The 98% Problem

The difference between a workforce forecast at ninety-five per cent and one at ninety-eight is not three points of arithmetic. It is whether the reforecast can be acted on in the room where it is presented, or has to be rebuilt from the bottom up first.

Those points are not available from a planning tool. They are held in a small set of records where the employment fact has changed and the cost consequence has not caught up. Governing that set is a different activity from planning, and almost nobody staffs it.

What actually happened

Polymath’s founder ran global human capital planning in a CPG group of roughly $2–3 billion in revenue that had been private and had since listed on the TSX: a workforce P&L above $200 million, payroll across approximately twenty-one countries. Workday held the employment record. He built the headcount model on top of it in Adaptive Planning.

The model took weeks. The definitions underneath it took years, and they are the reason the forecast held at about 98% across those twenty-one countries.

What matters commercially is not the figure but where the remaining error lived. It was not spread across the settled population, whose pay was known exactly and whose circumstances had not changed that month. Those records are arithmetic. The error sat in the records in transition.

It looked like this. A requisition approved and budgeted before the position existed in the establishment, so the plan carried cost from a date the system of record did not recognize. A leaver whose termination fell in one period while the vacation payout, notice and statutory contributions landed in the next. A promotion agreed in a meeting, effective the first of the month, entered three weeks later. Contract labour absent from the employment record and present in the labour line.

So the function stopped reconciling at the total and started at the event. A position became an object rather than a role description: an identifier, a level, a cost centre, an effective start and end, a country payroll calendar, and a rule about who could create one. Vacancy logic was written down rather than assumed: what a vacant position costs, from which date. Hiring timing became a dated commitment held by the hiring manager instead of an average applied by planning. The payroll integration carried effective dates rather than period totals, because the period a cost falls into is settled in payroll.

The planning tool did not change. Most of the rules fitted on a line.

Accuracy did not come from forecasting the workforce better. It came from defining the position and dating the handful of records where the workforce was actually changing. Define the position, date the event, let the tool do the arithmetic — that is what Polymath builds for a client before anyone looks again at the forecast.

Why it happens

Workforce cost is the only material line on the P&L produced by a system whose purpose is not financial. The ledger exists to state financial position. The employment record exists to administer employment: hire the person, pay them correctly, enrol them in benefits, keep them lawful wherever they sit. Cost is a consequence of that work, not its object.

One precise effect follows. An employment record becomes true when the employment fact becomes true: the candidate accepts, the manager approves, the resignation is tendered. A cost becomes true when an accounting period says it does. Between those moments sits an interval, and the interval differs by event type, by country and by payroll calendar.

For the settled population that interval is zero. Nothing is changing, so the employment date and the cost date coincide, and a planning tool with correct compensation data is simply right. The error is in the moving records, a small share of the population and the entire subject of any reforecast.

This is why aggregate remedies never reach it. A better model refines the part that was already correct. An attrition allowance applies an average to individually dated events, each with a country attached. More history helps only if transitions behave this year as they did last, and they do not: they move with policy changes, acquisitions and whichever hiring manager was in a hurry.

A second property makes the error worse than random. Seam error has a direction. Requisitions are optimistic about start dates, because optimism is what a hiring manager feels when raising one. Leaver costs are late, because payment follows departure by construction. Pay changes agreed offline are backdated, never forward-dated. Because each seam leans consistently, the errors do not cancel. They compound in the direction of whoever creates the record.

So the forecast becomes something people correct for privately, each with their own adjustment, and the discussion turns into an argument about whose adjustment to use. None of that is recoverable from the total, which carries no information about which transition produced the gap. Worse, the rules decay: every new country, acquired entity and policy change creates a transition nobody has written a rule for.

The strongest objection

The strongest case against this: we run at about ninety-five per cent and the last points cost more than they are worth.

For an annual budget submission and a board-level view of labour cost, that is correct. Three points on a workforce line sit inside the tolerance of the decisions those artifacts serve. If the plan exists to set an envelope once a year and defend it, the work below will not repay the effort.

It stops being correct the moment a reforecast is used for a decision smaller than the error. Real workforce decisions are small and specific: release eleven positions in two regions, extend a hiring pause, decide whether a soft trading month changed the labour run rate. When the decision is smaller than the range around the number, the number cannot settle it, so the meeting adjourns for a bottom-up check by the managers who own the roles.

The second objection is better: not every seam is worth closing. True. A country with forty people and a statutory calendar nobody can change deserves a convention, not a project. What we propose is narrower: the rules that repay themselves sit on three transitions, joining, leaving and off-cycle pay change. Those are also the three the business alters most often, which is why the work is never finished.

The method

We run six steps, in this order. Each narrows the next, which is why we hold the sequence.

1. We split the forecast into the settled population and the moving records, and we report them as two objects with two owners. We treat the settled half as arithmetic. The moving half is the forecast. The trade-off we accept: the moving set looks far too small to justify the attention it is about to receive. Skip it and every accuracy conversation will be about the model, which was never the half that was wrong.
2. We write a conversion rule for each transition type. Joiner, leaver, promotion, transfer, leave, contract labour, acquired population. For each we settle what evidence makes the event real, which date the cost starts and stops, and which record is authoritative when two disagree. The trade-off we accept: your HR and finance teams must disagree in the open about dates they have settled privately for years. Skip it and each analyst applies a private rule that changes when the analyst does.
3. We refuse a forecast line that has no establishment object behind it. If a role is in the plan, the position exists in your system of record with an identifier, a level, a cost centre and an effective date. We make no exceptions for urgency. The trade-off we accept: this is genuinely slower for your hiring managers in a growth year. Skip it and the plan carries headcount the establishment has never heard of, and the variance cannot be attributed.
4. We build the departure cost profile once, per country. Notice, accrued leave payout, severance practice, statutory contributions, benefit run-off, equity treatment. We treat leaving as the most predictable one-directional error in the line, and it is knowable in advance. The trade-off we accept: your payroll team must expose mechanics it treats as routine housekeeping. Skip it and every quarter closes with a labour line right about people and wrong about money.
5. We reconcile forward, record by record, at each close. We do not explain variance in aggregate. We take the moving records, match each to what actually happened, and we classify every miss by the seam that produced it. The trade-off we accept: it is tedious and it recurs monthly. Skip it and you will know you were wrong without learning where, which is how a forecast holds the same accuracy for four years.
6. We reopen the seam list after every acquisition, new country and policy change. Each creates transitions your rules do not cover: a different notice regime, a payroll calendar that lands elsewhere, a pay practice with no local equivalent. The trade-off we accept: we make the rule set a maintained asset with an owner rather than a document written once. Skip it and accuracy decays silently and returns as a surprise variance two quarters after the deal.

Note what we do not bring. No new planning platform, no forecasting technique. The work is definitional, and it is the part of job architecture nobody builds. The classic spine — job, family, level, skills — carries none of the attributes a forecast needs: effective dates, cost centre, payroll calendar, vacancy state, and the right to create a position. We extend it with those, and Forecast-to-Plan becomes arithmetic over well-defined objects. A planning tool earns its place on cycle time; it holds no view on whether a leaver’s final payment belongs in March or April, and that is where we find the last points.

Evidence and measures

Where an employment event and its cost part company, and the rule that closes each:

Five measures worth tracking over ninety days. None requires a new system:

— Share of last month’s workforce variance attributable to named records within five working days of close.

— Transition types with a written conversion rule, against the transition types that occurred last quarter.

— Days between an offer being accepted and the cost appearing in the plan, by country rather than averaged.

— Proportion of leavers whose full cost, including trailing payments, landed in the period the forecast put it in.

— Reforecasts sent back for a bottom-up rebuild before anyone would act on them. That tells you what the number is currently worth.

If those five move, the seams are governed. If only the headline percentage moved, you had a quiet quarter.

How Polymath solves this

We begin with two closes and a list of names. We decompose your recent workforce variance to individual records, which settles where the error actually sits, and then we redefine the position as an object: identifier, level, cost centre, effective start and end, payroll calendar, vacancy state, and who may create one. That is the Job Architecture 2.0 spine — job, family, level and skills extended with event and decision-right attributes — and without them Forecast-to-Plan has nothing to attach a cost to. We run it as Foundational Architecture, and what we hand over is a definition set rather than a forecast.

This happens in the close itself, month by month. We have the conversion rules for joiners, leavers and off-cycle pay changes written by your payroll leads, HR operations and finance in one room, because the disagreement is theirs. Then we configure them rather than document them: position rules in the HCM platform, vacancy and hiring-timing logic in the planning layer, effective dates carried through the payroll integration. We reconcile one close record by record with your analysts, and they run the next one.

What compounds is the position object, because it outlives the forecast that motivated it. The same definitions later date a restructuring or a hiring freeze, tell provisioning and approval routing what access a role carries, and give an acquired population somewhere to map into. Each use reads the same object instead of keeping a private list. Nothing after this has to define a position again: the restructuring, the freeze and the acquired population all read an object the forecast already paid for.

What it costs to do nothing

The first cost is rebuilt work. A reforecast that cannot be trusted on presentation is reconstructed by the managers who own the roles, in the weeks they should be interviewing.

The second is that imprecision forces blunt instruments. When the workforce line carries a range wider than the decision in front of you, the only defensible action is the broad one: freeze everything rather than release positions where the plan is short. That costs more, lands on the wrong teams, and takes two quarters to reverse.

The third is cash timing. Trailing termination costs and statutory contributions across twenty-one countries do not land in the period a total-level forecast implied. The quarter that absorbs them is the quarter with an unexplained labour variance, and in a listed business that is explained on the record.

A workforce number precise enough to act on is not a reporting virtue. It is the difference between deciding on the largest controllable line in the business and negotiating around it, and that difference begins with two closes and a list of names.