The next stage of HR AI adoption in Asia starts with two things

A few months ago, I sat across from a Chief People Officer who'd just watched her team pilot an AI tool for the first time. I expected her to talk about accuracy, whether the outputs were right. Instead, she asked me something different: "Where does the question actually go once someone types it in?"

That stuck with me. It's rarely the concern I hear when people talk about AI. I think much of the confusion around "AI trust" - a phrase that's become shorthand for every hesitation a leader has about adopting technology - comes from treating it as one problem when it's actually two.

The first is whether the output is accurate. The second is whether the input is safe to share with an AI tool in the first place. Both concerns are completely legitimate. And most companies, in my experience, are only building for one of them.

None of this is an argument against AI. Done well, AI tools can considerably improve impact across every department. The upside is real, and getting specific about the risks and their fixes is what will allow HR teams to confidently reap the benefits AI agents can offer. 

The accuracy problem is a data problem

Start with accuracy, because it's the one everyone talks about, and for good reason. AI does make mistakes. Anyone who tells you otherwise hasn't used it enough.

When AI gets something wrong, the instinct across HR and people teams I talk to is almost always to blame the model. Wrong prompt, wrong tool, wrong vendor. Rarely does anyone ask what the AI was actually working from to begin with.

Earlier this year, Omni HR's State of AI in HR report surveyed more than 400 HR leaders in Southeast Asia, and the numbers told a story that's less about AI and more about how companies store their data. 91% of surveyed leaders reported having a single source of truth for their employee data. But when we dug deeper with follow-up questions about what their actual HR setup looks like, only 26% were on a genuinely unified platform. The rest were running a patchwork or two, three, sometimes more disconnected tools, synced manually or not at all.

AI will work with whatever data you feed it, unaware of inconsistencies. If you feed it faulty data, you will get faulty answers. Half of the leaders in that same study told us fragmented data was already limiting how far they could take AI. 70% said fixing data accuracy was the single most important thing they needed to do before AI adoption could actually pay off.

That's less of an AI problem than it is a data problem. Every function in a business is going to run into some version of this in the next few years; the moment you use AI to help make a decision, whatever inconsistency was hiding in your data quickly becomes visible.

The exposure problem is an architecture problem

The second concern, the one this CPO raised, and the one I think gets far less airtime than it deserves, has to do with data privacy.

Every time someone asks an AI tool a question involving sensitive information, that information has to go somewhere to get answered. For a lot of AI products on the market today, "somewhere" means a third-party model, hosted outside the company's own environment, reading through whatever records it needs to construct a response. 

For HR data specifically, and any other sensitive company intel, that should give any leader pause. Salaries, performance reviews, medical leave, disciplinary records. This is some of the most sensitive information a company holds, and in many cases, people don't actually know whether their AI tools are reading it directly, retaining it, or passing it somewhere else entirely.

The fix here comes down to the architecture of the AI tool itself. An AI tool should only ever see what the person asking is already authorized to see, and the underlying data shouldn't need to leave a governed environment just to produce an answer. It's the same principle we built into Mino, Omni HR's AI agent: it works from what's already inside the platform, scoped to each user's existing access, rather than reaching for a wider view of the business than the person operating it has. Mino is the first AI agent built for multi-country teams in Asia, running on live, centralized HR and payroll data across every market its users operate in. It can check what a compliance rule is in a specific market, compare headcount or cost country by country, or catch a retention risk before it becomes a resignation, all without the data ever leaving the platform its teams already trust.

Two problems, two fixes

The fixes are as different as the problems themselves.

For accuracy, start by mapping where employee data actually lives: how many systems hold a version of it, and how often those versions stay in sync. Organisations that can answer this clearly are already ahead, regardless of which AI tools they eventually choose. From there, prioritize the functions where the data is cleanest and most verifiable, and expand into more complex ones as the rest of the data catches up.

For exposure, the work looks different. Before adopting any AI tool that touches employee data, ask a few pointed questions. Does it only see what the requesting user is already permitted to see, or does it default to broader access? Is any data retained at the point of connection? Is every query and action logged and tied to a specific person, so it can be reviewed later? A vendor able to answer these plainly has usually treated exposure as a deliberate design choice.

A tightly permissioned AI tool running on contradictory, stale data will still hand back confidently wrong answers, just safely. A model fed clean, well-organized data with no boundaries on who can see what will hand back accurate answers at a real security cost. Solve for clean, connected data and you fix what AI is capable of saying. Solve for governed access and visibility and you fix where that conversation is allowed to go. Solve for only one, and you've solved exactly half of the "trust" problem.

At Omni HR, we believe an AI tool should never know more about your business than the person asking it a question, and it should never be right by accident. That's the standard we hold our own AI agent, Mino, to: built on a single connected data layer, with access controls that travel with the data instead of sitting on top of it. It's also the standard we think the rest of the industry should be building toward. Trust gets built by systems specific enough, and honest enough about their own limits, to earn it.