Employee Turnover Causes: Which Leavers, and What HR Can Change

Your attrition rate came in flat, and it’s hiding three separate things.

Underneath it, a support team lost most of its senior people over eight months. A graduate intake left before their second anniversary. And the function you’re about to expand lost the two people who knew how the work really ran. Averaged in with everyone who left for reasons you were fine with, all of that reads as steady.

Employee turnover causes get written up as a list: pay, managers, growth, recognition, burnout. Every item is true somewhere in your company and none is true everywhere in it. The list only pays you back once you can say which leavers, from which segment, and whether anything you control could have gone differently.

Turnover is the tenure half of employee lifetime value, which is what our work has been about since 2017. The chain from a tenure lever to a number your CFO already reports is laid out in how HR contributes to business growth. This post goes at turnover itself.

A company-wide attrition rate answers a question nobody asked

One rate adds the exits you were relieved about to the ones that cost you a quarter of delivery, then divides by headcount. Whatever comes out, you can’t act on it.

It averages populations with nothing in common. Shift-based roles, a sales floor, a graduate intake and a twelve-person data team run at different speeds for reasons that don’t overlap, so a real problem in one gets canceled out by a good year in another. It flattens timing too. An annual figure looks calm across a run of exits that all landed inside two quarters.

Cut the same exits several ways and see which cut concentrates them.

Cut	What it tells you
Tenure band, from joining to beyond three years	A hiring and onboarding problem, or a ceiling problem
Function and level	Where exits concentrate, and whether that tracks the market for the skill
Hiring cohort, by source and start quarter	Which channels and intakes hold up
Manager and span of control	Whether exits cluster under a few managers or spread evenly
Pay-band position at exit	Whether the leavers sat low in band
Months since the last move or promotion	Whether exits follow a stall
Location and entity	Market, commute and policy effects, cheaper to fix at one site

Read a concentration against that segment’s own history, never against the company number. And in a small group the exit count says more than the rate does: four out of eleven is a fact somebody can hold in their head, while the same thing as a percentage starts an argument about the denominator.

Hold a minimum group size throughout, because slicing until a cohort is small enough to name people is how an analysis becomes a privacy incident. And settle the definitions before the first cut, since plenty of HR teams carry three live definitions of attrition across their dashboards without knowing it.

There’s a low-sensitivity agent for that second one. A metric definition and governance agent maintains the dictionary, flags dashboards running conflicting logic, and drafts changes with the before-and-after numbers. It reads report logic and no row-level personal data. The Head of People Analytics approves every definition, Finance co-approves the cost and FTE ones, and it changes nothing on its own.

Regretted or not, decided before the resignation lands

Most companies set the regretted flag at the exit, in a dropdown, filled in by the manager who just lost the person. That flag records how the last two weeks felt. A manager who got blindsided marks regretted. One who saw it coming marks the other box.

The fix is a definition and a calendar, in that order.

Write one definition, with the function heads in the room. Two tests: would we hire this person back into this role today, and did we plan to keep the role as currently scoped. Regretted needs a yes to both. An exit you’d have made yourself, or a role you were closing anyway, sits outside it whatever the circumstances.

Set the flag before anyone resigns. Attach it to a cycle you already run: talent review, calibration, a quarterly roster pass. Manager plus HRBP, dated, stored in the HRIS as a point-in-time record. Once it predates the resignation, nobody can reverse-engineer it.

Allow a third state. Forcing a binary manufactures noise. Let people mark unclear, then watch how big that pile gets. A large one says your talent review isn’t producing a view.

Sample and audit each quarter. Pull a handful of exits, set the flag on the exit record beside the one filed before notice, and take every mismatch back to the HRBP who owns it. Flags that moved after a resignation landed are the ones worth reading.

Two tells that your flag is broken. If almost every exit comes out regretted, the field is measuring politeness. If almost none does, someone is protecting a number.

If you want the definition applied the same way by every manager, a regretted attrition analysis agent applies the one you wrote, prompts the manager and HRBP for the inputs it asks for, and aggregates the result by segment. Those two confirm every tag. It never derives one from a performance rating, individual tags stay inside HR, and the reporting runs at segment grain.

Once the split is real, you can cost the half that matters. Skip the replacement-cost multiples that circulate online, because your CFO will ask where the figure came from and that conversation goes nowhere. Build it from your own line items: sourcing or agency spend, vacancy weeks covered by overtime, interview hours, the payroll spent while a replacement is still learning the job, and the work that stalls while nobody owns it. Use Finance’s assumptions, and it’s their number too.

The causes, sorted by whether you can move them

Every turnover article lists causes. Few say which ones sit inside HR’s budget, and that column is the point.

What drives exits in a segment	Can HR move it
Pay position against market for that segment	No. Comp and the CFO own it. HR sizes it and costs another year of it
Workload and staffing against the plan	No. The function head owns it. HR shows the pattern and the vacancy cost
The work itself, and whether the structure has a next level in it	No. Org design owns it, and saying so early beats running a program
A reorg, an acquisition or a leadership change	No. Already decided. Plan for the exits and staff the gap
Manager capability	Partly, with the function. The highest-value lever you own
Whether a career path exists in practice	Yes, on the mechanics: internal posting, promotion cadence, move approvals
Selection, and how the job was described at offer	Yes, with your hiring managers
Onboarding and time to productivity	Yes
Recognition, feedback and workload inside a team	Yes, through managers

Look at the four rows that say no. Where pay position or workload is driving exits in the segment you care about, a listening program won’t fix it, and running one costs you twice: the money, and the credibility of the next survey, since the answer people already gave comes back to them as a workshop.

Naming the constraint is a real HR output. A CHRO who tells the exec meeting that exits there track pay-band position, here’s the cost of another year of it, and here’s who has to fund the fix, has done the job. The rows below those four are where your budget belongs, and manager capability leads that group. It shows up as line managers handling the conversations they’ve been postponing, one team at a time.

The trust trap: scoring individuals for flight risk

Somewhere in your inbox is a demo where every employee carries a risk score, and every Monday a manager receives the names trending the wrong way. It’s the idea we get asked about most, and the one thing we’ve deliberately left off a map of 122 agents. Four reasons why, and the first has nothing to do with privacy.

The impact thins the moment you ask what happens next. Nobody argues with the target, since losing people you’d have kept is worth spending money to prevent. What the demo can’t answer is what a manager does differently the morning a name climbs, because the answer is a conversation they could have had anyway. The score carries almost none of the work.

A score arrives before judgment does. Once someone has read a number next to a name, they can’t go back to reading the person. The stretch project goes elsewhere, the promotion case waits a cycle, and none of it gets logged as caused by the score. When the person leaves, the model looks right.

It wants a join nobody should make. Individual prediction needs performance records, survey responses and movement history lined up against one named person. The moment survey answers feed a scoring pipeline, the confidentiality promise on your engagement survey is broken, and your response rates will notice before your legal team does. Below a few thousand people there aren’t enough past exits behind any one profile for the estimate to hold still, so the number moves while the person doesn’t.

The law has arrived. AI used to evaluate workers or to shape decisions about their employment is high-risk under the EU AI Act. And once a score is the thing that settles an outcome, Article 22 of the GDPR gives that person a route to have a human take the decision instead. The ICO’s guidance on monitoring workers sets the UK expectation, and where the EU information and consultation directive applies your works council gets the conversation first. They’ll read a per-person score as monitoring, and that reading is fair.

Trust to earn: The most. No impact estimate and no data-readiness answer rescues an idea sitting there. The scoring walk-through sits in how to prioritize HR AI use cases.

This is where our one fixed rule lands. We don’t build systems that make autonomous decisions about people. AI may inform, surface, rank, summarize or draft. Every call about a person gets made by a person, and the system is built so it can’t do otherwise.

Two designs get treated as one thing here. “A person decides” and “a person can override” are different builds. Override is a correction applied to an answer the system already reached, and corrections get rarer the longer the system looks right. Deciding puts the person ahead of the answer: they work the segment in full, including the part nothing flagged, and no outcome has a route that goes around them. Which controls end up in your build is a decision you make. We put our recommendation in writing.

What works at cohort grain

Three builds do the job the scoring dashboard promised. All three carry high sensitivity and sit late in a sensible sequence, because they need clean movement history, agreed definitions, and a works council that already trusts how you handle people data.

Build	What it produces	Who decides	The guardrails that make it safe
Attrition driver analysis	The cohort factors associated with leaving: tenure band, manager span, pay position in band, time since promotion, location, function	People Analytics lead reviews, HRBPs choose which hypotheses to test	Segment grain only, no per-person score or list, minimum group size on every cut, pay as band position with performance ratings excluded, every run and reader logged
Cohort attrition forecasting	Expected exits over the next two to four quarters by function, level and location, with ranges, assumptions and a backfill estimate	Workforce planning and TA leads review it and set the hiring plans	Cannot be queried below segment grain, movement and tenure data only, no performance or person-level survey data, assumptions and model version logged with every forecast
Exit interview synthesis	Quarterly themes by function, level, tenure and manager span, with de-identified quotes	CHRO and HRBPs review the themes, employee relations reads anything flagged	Aggregate only above minimum group size, no manager scores, comments naming a person or alleging misconduct reach a human unsummarized, leavers told upfront, interviews opt-in

The forecast is the one to backtest. Hand it history up to eight quarters ago, let it forecast the quarters that have since happened, and check whether the actuals land inside its ranges. Under a few thousand people, expect ranges wide enough that your planner may still prefer their own estimate. That’s an answer too, and it beats tightening the model until it sounds confident. When cohorts become funded programs, the retention investment review agent is the step after, ranking segments on Finance’s cost assumptions for HRBPs and the CHRO to fund. It names nobody.

The first thirty days

None of this needs a vendor.

1. Agree the definition. One written page, function heads in the room, the two tests for regretted, and the point in the cycle where the flag gets set.
2. Recut the last two years. Tenure band by function by level, minimum group size held throughout. Most teams find a couple of pockets carrying the weight, and they’re quieter than the ones getting attention.
3. Read the files by hand. For the leavers you’d have kept: manager, pay-band position at exit, months since last move, whether a career conversation ever happened, what the offer promised against what the job became. A dozen files read properly beat a year of exit codes.
4. Say the constraint out loud. Where the driver is pay position or workload, that sentence belongs in the exec meeting, with a cost and a name attached to it, not in a program plan.
5. Fund one segment, with an owner outside HR. One intervention, one date, one number somebody else reports, and a note naming the result that would end the funding.
6. Write down what you won’t do with attrition data. No per-person risk scores. No manager-visible lists of who might leave. Minimum group sizes on every published cut. Having that page written makes you a faster buyer of everything else.

Start with the definition, then the cut

The distance between a turnover number and a turnover decision is one afternoon of definition work and one afternoon of cutting. Do both before anyone builds anything, since every AI idea in this space runs on the same history and inherits the same definitions.

Then put the retention idea on your desk through the 10X HR-AI Framework. Eight questions, and it lands on one of four tiers. The fourth tier tells you to wait, with the thing to fix first written next to it. For the wider view, 122 HR agents across 12 functions sets out who decides on each.

Sequencing that whole list, deciding which builds your movement history can support, and attaching an eLTV number to each is the strategy month. The steps and the commercials are written out here, next to the four kinds of employee experience consulting if a survey program or a platform rollout is the other thing on your desk.

Frequently asked questions

What’s a good employee turnover rate?

There’s no benchmark worth managing to, because the published ones blend industries, geographies and job types that behave nothing alike. What helps is your own segments against their own history, cut the same way each time, with the definition frozen. A group whose rate doubled while the company number stayed flat is worth a morning. Beating an industry average tells you nothing.

Do exit interviews tell us the real reason people left?

Partly. People leaving stay polite, they want the reference, and the real version often arrives months later through someone who stayed. Treat exit reasons as one weak signal, then check them against things nobody can edit: pay-band position at exit, months since last move, whether that manager had exits elsewhere.

Can AI predict who’s going to quit?

Something will always return a number when you ask it to. The question is what you’d do with that number and what having it costs you. Per-person scores buy an action a manager could take anyway, and they spend more trust than anything else on the list. Cohort forecasting answers the planning question at segment grain, with ranges and assumptions your planners can argue with.

Should managers see attrition data for their own team?

Aggregate patterns above a minimum group size, yes, and it makes them better at the job. Per-person risk flags, no. The moment a manager holds a prediction about someone reporting to them, the relationship changes, and nothing you write in a training deck changes it back.