
| Scope | Which tools and account types are approved | One business AI subscription, company accounts only |
| Approved uses | Tasks people may use it for now | Drafting emails, summarising internal reports, meeting notes |
| Red lines | Uses that stay off limits | Client fee decisions, HR matters, anything sent unreviewed |
| Data rules | What may and may not be entered | No client personal information until the privacy review is done |
| Pilot group | Who starts, by name | Six people across admin, engineering and accounts |
| Phases and gates | When the next group starts, and on what evidence | Review at week 5; widen only if gate criteria are met |
| Stop conditions | Events that trigger a pause | Personal information in an unapproved tool; an AI error reaching a client |
| Support | Who answers questions and logs problems | One named champion, one shared channel |
| Communication | What staff are told and when they can comment | Plan shared two weeks before the pilot, with a feedback date |
How do you set gates and stop conditions?
A gate is a scheduled decision point with criteria written in advance. A stop condition is an event that pauses the rollout immediately, whatever the calendar says. Both need an owner.
Good gate criteria are things you can check in a 30-minute meeting: no unresolved data incidents, pilot users can name at least one task where the tool helps and one where it does not, the support champion has answered the common questions, and costs are inside budget. Good stop conditions are concrete events, not feelings. If a condition cannot be checked, rewrite it until it can.
The Privacy Commissioner’s expectations for generative AI fit naturally into gates: senior leadership approval, a privacy impact assessment before personal information goes in, and human review of outputs before acting on them.
Exercise: draft your plan with AI
Use the fictional fact sheet below, or replace it with notes about your own business. Keep real client names and staff details out of personal or free AI accounts.
Illustrative fact sheet. Fictional business.
Business: Harakeke Civil Ltd, civil engineering consultancy, Hamilton
Staff: 22 (14 engineers, 4 drafters, 2 admin, 1 accounts, 1 director)
Systems: Microsoft 365, Xero, a project folder on SharePoint
Current AI use: several engineers use free AI tools on personal accounts
Wanted: faster report drafting, meeting notes, tidier client emails
Worries: client confidentiality, wrong figures in reports, cost
Constraints: reports are signed by a chartered engineer; client
contracts include confidentiality clauses
Budget: up to $1,500 (NZD) a month for tools
- Paste the fact sheet and the worked example prompt into ChatGPT, Claude, Copilot or Gemini and run it.
- Check every red line. Ask yourself whether each one matches a real risk in the fact sheet. Add any that are missing, such as signed engineering reports.
- Tighten the stop conditions. Rewrite any condition that cannot be checked as a yes-or-no event, and add who decides to pause and restart.
- Name the pilot group. Replace role descriptions with real names in your own version, and make sure the group includes at least one sceptic.
- Set the gate date and criteria. Put the first review in the calendar now, four to six weeks after the pilot starts.
- Plan the conversation with staff. Decide when the plan is shared, how people can comment, and whether any duties change. If they do, talk to staff before deciding.
Worked example: the planning prompt
You are helping a New Zealand business plan a staged rollout of a
business AI assistant. Using the fact sheet below, draft a one-page
rollout plan with these headings: Scope, Approved uses, Red lines,
Data rules, Pilot group, Phases and gates, Stop conditions, Support,
Communication.
Rules:
- Be specific to this business. No generic advice.
- Every stop condition must be an observable event, with who
decides to pause and who decides to restart.
- Gate criteria must be checkable in a 30-minute meeting.
- Flag anything in the fact sheet that needs a decision by the
director before the pilot starts.
[paste fact sheet]
Example output (illustrative and abridged; your tool’s wording will differ):
Red lines
- No AI-drafted figures, calculations or design statements go into
a report without check and sign-off by the chartered engineer.
- No client documents in personal AI accounts. Free accounts are
retired at the start of the pilot.
Stop conditions (pause decided by the director; restart by the
director after review with the champion)
- Client-identifiable information found in an unapproved tool.
- An AI-introduced error found in anything sent to a client.
- Monthly tool cost exceeds $1,500 (NZD).
Decisions needed before the pilot
- Confirm the client contract clauses allow use of an AI service
that processes data offshore.
The useful part is the last block. A good draft surfaces the decisions you have not made yet.
Where this goes wrong
- Personal accounts carry on. If staff keep using free tools alongside the approved one, your data rules apply to half the work. Retire them on day one and say why.
- The pilot is all enthusiasts. Results look great and the questions the rest of the team will ask never come up.
- Roles change without consultation. Employment New Zealand says changes that affect employees’ jobs need genuine business reasons, consultation and a fair process. Improved technology can be a genuine reason, but it does not replace the process.
- No stop rule. Without one, a serious incident turns into a debate. The NCSC’s AI guide for small businesses treats data leaks and unreliable outputs as core risks; plan for both.
- The plan is never reopened. Bring it to every gate meeting and update it. Record risks in an AI risk register.
Check your work
- My plan fits on one page and covers all nine sections.
- Every red line matches a real risk in my business.
- Every stop condition is an observable event with a named decision-maker.
- The pilot group is named and includes at least one sceptic.
- The first gate meeting is in the calendar with written criteria.
- I know whether any roles or duties change, and how staff will be consulted.
Where to go next
Next in the pathway: measuring adoption after training, which shows how to tell whether the rollout is working. The plan’s rules belong in a standing policy; our lessons on creating a team AI use policy and rolling out AI to a small team cover that side, and the AI use policy template gives you a starting draft. For what to do when a stop condition fires, see our engineering guide to AI incident response. Larger organisations with several teams to bring on board can use our AI governance advisory.
Sources
- Quickstart: Businesses using AI tools, Ministry of Business, Innovation and Employment, Accessed 19 October 2026
- Responsible Artificial Intelligence guidance for businesses, Ministry of Business, Innovation and Employment, Accessed 19 October 2026
- Generative Artificial Intelligence, Office of the Privacy Commissioner, 2023, accessed 19 October 2026
- The basics of workplace change, Employment New Zealand, Accessed 19 October 2026
- Artificial intelligence guide for small businesses, National Cyber Security Centre (NZ), 14 January 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.