Every major HCM platform has spent the last 18 months building the same thing.
Workday launched Illuminate. UKG rolled out Bryte. SAP embedded Joule. Oracle refreshed their Fusion AI narrative. The pitch is consistent across all of them: we’ve added AI to people analytics. Now you can predict attrition.
Impressive product marketing. Expensive R&D. The wrong answer to the right question.
Here’s why — and why it matters to anyone about to extend a licence, sign a new contract, or bet their retention strategy on it.
The data problem nobody’s talking about
Every one of these platforms has added AI on top of the same foundational inputs they’ve always used. Engagement survey scores. Self-reported satisfaction ratings. Performance review data. Voluntary feedback tools. Manager assessments.
These are self-report signals. Employees decide what to tell you, when to tell you, and how to tell you. The AI processes whatever comes in.
Garbage in. Garbage out. Just faster now.
The structural problem with self-report data isn’t measurement frequency — pulse surveys attempted to fix that. The problem is that the employees most at risk of leaving are the least likely to tell you the truth. They’re disengaged from the process of being asked. They filter. They say what’s safe. Or they simply don’t respond. Response rates for enterprise engagement surveys have already fallen below 50% in many large organisations.
When people are gaming the input, the model can’t produce a reliable output — regardless of how sophisticated the AI sitting on top of it has become. The platforms know this. They just don’t have a structural solution for it. So they ship better dashboards instead.
What the market is actually celebrating
Workday’s Illuminate will surface beautiful risk visualisations. UKG Bryte will generate AI recommendations in the flow of work. Oracle will show you attrition heat maps by department. SAP Joule will suggest retention interventions via its copilot interface.
None of this is useless. For some workforce problems — headcount planning, skills gap analysis, internal mobility, pay equity — HCM analytics is genuinely improving. If you have high-quality input data, better AI processing gives you better output.
But for predicting which specific frontline employees are 60 days from handing in their notice, the data layer doesn’t support it. An HCM system is built to record what HR wants to know. It isn’t built to observe what employees are actually doing.
The inconvenient truth is that HCM analytics can only see what the HCM system was built to capture. It’s a closed loop. The insights it produces are drawn entirely from the data that flows through HR processes — and that means it can only tell you about the employees who were actively engaging with those processes in the first place. It captures what happened to the people who showed up to be measured. It tells you nothing about what’s unfolding beneath the waterline.
Where the real signal lives
Disengagement doesn’t announce itself. It leaks.
An employee who has mentally checked out starts shifting their behaviour weeks before they start looking for a new role. Shift acceptance patterns change — subtly, incrementally. Unplanned absence frequency ticks up. Voluntary overtime take-up disappears. Productivity drifts. After-call work stretches. Escalation rates inch higher. Tone changes in ways that are invisible to a survey but visible in operational data — if you’re set up to read it that way.
None of these signals appear in an engagement tool. None of them require an employee to report how they feel. They’re the natural output of someone still showing up but who has already left emotionally.
This is where predictive signal lives. Not in what people choose to tell you. In what their behaviour reveals whether they mean to or not.
That data already exists in almost every large organisation. It sits in scheduling systems, WFM platforms, CCaaS infrastructure, attendance records. It has been there the whole time. The HCM analytics layer just isn’t connected to it — and wasn’t designed to read it the way it needs to be read.
A $47 billion category with a gap at its centre
The global HCM software market is now valued at close to $47 billion. It’s growing at nearly 9% annually. Every major vendor is doubling down on AI investment. Workday acquired an AI knowledge management platform for $1.1 billion. ADP launched an AI-augmented analytics suite. The capital flowing into this space is real, and the intent is genuine.
But the single most commercially costly question any of these platforms should be able to answer — which of your people are about to leave, and when — remains largely unsolved at scale.
Because solving it requires a different kind of data than the systems were originally built to collect. The signals that reliably predict flight risk live in the operational layer, not the HR layer. And closing that gap isn’t a feature update. It’s an architectural rethink.
Where this goes next
The platforms that win the next cycle of people analytics won’t win by building better AI on top of better surveys. They’ll win by changing the input entirely.
The organisations that crack frontline retention in the next three years won’t do it by asking employees better questions. They’ll do it by learning to read the signals that are already there — in scheduling data, attendance patterns, productivity shifts, and call behaviour. The data that reflects what people actually do, not what they decide to report.
The HCM market is moving fast. The AI investment is real. But the structural flaw in the data model isn’t being solved by any of the major platforms — it’s being papered over with better interfaces. And the CHROs who’ve been burned by engagement tools that promised prediction and delivered retrospective dashboards are running out of patience for another version of the same thing.
The next breakthrough in people analytics won’t look like an upgrade.
It’ll look like a different category entirely.
Jonathan Hawkins is the Founder and CEO of Anthrolytics, a predictive behavioural analytics platform that identifies employee flight risk using operational data — no surveys, no self-report, no new data collection required.