NZ Police's AI transcription rollout: what businesses should copy

What NZ Police switched on

On Monday 2 November, New Zealand Police started using Auto-Transcribe, a speech-to-text product from Axon, the Arizona company best known for tasers. The tool turns recorded interviews into text for investigators and prosecutions staff, RNZ reported. Axon told RNZ the product averages 80 percent accuracy on public safety data and goes above 90 percent as the audio gets clearer. Police spokesperson Stuart Mills was candid about the limits: "These speech-to-text records are not always accurate."

Police call Auto-Transcribe a supporting tool. Investigators have to confirm and corroborate anything taken from a machine transcript before they act on it, and they have to record that check on file. The court transcripts will still be prepared, reviewed and verified by people in the police Transcription Services team. Police say they wrote a privacy impact assessment under their technology policy. When RNZ asked for a copy, police said it would need a formal Official Information Act request. RNZ also reports that Axon has already stored some New Zealand police interviews and footage in Australian datacentres, and that police aren't using Axon's separate AI report-writing product, Draft One.

The earlier trial is the more useful part

A second RNZ report published on 4 November fills in the history. Police tested OpenAI's Whisper in mid-2024 on intercepted Chinese-language phone calls. A six-month Whisper trial followed in early 2025, and officers broke the usage safeguards during it, so police largely suspended Whisper for about six months. Police have since added a second translation and transcription tool, from Meta. The Police Association, which has about 280 members in the restructured unit, says nobody consulted it on Whisper or the other AI transcription tools.

The same report says police cut about two dozen jobs in the file management unit in October. Police say AI played no part in those cuts. A July restructure document still states that new technologies will change the size, structure and skill mix of the workforce as the tools bed in. Transcription Services was mostly left alone in the restructure, but the team now has a modernisation programme of its own.

Any New Zealand organisation buying AI for routine work can learn from these two reports. Police wrote rules for the earlier trial, and staff didn't follow them closely enough to keep the trial running. The people whose jobs sit next to the tool say they weren't asked. The vendor's accuracy figure is the only performance number anyone has published.

Why a police project applies to your business

Few businesses transcribe criminal interviews. Many record speech and then act on it: contact centre calls, HR investigations, insurance claim interviews, site meetings and recorded Teams calls. Speech-to-text features now come bundled with meeting software and call platforms, so a lot of firms are already running AI transcription whether or not anyone formally signed off on it.

Axon's headline accuracy number shows what a buyer is up against. The RNZ article doesn't say how Axon measured the 80 percent, and Axon didn't give a method. A buyer can't tell from the figure whether the errors land on filler words or on names, dates, amounts and negations, and in a disciplinary file or a disputed claim a single missing "not" changes the record. Your own audio is the only reliable test, because accents, background noise, phone quality and te reo Māori place names all affect results.

The business case for transcription is real. Axon's 2020 launch announcement said manual transcription can take four to ten times the length of the recording, and the product was built so staff correct a machine draft instead of typing from scratch. The time saved on typing turns into checking work, and somebody has to own the checking. Auckland barrister Chris Wilkinson-Smith told RNZ that officers under time pressure won't check transcripts consistently, and he expects the work to shift to prosecutors and defence counsel. In a business the same shift happens quietly: the verification job lands on whoever reads the transcript last, often a manager or a customer who never agreed to do it.

The Privacy Act already covers your organisation here. The Office of the Privacy Commissioner's AI guidance, dated 21 September 2023, tells agencies to complete a privacy impact assessment before they start using an AI tool and to keep the assessment up to date. The guidance asks organisations to test whether a tool is accurate and fair for the purpose they intend (IPP8) and to keep personal information secure, including prompts and training data (IPP5). The Commissioner's advice for uncertain cases is blunt: if in doubt, don't use AI tools to handle personal information. Recorded interviews and calls are full of personal information, so a transcription rollout falls inside that guidance.

Where the data lives deserves the same attention. Both barristers quoted by RNZ raised the risk of depending on an overseas vendor, and Upper Hutt barrister Michael Bott pointed to Axon's US base as a data security concern. A business processing staff or customer recordings through an offshore service should know which country stores the audio, whether the vendor keeps it after transcription, and how hard switching providers would be. For sensitive recordings, organisations can also run open-source speech models such as Whisper on hardware they control, and a private AI server keeps the audio onshore.

Practical next steps

• Find what's already running. List every place speech is being transcribed today, including the meeting, phone and CRM features staff switched on themselves, and note where each one stores its audio and text.
• Test on your own recordings. Take a sample of real calls or meetings, have a person produce a reference transcript, and count the errors that change meaning: names, numbers, dates and negations. Don't rely on a vendor's average.
• Write the verification rule down and make it auditable. The police rule of confirming before acting and recording the check is a sensible model. Decide which uses need a human-checked transcript, such as disciplinary matters or disputed claims, and which can use the raw output.
• Do the privacy impact assessment before rollout. Follow the OPC's guidance, cover data location and retention, and keep the assessment ready to show staff and customers who ask.
• Talk to affected staff early. The Police Association's complaint about consultation is a warning sign for any employer. Tell people what the tool does, what stays with them and how errors get reported.
• Plan for safeguards being broken. The suspended Whisper trial shows that rules on paper aren't enough. Monitor how staff use the tool and set out in advance what triggers a pause.

If you want an outside view of how your AI tools line up with the Privacy Act and your own risk settings, DataForge's AI consultancy team works with NZ and Australian organisations on AI governance.