Launch ChatGPT with a controlled first week of team work

Before the week starts: choose a bounded pilot

Select a small group that includes the task owner, someone who does the work and someone who can review it. Invite a sceptical colleague as well as an enthusiastic user. Pick tasks with available source material and reversible outputs. Public service descriptions, internal agendas and summaries of approved product information make useful candidates. Keep decisions about individuals outside this initial pilot.

Write a pilot card for each task. Include the problem, current process, permitted input, expected output, reviewer and stop condition. A generic Tauranga wholesaler could pilot a supplier-query draft based on public catalogue information. The purchasing manager checks the draft and sends it through the usual email system. ChatGPT has no authority to place an order. This scenario illustrates a workflow; it is not a client success claim.

Review the data before invitations go out. The Privacy Commissioner's AI guidance recommends assessment before use and revisiting it as the use changes. The Privacy Act 2020 still governs personal information used with AI.¹ Prepare approved practice files and remove hidden sheets, comments and identifiers. If the privacy review is unresolved, keep the pilot on public or wholly fictional material.

MBIE identifies business oversight and clear accountability as foundations for responsible AI.² Assign a sponsor who can pause the pilot, an administrator who can control access and a champion who can answer ordinary questions. These can be existing roles. Ensure somebody remains available when the owner is visiting customers.

• Agree a task-specific acceptance test before participants see the first generated answer.
• Create a help channel and a separate route for reporting sensitive incidents.
• Put the temporary data rules in writing and ask managers to explain them.
• Prepare an ordinary completed example so reviewers have something concrete to compare.

Monday: get everyone into the right workspace

Use ChatGPT Business for this walkthrough, or adapt it to your existing Enterprise controls. On the web, users switch workspaces from the profile menu. In Business, invitations are under Workspace settings, Members, then Invite member. Review the role and seat type before sending. Current Business documentation distinguishes Standard and Premium ChatGPT seats; a role controls permissions rather than usage limits.³

Ask each person to confirm the business workspace name before opening the practice task. Resolve invitation and access problems before the training session. Do not share a manager's login to get someone through the exercise. Record who accepted access and which approved workflow they will test. Let late joiners use the same onboarding checklist rather than relying on a colleague's remembered explanation.

Keep personal content separate during the initial launch. OpenAI allows a personal workspace to remain separate from Business. A merge transfers supported content into the organisation's workspace, removes the personal workspace from the switcher and has no undo option.⁴ Do not treat merging as a routine box that staff must tick to join a pilot.

Check your settings against the written pilot rules. OpenAI excludes Business and Enterprise workspace content from model training by default, while retention, memory and access are separate controls.⁵ Record what you verified for this workspace and what the plan does not provide. Avoid presenting controls from an Enterprise help article as controls every Business administrator will see.

• Monday output: an access list, a settings record and successful sign-in for each participant.
• Demonstrate where staff can find the workspace name and the pilot rules.
• Check the intended administrator can remove access without help from a departing employee.
• Pause onboarding for anyone whose account or workspace cannot be confirmed.

Tuesday: practise a task from input to approval

Teach one complete cycle: prepare the source, state the task, generate a draft, inspect it, revise it and obtain approval. OpenAI's prompt guidance recommends clear context and iterative refinement.⁶ Let participants see a weak result and identify the missing instructions. The lesson should finish with a checked output, not with applause for a fluent answer.

For an illustrative Christchurch building-services office, use a fictional enquiry that asks about servicing and availability. Give ChatGPT a public service description and ask for a draft that identifies unanswered questions. Instruct it to leave appointment times and prices for the manager. A useful output separates the draft message from a list of details staff must confirm.

Have the reviewer compare the message with the source line by line. Check that it has not promised an inspection, a warranty or an urgent visit. Then ask the participant to correct the prompt and rerun it. Save the accepted instructions and the review notes in your existing team document system. Keep the original source beside them so another person can reproduce the exercise.

Use a deliberate gap as a test. Leave a cancellation condition or delivery detail out of the training source and see whether the answer flags it. The GCDO's Public Service guidance explains that generated text can be convincing and wrong, and recommends checks against trusted evidence.⁷ That verification habit is useful for private employers too; the guidance itself is directed at government agencies.

• Tuesday output: a reviewed draft and an improved prompt for each pilot task.
• Make the task owner state which facts are supplied and which remain unknown.
• Ask a colleague to find an unsupported statement before giving any score for writing quality.
• Record the correction needed when the output fails the acceptance test.

Wednesday: build a Project with controlled sources

Create the pilot Project using New project in the sidebar. Add its approved reference material and enter project instructions through the more options menu, then Project settings. Those instructions override global custom instructions within the Project.⁸ Name the Project by its purpose so staff can recognise the correct working area without opening every file.

Choose a source owner who can replace outdated files. A generic Wellington consultancy could build a proposal-drafting Project with an approved capability statement, a blank response structure and public procurement requirements. Its instructions should tell ChatGPT to separate supported capabilities from questions for the director. Do not load confidential bids just to make the demonstration look more impressive.

Test access with an ordinary participant. OpenAI distinguishes chat and edit permissions, and shared Projects use project-only memory.⁸ Give edit access only to people who maintain the source material. Explain that participants can see shared project content. Keep client matters with different access needs in separately assessed workflows rather than adding them to the pilot hub.

Ask someone who did not create the Project to complete the exercise. They should identify the current source, obtain a draft and locate the review checklist without calling the champion. If the result depends on a private explanation or an undocumented file, improve the setup before calling it repeatable.

• Wednesday output: a working Project with approved material, tested access and a named maintainer.
• List the source version and its owner in the pilot record.
• Add a prompt that tests how ChatGPT responds when the reference does not contain the answer.
• Remove superseded reference material before another participant uses it.

Thursday: test one optional app connection

Skip this step if file uploads already meet the pilot need. An app connection should solve a known problem, such as finding the approved public brochure in a controlled folder. The NCSC's small-business AI guidance identifies data leaks, unreliable or manipulated outputs and supply-chain risks.⁹ Add the app provider and access path to your review rather than assuming that approval of ChatGPT covers every connected service.

OpenAI currently directs users to Settings, Plugins, then the relevant app or plugin. Complete any connection flow and inspect the requested permissions. App access depends on the plan, region, workspace, role and interface. An administrator can review the selected workspace's Plugins in Admin Console.¹⁰ Use the menus visible in your account and record a missing capability as a limitation.

Connect a test account with deliberately narrow source access. Ask it to find a known document and report its title and version before drafting anything. Check that the result is the intended source. Next, try a document the account cannot normally open. Stop the exercise if it exposes material outside the approved boundary and ask the administrator to investigate.

Do not enable automated sending or record changes during this introductory test. Write down who disconnects the app if the pilot ends. Give users an explicit instruction to stop when an approval prompt does not match the task they requested. An employee should not feel they must grant broader access just to complete the training exercise.

• Thursday output: a reviewed connection record, or a documented choice to continue with uploads.
• Record the provider account, allowed sources, available actions and responsible administrator.
• Verify the source version in the connected system before using its contents.
• Keep writing actions disabled or outside the pilot's authorised scope.

Friday: decide using completed work and staff feedback

Review the evidence task by task. Count accepted drafts, unsupported claims, corrections and unresolved access issues. Measure the whole process, including preparation and checking. Do not announce a productivity gain from the generation step alone. Ask whether the workflow remains useful when the champion is absent and when the source has an awkward gap.

Keep feedback about the tool separate from judgements about individual performance. Employment New Zealand explains good faith obligations under the Employment Relations Act 2000, including responsive and honest communication.¹¹ If expansion would change duties or affect jobs, use the relevant workplace-change process and obtain advice before deciding. A training pilot should not become an undisclosed employee ranking exercise.

For safety-related work, require the safety lead's approval before adding it to the next phase. WorkSafe describes the business's primary duty under the Health and Safety at Work Act 2015 to protect workers and others so far as is reasonably practicable.¹² An accurate summary of an approved procedure does not establish that a suggested new procedure is safe.

Give each task one decision: continue under the same controls, repeat with changes, or stop. Specify the reason, owner and next review point. TheColab offers a ChatGPT essentials workshop for teams that need guided practice and a Business and Enterprise rollout course for managers establishing the workspace and governance process. Training should support the approved task and its review standard.

• Friday output: a written decision for every pilot workflow and an action owner for each gap.
• Continue only when the source, permissions and reviewer are ready for ordinary use.
• Repeat the trial when a useful task still produces errors that staff can clearly diagnose.
• Stop a task when the necessary checking exceeds the team's capability or the input approval remains unresolved.

Prepare the next group without expanding the risk

Package the accepted prompt, current source, approval steps and a failed example for the next group. A failure teaches staff what to look for when a plausible answer departs from the evidence. Keep the help route active and brief the next manager on the decisions that remain with them. Gradual expansion works best when the process can be explained without the original trainer.

Review the workspace membership and Project permissions again when people move between roles. Retire practice files that could be mistaken for live documents. Schedule a follow-up review of accepted outputs and update the task register when the provider changes a feature or menu. Keep each expansion within an explicit scope instead of silently turning a drafting pilot into an automated operating process.

• Give the next cohort a fresh practice source and the same acceptance standard.
• Assign ongoing source maintenance before the pilot champion returns to normal duties.
• Record which tasks still need specialist, privacy or technical review.