DLPA’s new research shows AI has moved into performance reviews, feedback and terminations before anyone set rules for it.
There’s a question in our survey asking people why they use AI at work, and most of the answers are what you’d expect. Saving time, writing better emails, coming up with ideas when the well’s dry. Then about one in eight managers (11.9%) ticked “to avoid difficult conversations”, and I’ve been thinking about that box ever since.
Most managers got promoted for being good at the work, and nobody checked whether they could tell someone their work isn’t up to scratch without either flattening them or saying nothing useful at all. Now there’s a tool that will write a calm, reasonable, personality-free paragraph about “areas for development” in about four seconds, and if you dread the conversation, you can see why you’d reach for it. It works, too. For the manager.
The person on the other end still gets the bad news, they just get it from something that has never met them and doesn’t understand their nuance, and people can smell that. Just over eight out of 10 of the people we surveyed would rather be managed by a human, and almost half of those working in organisations trust AI-written workplace communication less than something a person wrote. A performance conversation is the one place where trusting the person talking matters more than what they actually say, and it seems we’ve handed a chunk of them to software.
How big a chunk is the bit that should worry HR. More than half of the managers in our data who use AI (54.2%) have already used it on a people decision, whether for hiring, promotion, performance or a restructure. About one in five has used it to help with a performance review, and 16 per cent (85 managers in our sample) have used it on a redundancy or a termination. Every one of those is a decision someone can take to the Fair Work Commission, and I wouldn’t want to be the HR manager sitting there trying to explain how it was made. “Did a person actually make this decision?” is a fair question for a member to ask, and “the AI drafted it, but I did read it” is not what you want on the transcript.
It would be easy to make managers the villains here, and I don’t think they are. They’ve been told from every direction that AI is the future, and they’ve taken that seriously. 37.4 per cent of everyone we surveyed believe that refusing to use AI, or refusing to use more of it, could hurt their career, and among people who use it often that rises to 58.3 per cent. Meanwhile, only 31.3 per cent of managers have had any formal AI training, three in four people working in organisations have had none, and 69.1 per cent can’t confirm their employer has an AI policy at all. So we have managers who feel they can’t afford not to use the tools, using them on the hardest part of their job, with very little to go on. Safe Work Australia’s model Code of Practice lists poor support and lack of role clarity among psychosocial hazards, and you don’t have to squint to see both in that picture. The hazard sits with the manager writing the performance review as much as with the person receiving it.
The obvious fix is training, and this is where the data got interesting. People who’d had formal AI training at work were more likely to have used an unapproved tool than people who hadn’t (55.2% against 42.5%), and far more confident they understood the privacy and security risks (81.2% against 36.1%). There are a couple of ways to read that. Trained people may simply recognise a breach when they see one, while untrained people breach without realising it. Or the training made them capable enough to go looking for better tools than the approved ones. It’s survey data, so I can’t tell you which. What I’m fairly sure of is that training, the way most organisations are doing it, is producing confident users. Whether it’s producing good judgement about when to keep AI out of a decision altogether is a different question, and it’s the one that matters when the decision is about a person.
So if your plan is an AI training module, I’d look hard at what it actually teaches. Most of what I’ve seen covers the tool, which buttons to press and what not to paste in. Very little of it covers when a manager should close the laptop and have the conversation themselves.
For what it’s worth, I wouldn’t write another policy. I’d find out where AI is already sitting inside your people processes (it is, trust me) and decide which of those decisions need a named human who owns them. Then do the unfashionable thing and teach your managers to have the conversation themselves. It’s slower and it costs money, but it’s still a lot cheaper than the claim.
About the research: DLPA surveyed 1,008 Australians through the Pureprofile research panel in August 2026. Just over half (53%) hold owner, executive or management roles. Each figure is reported against the respondents who were asked that question; some questions went only to AI users or excluded sole traders.
Karlie Cremin is CEO of Dynamic Leadership Programs Australia and author of Don’t Lead Like a Jerk and Your Computer Is Trying To Eat You.
