Air NZ and OpenAI create 'virtual customers' to test AI service
In plain words
Air New Zealand signed a direct partnership with OpenAI and is using simulated customers to check whether GenAI is ready to talk to real passengers.
Still a trial: Still being tested, not yet used with customers.
Source: nzherald.co.nz ↗
How to do it yourself
Our suggested steps. The company didn't publish theirs.
- 1
Write 10 customer personas (rushed, angry, confused, non-native speaker).
- 2
Use the prompt below to have ChatGPT act as each one.
- 3
Paste its messages into your chatbot and paste replies back.
- 4
Score each conversation: solved, tone, accuracy.
- 5
Fix the bot before real customers see it.
The instructions to give the AI
Paste this into ChatGPT (the free plan is fine), swap the [brackets] for your details, and test it on a few real examples.
Agent instructions (prompt)
Pretend to be this customer: [persona]. Your problem: [issue]. Talk to a support bot one message at a time, realistically (typos, impatience). After 6 turns, score the bot 1-5 on solving it, tone and accuracy, and say what it got wrong.
[Highlighted] = replace with your own details before you run it.
What you should get back: A full test chat as a grumpy customer, then a list of where your bot got it wrong.
Written by Quickly and Good, inspired by this case.
More detail: tools, prices and what you'll need
The company used: OpenAI
You could use: ChatGPT to role-play test customers against your own chatbot
- ChatGPT
General AI assistant: writes, summarises and answers from files you give it.
Free plan; paid plans available · Official site ↗
Prices from each tool's own pricing page, checked Oct 2026.
What you'll need
- ☐A ChatGPT or Claude account
- ☐Your own documents, templates or examples
- ☐The agent instructions (prompt)
- ☐A habit of checking every output before you use it
Can I do this myself?
Yes. You can try this today with a chat assistant and your own files.
Similar cases
ShipStation's Slack IT agent answers 97.5% of staff questions
Replied to 97.5% of messages, handled 812 of 1,221 conversations without a ticket (about 1,600 hours a year saved), and positive ratings rose from 82% to 96%.
Sophos cut security investigation time from 38 minutes to 89 seconds
Average response time on agent-handled cases fell from about 38 minutes to 89 seconds, and 52% of cases are resolved end to end by AI.
AGL's AI agent drafts security review documents for architects
Architects start from an AI draft instead of a blank page; deployed in the security architecture team.