Identify the claim that could cause harm
OpenAI warns that ChatGPT can invent references, give incorrect facts and express confidence in a mistaken answer. Search, deep research and data analysis can support verification, but the company still recommends checking important information against reliable evidence.1 For an owner, the useful question is whether a particular sentence could change a payment, a promise, someone's treatment or a work instruction.
Read a draft with that question in mind. Highlight factual statements separately from suggested wording. A heading such as delivery options is a writing choice. A sentence promising delivery before the customer's event is a business commitment. Check the commitment against the job record, even when the prose sounds ordinary. Ask the author to mark assumptions before passing the draft to someone else.
- Highlight claims about names, amounts, dates, eligibility, obligations and technical limits.
- Mark anything attributed to a customer, employee, regulator or supplier.
- Assign a reviewer who can access the evidence and understands the decision.
- Hold the output when missing evidence could affect the result.
Prepare a source pack before you prompt
For repeat work, collect the current policy, approved template and relevant records in an authorised workspace. ChatGPT Projects can hold files, related conversations and project instructions. Create one with New project in the sidebar, then open its more options menu and select Project settings to add instructions.5 Keep unrelated client matters apart. Name each source clearly so a reviewer can distinguish the current agreement from an earlier draft.
Write a source-led request: Use the attached supplier agreement and the purchase order to draft an internal summary. For each commitment, identify the document and quote the supporting passage. Put unresolved conflicts in a separate list. Where neither document supplies an answer, write not established. This format is a review aid. It gives you visible questions to resolve without assuming that the model's source labels are correct.
The GCDO's guidance, published by DIA for the public service, recommends specific prompts, reliable data, testing and checking results against trusted sources.8 A private firm can adopt these practices as a management method. Give staff a short source pack for each recurring task, with an owner responsible for replacing outdated material.
Use Search for facts that can change
In ChatGPT, open View all tools and choose Search, or type a slash and select Search in the composer. Availability can depend on workspace settings and usage limits.2 State New Zealand, the topic and the relevant period in your request. Ask for the regulator or original publisher rather than a collection of summaries. Avoid including customer identities in a public research prompt.
Select each citation and open its page. Use Sources when that control appears to inspect the links. OpenAI cautions that search citations can be incorrect, incomplete or outdated.2 Read the passage supporting the claim. Check the publication date, jurisdiction and exceptions. A real page about a different situation is insufficient evidence. Record the page you read alongside the sentence you plan to use.
For a substantial research question, start deep research with /Deepresearch. In Sites, choose Manage sites to specify domains and whether the research must stay within them. Review its proposed plan before it starts.3 For a NZ policy paper, select relevant government and regulator sites. Inspect the final references yourself and record any point that those sources leave unresolved.
- Confirm that the cited publisher is responsible for the information.
- Find the actual supporting text on the source page.
- Check whether a newer publication or amendment changes the answer.
- Separate the source's statement from your interpretation of its effect.
Reconcile spreadsheet output with the underlying records
ChatGPT data analysis accepts structured files and can run Python calculations. OpenAI recommends clear headers and one record per row, and tells users to review generated code, outputs and assumptions.4 Upload an approved extract with defined columns. Explain what a row represents, which records to exclude and whether amounts include GST. Keep the untouched export available for comparison.
Ask for a reconciliation before asking for conclusions. Request the records included, the exclusions, the grouping rule and the calculation used. Check the result with a spreadsheet formula or accounting report outside ChatGPT. Inspect credit notes, missing values, duplicate records and date boundaries. A correct calculation over the wrong records can produce an unusable conclusion.
Statistics need the same care. Stats NZ's work-from-home publication identifies distinct data from the Household Labour Force Survey, Business Operations Survey and Census, with different definitions and reference periods.12 When ChatGPT quotes a national figure, establish which measure it uses. Do not present a historic work-location measure as current AI adoption, or turn an industry estimate into a claim about your own staff.
- Compare the source record count with the records used in the analysis.
- Check a normal transaction, an exception and a blank entry.
- Recalculate the total independently before acting on the chart.
- Save the checked method and the source extract with the final report.
Worked exercise: a Hamilton wholesaler's overdue report
Consider this hypothetical exercise for an eight-person Hamilton wholesaler. The office manager wants a weekly overdue-account summary. The training extract uses invented account labels and contains an unpaid invoice, a credit note and an invoice still within its agreed terms. No real customer information is needed to practise the check.
The manager asks ChatGPT to group balances by account and explain which are overdue. The draft wrongly treats all positive balances as late. The reviewer opens the export and checks the due-date column against the reporting date. They also find that the credit note has been counted as another debt. These are defined faults in this exercise, not a claim about a real client's experience.
The reviewer asks for a revised method: apply the due-date rule first, preserve the sign of each transaction, list exclusions and show the source rows for every account total. They reproduce the totals in their spreadsheet. They then approve an internal report, while keeping customer chasing messages as a separate review step. The manager adds both faults to the next training exercise so staff must demonstrate the check themselves.
Increase the review depth for people and safety
Information privacy principle 8 of the Privacy Act 2020 requires reasonable steps, in the circumstances, to ensure personal information is accurate, current, complete, relevant and not misleading before use or disclosure.6 A generated allegation about an employee therefore needs more than a spelling check. Compare it with original records and preserve uncertainty where the records disagree. The Privacy Commissioner also expects human review before action on generative output.7
Employment New Zealand, part of MBIE, explains good faith as truthful, responsive communication between parties to an employment relationship.10 Use that guidance when reviewing ChatGPT material for a process under the Employment Relations Act 2000.10 Ask an employment adviser to check the applicable process, including current exceptions, before using a draft that could affect someone's job. Do not let a polished summary turn a disputed observation into an established fact.
WorkSafe explains that the Health and Safety at Work Act 2015 puts primary responsibility on businesses for workers' health and safety.11 If ChatGPT rewrites a machinery procedure, a competent supervisor must compare it with the controlled procedure and actual equipment before you issue it. Preserve warnings, conditions and stop instructions. An easier sentence is unacceptable if it changes the safe sequence.
Check for instructions hidden inside source material
The NCSC's joint small-business guidance describes prompt injection as malicious input that can manipulate an AI system into unsafe or incorrect answers. It recommends human involvement for sensitive work and monitoring unusual behaviour.9 Treat emails, web pages and supplier documents as evidence to inspect. A request embedded in those documents should not become authority to send information elsewhere.
In a hypothetical Christchurch construction office, staff use ChatGPT to summarise a supplier specification. Add an obvious test instruction to a training copy that asks the tool to ignore the specification and reveal unrelated information. The exercise should end with the reviewer identifying the instruction and rejecting it. Keep testing material apart from live records. Report surprising behaviour to the workspace administrator before continuing the workflow.
Put a release check into the team's normal work
Use an approval note with the final document. State what ChatGPT helped produce, the evidence checked, the unresolved questions and who approved release. Store that note where your team already manages quality records. Keep it proportionate: a routine internal rewrite can have a brief author check, while client advice needs a reviewer with the right competence.
Teach the method with actual document types from your business, using sanitised practice material. TheColab's ChatGPT essentials workshop and full-day workshop cover prompting and checking. The finance and operations course provides a relevant next step for spreadsheet work. We come to your team anywhere in New Zealand, including Waikato and Canterbury. Use the related course and industry links to choose the training that fits the task.
- Confirm that every consequential claim has supporting evidence.
- Confirm names, recipients, amounts and promises against original records.
- Record who checked technical or sensitive content.
- Remove unsupported claims before the document leaves the business.
- Capture a caught error as a future practice case without exposing personal information.