Author: David Jenkins, NZPPA CEO
Consider a familiar payroll situation. It is nearly cut-off, an employee needs their bank details changed, and the message sounds entirely reasonable. Then a manager calls to say it must be done today. This is an illustrative scenario, not a reported incident. It shows why the important question is no longer simply whether a request looks or sounds genuine. It is whether the change has been independently verified.
Artificial intelligence can make deception more convincing, but payroll practitioners do not need to become AI experts to respond. They need reliable controls, enough time to apply them and management support when they question a request. Businesses also need to examine the risks created when their own staff use AI with payroll information.
Recent news about hackers and criminals using artificial intelligence has made me think about what payroll practitioners and businesses should be doing to protect payroll. The threat is real, although it is important to separate the facts from the hype. I should also acknowledge that I used AI to help research this article and find references. AI can be a useful tool for payroll practitioners when it is used responsibly, confidential information is protected, and the results are checked.
What the evidence actually tells us
New Zealand’s National Cyber Security Centre reports that criminals are already using AI to increase the speed and scale of attacks, research their targets and produce personalised phishing messages and false audio or video. That is a credible reason to strengthen payroll protection. It is not evidence that every payroll scam is AI-generated, or that an AI system can effortlessly defeat every security control.
New Zealand has already experienced the underlying payroll fraud. In 2019, the Office of the Auditor-General described incidents in which scammers impersonated employees and asked employers to change their bank details. The changes were made without independent verification, and the fraud often emerged only when employees asked why they had not been paid. AI involvement was not established in these incidents. They show that the weakness AI can exploit is already familiar: a believable request being accepted without a separate check.
There is also direct evidence of payroll diversion. In October 2025, Microsoft described attacks against US organisations in which criminals compromised employee accounts, changed salary payment details and hid warning emails. Microsoft did not identify AI as the cause. The distinction matters: payroll already has weaknesses worth fixing, without attaching an AI label to every incident.
For payroll, the potential harm is personal. An employee whose wages reach the wrong account may miss rent or a mortgage payment while the organisation investigates. A stolen payroll file can expose income, bank details and other private information. A system outage can prevent an otherwise accurate pay run from being completed. These consequences deserve attention regardless of the technology used by the attacker.
The misconceptions that weaken our response
“We do not use AI” offers little reassurance when the person targeting the business does. Good spelling is no longer a useful test of trustworthiness, and a recognisable voice or convincing video should not be treated as proof of identity. A request that appears to come from a senior manager still needs the same checks as any other request.
Cloud software and multifactor authentication, which asks for an additional proof of identity at sign-in, are valuable protections. Neither removes the need for verification. Microsoft’s investigation showed how some authentication codes could be captured through deceptive sign-in pages. Its recommendation was to use phishing-resistant authentication, such as suitably configured security keys or passkeys.
There is a danger in the opposite direction too: suggesting that AI makes fraud unstoppable. That can leave staff anxious and businesses shopping for a technological cure while basic controls remain weak. A more useful question is whether one convincing message, one compromised account or one rushed approval could redirect someone’s wages.
Where payroll practitioners can make a difference
Start with changes to bank details. Verify them through an established, approved process using contact information already held independently of the request. Do not use a phone number supplied in the suspicious message. Calling a known number is a useful safeguard, but the sound of a familiar voice is not enough on its own. The process should establish who is requesting the change and record how it was checked. New Zealand guidance similarly recommends independently checking emailed payment requests.
Then look at what happens before payment. Review changes to bank accounts, new starters, pay rates and unusual one-off payments, rather than relying only on whether the total payroll looks right. A salary can be diverted without changing the total by a cent. Someone independent should check significant changes and the final payment instructions. Where a small business cannot fully separate duties, the owner or another authorised person needs a meaningful compensating check, not a routine click to approve.
The most awkward test may be an urgent request from the chief executive. Payroll needs explicit backing to pause it. An approval process that senior managers can bypass whenever they are in a hurry is an unreliable control. The same applies to employee self-service: convenient access should not mean important changes escape review.
These are practical recommendations, not a guarantee against every attack. Their value is that they reduce dependence on one person spotting a convincing fake. Staff should also know how to report a doubtful request or an accidental click quickly, without fear that admitting a mistake will make matters worse.
The risks created by our own use of AI
A practitioner asking AI to improve an explanation of a payroll process faces a different risk from someone uploading a complete payroll export. The latter may disclose information the business has no good reason to share with that service. The Privacy Commissioner warns that information entered into AI tools may be retained or disclosed, and that confident answers can contain factual or logical errors.
Use only tools approved for the information and task involved. Check what the provider retains, who can access the data and whether it can be used for training. Remove unnecessary personal information and use fictional examples where possible. Removing employee names alone does not necessarily make a detailed payroll record anonymous.
Accuracy needs its own controls. A polished answer about leave, deductions or tax is not evidence that it is correct for New Zealand, current, or applicable to the employee concerned. Check the underlying official source and the relevant facts. Test calculations through established payroll procedures. An AI answer should support professional judgement, not replace the evidence on which a payment depends.
Giving an AI assistant access to email and payroll systems adds another concern. An attacker can hide instructions inside material the assistant reads and try to make it disclose information or take an unauthorised action. The UK NCSC warns that this problem cannot be treated as a fully solved security issue. Keep access limited and require separate human approval before changes to employee bank details or payment release.
What the business must own
Payroll cannot carry this responsibility alone. Management should agree who owns access security, employee verification, payment approval and incident response. IT or the service provider should enforce strong sign-in protection, remove unnecessary access, keep systems updated and monitor important changes. These measures align with the NCSC’s advice to strengthen basic security and prepare for incidents.
Supplier assurance needs practical questions. Can the business see who changed bank details and when? How quickly will the supplier report a security incident? Can records be restored, and has recovery been tested? How will employees be paid if the usual system is unavailable? Outsourcing software does not remove the employer’s need to understand those answers.
The organisation also needs a proportionate AI policy that people can follow. It should identify approved tools, prohibited data, required checks and who is accountable for the result. The Privacy Commissioner expects leadership consideration and a privacy impact assessment when organisations introduce generative AI involving personal information. A ban without a workable alternative may simply leave managers unaware of what staff are already using.
Make the next pay run safer
A useful starting point is to walk one bank-account change through the entire process. Ask who verifies the request, who can enter it, who reviews it and what evidence remains. Then test an urgent executive request and a payroll outage. Any step that depends on “we would probably notice” deserves attention.
If something goes wrong, stop the affected payment where possible, contact the bank promptly if funds have moved, involve the security team or provider and preserve relevant messages and records. Include privacy assessment and employee communication in the response plan rather than improvising them under pressure.
In conclusion, AI will continue to change the quality and volume of deception. Payroll’s response should be specific and practical. Give practitioners the authority to question requests, protect the systems they rely on and make independent verification part of ordinary work. The employee waiting for payday needs those controls to work, not a confident promise that the latest technology has everything covered.
