The first surprise: you cannot see last year
Most workforce planning projects stall in week two on a discovery nobody expects. Someone asks for headcount by department as of last March, and it cannot be produced. In a great many human resources systems the employee record is updated in place: a transfer overwrites the department, a promotion overwrites the title and the grade, a location change overwrites the location. Today's roster is perfect. History is a set of effective-dated rows if you are lucky, and a single current-state record if you are not.
The fix is the cheapest high-value thing in this article and it takes an afternoon. Write a scheduled job that copies the full roster into a dated table on the first of every month, forever. Every field, every person, one row per person per month, append only. In three years you own a longitudinal asset almost nobody in your industry has, and you can answer questions about tenure, movement, span of control and cost that are currently unanswerable at any price.
Do it before the planning project rather than as part of it. If it is already running, you are a year ahead of where most companies start.
You are probably here because
- The finance headcount plan and the recruiting plan disagree and nobody can reconcile them
- You are always a quarter behind on hiring for roles that take four months to fill
- Somebody asked which skills the company will be short of in eighteen months
- Attrition jumped in one function and nobody can say whether that is a trend or noise
The reconciliation is the deliverable people end up valuing most, and it is the least technical thing on the list.
Job titles are not data yet
Pull the distinct titles. A company of two thousand eight hundred people routinely has three to four thousand of them. Analyst II, Senior Business Analyst, Sr. Bus. Analyst, Business Analyst - Senior and Analyst, Business (Sr) are five strings and one job. Free-text title fields, acquisitions that brought their own scheme, and managers who write a title to make an offer more attractive all contribute.
You cannot plan on that, and no amount of clever clustering fixes it unaided. What works is a two-stage pass. Group titles mechanically first — normalise the text, embed it, cluster it, and sort the clusters by headcount. Then have people who know the business review the mapping, top-down by headcount. The top two or three hundred titles usually cover eighty to ninety percent of your people, which turns an impossible task into a few afternoons with business partners in a room.
Keep the taxonomy shallow. Job family, sub-family, level. Three tiers is enough for planning and a five-tier scheme collapses under its own maintenance within a year. Write down the rule for what makes a new node, and make one person responsible for approving them.
Three plans, three grains, one reconciliation
Most companies are running three planning exercises and treating them as one, which is why the numbers never agree.
The financial plan is full-time equivalents and cost, owned by finance, at department and quarter grain. It is a budget, and it is the number the board sees.
The position plan is individual roles, owned by the business and executed by recruiting, at requisition grain with start dates. It is an operational schedule.
The capability plan is what skills exist and what will be needed, usually owned by nobody, at role-family grain over years.
The first two are the same population counted differently, and they disagree because of timing. Finance books an approved role from the start of the quarter; recruiting knows it will not start until the eleventh week; the department is carrying an open requisition it does not intend to fill. Gaps of five to ten percent between the two plans are ordinary, and the point of a planning system is not to eliminate that gap but to explain it, line by line, on a schedule. Publish the bridge monthly. That single artefact removes more friction than any forecast.
Predicting how many is realistic. Predicting who is not.
Aggregate attrition is genuinely forecastable. Voluntary departures by role family, tenure band and location have stable structure. Tenure dominates: the hazard of leaving is high in the first year, falls, and rises again at recognisable points. A discrete-time survival model is the right tool because the risk changes over time, and a well-fitted one predicts next quarter's separations for a function of two hundred people closely enough to plan around.
Individual flight risk is a different matter, and the gap between the two is where these projects lose credibility. A typical individual-level model produces a top decile in which perhaps twenty to twenty-five percent leave within a year, against a base rate around fourteen. That is real lift and it is nowhere near enough to act on for a named person. Four out of five people on that list are not going anywhere, and the model cannot tell you which.
There is also a practical argument, separate from the statistics, for not distributing individual scores to managers. It changes behaviour in ways you cannot govern, it leaks, and the day an employee discovers they were on a list, you have created the outcome you were trying to avoid. Report aggregates to planners. If you want to give managers something, give them the drivers — time since last compensation change, span of control, tenure in seat, internal move history — which are actionable, and let them apply judgment to their own people.
| Question | How well it can be answered | What to build |
|---|---|---|
| How many will leave this function next quarter? | Well, with an interval, from twelve to twenty-four months of history | Survival model by family, tenure and location. Plan against the interval |
| Which named person will resign? | Poorly. Real lift, not enough precision to act on | Do not distribute. Publish drivers instead of scores |
| How long will this role take to fill? | Well, if requisition history was retained | Median and 80th percentile days to fill by family and location |
| How many people do we need to serve the demand? | Well, if you can name the driver | A ratio to an operating driver, back-tested for stability |
| What skills will we be short of in two years? | Poorly from self-reported data; moderately from what people did | Use assignments, certifications and systems access, not surveys |
| What will next year cost? | Well, and it is mostly arithmetic once the roster and rates are right | Cost model on the same taxonomy, with the timing assumption visible |
Tie headcount to a driver, not to last year plus five percent
The single change that makes a plan defensible is anchoring demand to something operational. Nursing hours to patient days. Field technicians to installed base and service call volume. Support headcount to ticket volume and to the release calendar, since a big release moves tickets more than customer growth does. Billable consultants to signed backlog and target utilisation. Warehouse labour to units shipped and to the mix, because a small-parcel unit and a pallet are not the same work.
Then check the ratio's history before trusting it. Plot it for three years. If it has drifted, find out why, because there is always a reason and it is usually a product change, an automation project or a policy change that nobody connected to headcount. A ratio that moved twenty percent over three years is not a constant, and using its current value as a planning assumption bakes in an error you will discover at the worst moment.
Where the work goes in a first planning build — our allocation
Weights sum to 100. Our starting allocation for a company of one to eight thousand people. Judgment, not measurement.
Time to fill is why planning exists at all
If every role could be filled next week, planning would be a monthly conversation. Median time to fill for common roles commonly sits somewhere between forty and sixty days from requisition approval to accepted offer, and scarce roles run past a hundred and twenty. Add notice periods and onboarding and a role approved today is productive well into the following quarter.
That lag has two consequences worth designing around. Plan at a horizon longer than your longest fill time plus ramp, which for most companies means five to seven quarters rather than four. And use the eightieth percentile, not the median, for anything on a critical path; half your critical hires arriving later than assumed is not a plan, it is a coin flip repeated.
Internal moves are the lever nobody models
In many companies a large share of role fills come from inside, and every internal move creates a backfill somewhere else. Model the chain or the plan is wrong in a specific direction: it understates recruiting load and overstates net capacity, because one external hire at the bottom of a three-move chain is what actually resolved three openings, over three times the elapsed period.
The data for this is in the monthly snapshots, which is another reason to start them today. Movement between two snapshots is a transfer; disappearance is a separation; appearance is a hire. That is the whole method, and it works without any additional instrumentation.
Skills inventories are a second project, not part of the first one
A skills graph is attractive and it is a twelve to eighteen month programme with an ongoing maintenance cost. Self-reported skills decay fast and are inconsistently rated, so the useful sources are traces of what people actually did: project assignments, certifications with expiry dates, systems access, categories of work completed. Get the roster, the taxonomy and the driver ratios working first. If capability planning is genuinely the pressing question, scope it separately with its own sponsor.
Scenarios, not a number
A plan that produces one number will be wrong, and being wrong is fine; being wrong with no visible assumptions is what destroys trust. Produce three scenarios that differ in the two or three assumptions that actually move the answer — usually demand growth, attrition, and the fill rate you can sustain — and show the sensitivity. If a plan collapses when attrition moves two points, the leadership team should know that before they approve it, not in April.
Keep the scenario set small. Three named cases people can hold in their heads beat a tool that can generate a hundred and therefore commits to none.
When you do not need software for this
Under roughly three hundred people in one or two locations, a well-kept spreadsheet plus the monthly roster snapshot is the right answer, and building anything more is a distraction. The leadership team can still hold the organisation in their heads and their judgment is better than any model you could fit on that little data.
The threshold is organisational rather than statistical. Build when no single person can name every open role and who is likely to leave, which usually lands somewhere between four hundred and eight hundred people, or earlier if you are spread across many sites or have grown by acquisition. Even then, the first build is a monthly refreshed model and a dashboard, not a platform. We have talked companies out of the platform version more than once, and the ones who took that advice were still using the simple thing three years later.
What a usable planning model contains
- Monthly roster snapshots, append-only, every field, running before anything else starts
- A three-tier role taxonomy with a named owner and a written rule for new nodes
- Attrition rates by family, tenure band and location, with intervals, refreshed quarterly
- Time to fill at median and eightieth percentile by family and location
- A named operating driver per function, with three years of ratio history plotted
- Internal moves and their backfills modelled as a chain, from the snapshot differences
- A monthly bridge between the finance plan and the recruiting plan, explained line by line
- Three scenarios with explicit assumptions and a sensitivity on the two that matter
The mistakes we get called about
- No history, because the system overwrote it and nobody was snapshotting
- Planning on raw job titles, producing families of one person
- Individual flight-risk scores sent to managers at a precision that cannot support the action
- Median time to fill used for critical roles, so half of them are late by construction
- Headcount set as last year plus a percentage, with no driver behind it
- Internal mobility ignored, understating recruiting load and overstating capacity
- A four-quarter horizon in a company whose scarce roles take five months to fill
- A single-number plan whose assumptions were never written down
Bottom line
Start the monthly roster snapshot today, whatever else you decide, because it is an afternoon of work and it is the one thing you cannot buy retroactively. Fix the titles next, top-down by headcount, with people who know the business. Then model the aggregate rather than the individual, tie every function's demand to a driver you can name and whose history you have checked, plan at a horizon longer than your longest fill time, and publish the bridge between the finance plan and the recruiting plan every month. Almost none of that is modeling. All of it is what makes the modeling worth doing.
Frequently asked questions
Twelve months gives a usable base rate for large functions; twenty-four to thirty-six lets you separate seasonality from trend and fit tenure structure properly. For small functions, pool across similar families and report an interval rather than a point number. If you have no history at all because the system overwrote it, start snapshotting now and use industry benchmarks with wide bands in the interim.
With real but modest accuracy. A well-built model concentrates departures in its top decile, typically raising the rate there by half again or so over the base rate, which still means most people flagged are staying. That is useful for planning totals and not sufficient to act on for a named individual. Report aggregates for planning and give managers the drivers rather than the scores.
Because they are counting the same people on different clocks. Finance books an approved role from the start of a period; recruiting knows the start date will land weeks later; departments carry requisitions they do not intend to fill. A five to ten percent gap is normal. The deliverable is a monthly bridge that explains the gap line by line, not a project to make the two numbers identical.
Normalise and cluster the titles mechanically, sort clusters by headcount, then review from the top with people who know the work. The largest two or three hundred titles usually cover most of your population, so a few working sessions gets you to good coverage. Keep it to three tiers, give it an owner, and accept a long tail of unmapped titles rather than delaying the whole thing for completeness.
Longer than your longest time to fill plus ramp, which for most companies means five to seven quarters rather than the usual four. Roles that take four months to fill and three months to become productive are decided more than half a year before they contribute, and a four-quarter plan structurally cannot see them.