| SCOPE NOTE This post is general context on UK data protection, automated decision-making and profiling in an employment setting. It is not legal advice. For questions specific to your organisation, take advice from a qualified data protection or employment law specialist. UK GDPR Articles 22A to 22D came into force on 5 February 2026 through section 80 of the Data (Use and Access) Act 2025. |
The first sign is almost never the resignation letter.
Most people who leave a job have been leaving for months. There is a quieter meeting, a passed-over promotion, a manager change, a silence in the one-to-one that was not there last year. By the time HR sees the notice, the decision is old. That is the gap AI can help close. No magic score. Just a more honest read of the signals your organisation was already producing.
I have been in enough post-mortems on lost hires to say this plainly: the surprise is almost always avoidable. And in a UK SME, where the loss of one senior person can wipe out a quarter’s plans, the case for looking earlier is not sentimental. It is commercial.
Turnover prediction is the use of workforce data to estimate, ahead of time, who is most likely to leave and when.
It uses pattern recognition on a set of signals that historically travel with people who go on to resign: tenure, time in role, time since last promotion, engagement scores, absence changes, manager changes, pay position against the market, learning activity, and collaboration patterns. On a small team, an experienced people leader can hold most of that in their head. Across 80, 120 or 200 employees, they cannot.
AI earns its place here because the patterns rarely live in one place. A dip in engagement matters more when it lands two months after a promotion cycle where someone was passed over. A change in manager matters more for a mid-career specialist than for a first-year hire. Turnover models weight those signals together and show where to look.
The point is direction, not prophecy.
Because the cost of one avoidable exit is bigger than most SMEs budget for, and 2026 UK data shows quit intent is not falling.
The CIPD Good Work Index, based on a survey of over 5,000 UK workers, has consistently found around 20% of employees say they are likely to quit in the next 12 months. That number has been broadly stable since 2019 outside the pandemic dip. If your headcount is 120 and your team looks like the national average, roughly 24 people are already thinking about it.
The cost side is harder to see. Recruitment agency fees are only the visible line. Add manager time, interview panels, notice-period handover, ramp-up on the replacement, missed revenue on delayed work, and the compounding effect on the team the leaver sat in, and a single mid-level departure regularly runs into the tens of thousands of pounds. For a business without a dedicated People team, that is real money spent on a fire that was preventable.
A prediction system does not stop everyone from leaving. It stops you finding out last.
The strongest signals are almost always a change from that person’s own baseline, not a fixed threshold.
A model that flags anyone below a benchmark misses the point. What matters is the delta. Some of the signals worth watching:
None of these on their own means someone is leaving. Together, and with movement over time, they narrow the shortlist of who to talk to.
Take a scenario I have seen more than once. A senior developer in a 90-person business, five years in, strong performer, no drama. Her engagement score drops by 8 points on the January pulse. Not a red flag on its own. Two weeks later she takes three consecutive Wednesdays as annual leave. Not unusual. In February her manager is reorganised into another team and she gets a new one. Individually, none of these signals warrants action. Together, on the same person in the same eight weeks, they warrant a coffee. The point is not that she is definitely leaving. The point is that the shortest way to find out what is going on is to ask her, before someone else does.
The rules changed on 5 February 2026. Automated decision-making about employees is now permitted with safeguards, not prohibited by default.
Section 80 of the Data (Use and Access) Act 2025 replaced the old Article 22 of the UK GDPR with new Articles 22A to 22D. Article 22C sets out four safeguards that must accompany a solely automated decision with legal or similarly significant effect: information about the decision, the opportunity to make representations, human intervention, and the ability to contest the outcome.
The ICO published a report in March 2026 on how employers use automated decision-making in recruitment. Its central finding was that many employers were relying on solely automated systems without meaningful human involvement, and did not realise it. The same trap applies to retention. A “flight risk” score that a manager acts on without understanding, or that flows into performance conversations, can quickly move from support to decision.
Two practical implications for a UK SME thinking about turnover prediction:
The statutory AI code of practice under UKSI 2026/425 came into force on 12 May 2026. The ICO’s final guidance on automated decision-making, including profiling, is expected in Winter 2026. The direction of travel is set. Human involvement has to be real, not decorative.
Most of the value comes from combining data you already hold and running a short, repeatable review.
You do not need a vendor platform to start. You need three things.
First, a shortlist of signals your business will watch. Pick six to eight from the list above. Assign a source for each: the HRIS for tenure and absence, the pay review file for market position, the pulse survey tool for engagement, one-to-one notes for manager changes and mood.
Second, a monthly retention review. One person, one hour, one page. Who has crossed a tenure milestone this month. Whose engagement has moved. Whose pay is now below market. Where has a manager changed. The output is a short list of names for conversation, not a score attached to individuals.
Third, a clear line between watching and deciding. The review flags who to talk to. It does not decide anything about them. The person and their manager keep the decision, informed by the conversation. That line is what keeps the exercise inside the safeguards Article 22C requires.
A workable one-pager looks like this. Top of the page: the review month, the reviewer, the total headcount. Below that, a table with five columns.
One: employee name.
Two: which signal was triggered.
Three: what has changed compared to their baseline.
Four: who is best placed to have the conversation.
Five: the date the conversation is booked for.
Nothing else. No score. No risk category & no dropdown of predicted outcomes. The document does one thing: it moves a name onto a manager’s calendar. That is the whole point of the review.
For an SME under 200 people, this is often enough to move from surprise exits to early conversations. When you outgrow the manual review, you have a defensible baseline before you buy anything.
The best intervention is a good stay conversation, not a counter-offer.
A well-run one-to-one where the manager asks what is working, what is not, and what would make the next 12 months worth staying for will surface most of what a model was hinting at. It also gives the person the sense that they were seen before they were leaving. The manager needs a light structure, permission to hear an uncomfortable answer, and the authority to act on at least one thing that comes out of it.
Counter-offers to resignation letters are the worst version of retention. They are expensive, they signal that reward moves only when threatened, and the research on their long-term success is not kind. Stay conversations are the version that scales.
Track prevention, not prediction accuracy.
The vendor version of this question is model accuracy. The business version is whether the number of avoidable regretted exits went down. Useful measures for an SME:
If the model gets sharper and nothing changes in the numbers above, the model is not the problem.
Predicting turnover is not the goal. Losing fewer people you did not need to lose is.
Turnover prediction is the use of workforce data to estimate, ahead of time, which employees are most likely to leave. In an SME it is usually a monthly review of signals such as tenure, engagement, pay position and manager changes, rather than a full machine learning model.
Yes, subject to UK data protection law. UK GDPR Articles 22A to 22D, in force since 5 February 2026, permit solely automated decisions with legal or similarly significant effect provided the safeguards in Article 22C are met. A data protection impact assessment is likely to be required. Take specialist advice for your circumstances.
No. An SME can start with a monthly review of six to eight signals drawn from the HRIS, pay data, pulse surveys and one-to-one notes. Dedicated platforms add scale, not the underlying insight.
There is not one. The strongest predictor is usually a change from the individual’s own baseline across several signals over time, not a single threshold.
Monthly is usually the right cadence for a business under 250 employees. Weekly is overkill. Quarterly is too slow to prevent an exit already in progress.
Turnover prediction narrows the shortlist of who to talk to. A stay interview is the conversation itself. The prediction is only useful if it is followed by a real conversation the manager is trusted to have.
AUTHOR
Sabiha is a Talent Acquisition Director, Speaker and Author with 16+ years of experience helping UK organisations build smarter, more inclusive hiring and retention systems. She is the author of How to Use AI to Win Talent and Retain People (Trotman, 2026), a CIPD-aligned guide for HR leaders navigating AI in the workplace. Connect with her at meetsabiha.com.
Global Talent. Ethical AI. Strategic Hiring. Sustainable Retention.
