People Analytics & Headcount Data with Cole Napper


Podcast Overview


    This Podcast treats headcount as a dataset, not a report. It is a companion to our Headcount Experts conversation with Cole Napper, author of People Analytics and host of Directionally Correct. You will get a data model for headcount, his view on what AI changes, and a method for timing backfills.

    What is people analytics?

    People analytics is the practice of using data about a company's workforce and hiring to answer business questions: how many people the company needs, what they will cost, where hiring stalls and which plans are at risk. It differs from HR reporting in one way. Reporting describes what happened; analytics explains why and forecasts what comes next.

    The discipline sits between HR, finance and recruiting. Each of those functions owns part of the data, and none of them owns the whole picture. That is why the first deliverable of most people analytics teams is not a model. It is a reconciled count.

    From thermometer to thermostat

    Cole Napper frames it like this: A thermometer reports the temperature. A thermostat sets it and acts to hold it. "The thermostat says it's 72 degrees in the room. If it gets too hot out, we're gonna turn on the air conditioner. If it gets too cold outside, we're gonna turn on the heater, but we're staying at 72."

    • A thermometer function reports headcount, hires and departures after the fact.

    • A thermostat function sets the target, usually the approved hiring plan, and flags when reality drifts from it: a start date slipping, an offer landing above budget, a seat about to go empty.

    A thermostat still needs an accurate thermometer. If the count of positions is wrong, the function acts on the wrong signal. That is why Cole's "people intelligence maturity curve" starts with foundations rather than with models: in his words, it is about "how to set up the right foundations to be able to scale to solve any type of problem."

    Where analytics adds the most value

    Analytics pays off where there is enough repetition to learn from. Cole puts it bluntly: "Almost all of analytics is only useful in multi-incumbent, high-turnover roles." A company of 500 people in 499 different roles gives a model little to work with. A company of 500 people where 400 share one role, such as inside sales or customer support, gives it plenty.

    What people analytics is not

    It is not a dashboard tool, and it is not a survey program. Dashboards are an output. Engagement surveys are one input. The work is defining the entities, keeping them consistent over time and turning changes in them into decisions a business leader can act on.

    Why is headcount data the foundation of people analytics?

    Headcount data is the foundation because almost every people analytics question is a question about positions over time: how many, at what cost, filled when. If the count of positions is wrong, or if it only records the current state, every metric built on top inherits the error. A good headcount dataset records the plan, the forecast and the outcome separately.

    Count positions, not just people

    An HRIS counts employees. A plan counts positions, including ones nobody holds yet. People analytics needs both, joined. The useful unit is the position: a seat the company has decided to fund, which may be filled, open, being recruited for or closed.

    Three position types cover most cases:

    • New hire: a seat that did not exist before.

    • Backfill: a seat vacated by a departure and funded again, linked to the person who left.

    • Vacant seat: a departure with no backfill yet. It is still part of the funded organization until someone decides to close it.

    The last one is where many models undercount. If a departure simply drops out of the count, the company's committed cost looks lower than it is.

    The people in those positions are not all W-2 employees either. "I wouldn't necessarily say headcount is payroll," Cole says. "Payroll is a component of it." He thinks in full-time equivalents: full-time and part-time staff, contractors, contingent workers, consultants and, increasingly, AI agents tied to a workflow with money flowing to them. A position-level dataset can hold all of these, as long as each one is tied to the funded seat it fills.

    Keep plan, forecast and actual values apart

    A position's title, level, salary and start date change between the day it is budgeted and the day someone starts. Overwriting those values destroys the most useful signal in the dataset: how far reality drifted from the plan. Store them as separate families instead.

    Headcount data model for people analytics: initial plan, current plan, actual offer and HRIS record for each position

    Each position carries four versions of the truth: the initial plan, the current plan, the accepted offer and the HRIS record. Cole Napper explains why analytics built without one source of truth "is gonna fall down" at 24:55

    Initial plan against current plan measures slippage. Current plan against actual measures offer variance. Actual against the employee record catches data quality issues between recruiting and HR systems.

    Track the state of each position

    A position moves through a lifecycle: drafted, pending approval, approved, open for recruiting, posted publicly, offer pending, offer accepted, filled. It can also be rejected, put on hold or removed from the plan. Recording when each transition happened turns a static count into a time series. That is what makes time-to-approve and time-to-fill measurable without a manual tracker.

    What happens when the foundation is missing

    To an analyst, the need for one source of truth sounds obvious. Cole agrees: "If you're trying to build the foundation of your house on sand, in terms of not having a core source of truth, that house is gonna fall down." Automation makes the problem worse, not better. If the process has a fault, "you're not going to automate the process, you're going to automate the fault."

    He lists what a sound foundation needs: standardized data, core definitions, a system where the source of truth lives, a way to audit it, and an agreed tolerance for error. Backdated changes are the classic example. Someone moves teams on the last day of the month, the change is entered five days later and backdated, and last month's report is now wrong. That is tolerable when the procedure expects it. It is not tolerable when it is a symptom of "a completely leaky bucket of headcount information."

    HR data makes this harder than in most functions. "As far as data goes, HR has some of the most challenging data of any function," Cole says. It is spread across the HRIS, the ATS and finance tools, and it changes every time someone is hired, moves or leaves.

    How is AI changing headcount planning?

    AI is breaking the assumption that one unit of extra work means one more hire. Headcount planning used to be fairly linear once demand was known. Now each new need may be met by a person, an AI agent or both, so the plan has to track cost and output across people and agents, starting from an accurate count of people.

    The hard part was always demand

    "I've always thought that headcount planning was a quite linear process as long as you had a good demand function," Cole says. The difficult part is forecasting demand, which requires being embedded with the business. He compares good people analytics teams to forward-deployed engineers: they sit with finance, sales, operations and R&D, not behind a laptop.

    The one-to-one ratio is broken

    Companies used to hire on a one-to-one basis: one unit of extra work, one more role, posted, promoted or trained. "That one unit to one unit ratio is broken because of AI," Cole says, "and so I think actually for the first time headcount planning, the linear part of it is not so linear anymore." Planning now means weighing what he calls tokenomics, the spend on AI usage, against the spend on headcount.

    A bottleneck caps the gain

    Cole draws on Lean Six Sigma to explain why AI gains often fail to show up. If AI speeds up steps one to three of a process but step four is still manual, "step four becomes the only thing that matters anymore." He has not yet seen a process automated end to end. Until that happens, the human positions around the bottleneck still set the pace, and the headcount plan still decides the outcome.

    What this means for headcount data

    Every comparison between people and AI needs a human baseline: what the positions cost, when they start and what they produce. Without an accurate headcount plan, there is nothing to compare AI spend against. [DATA NEEDED: a credible survey figure on how many companies track AI spend alongside headcount spend, with source.]

    Worked example: measuring plan drift in an engineering hiring plan

    Plan drift is the gap between the hiring plan a company approved and the one it is executing. Measuring it takes three numbers per position: the initial plan value, the current plan value and the actual value. The example below uses a hypothetical company to show the method; the figures are illustrative, not benchmarks.

    The setup

    A software company approves 40 new engineering positions for its plan year, each budgeted at $160,000 base salary. Six months in:

    • 12 open positions have had their expected start dates pushed back by an average of 60 days.

    • 18 positions have accepted offers. Their combined actual base salary is $2,995,200.

    • Of those 18, the salary was changed on 7 positions. Two changes were made before recruiting started; five were made while the offer was being finalized.

    Plan drift example: 18 of 40 engineering roles filled, yet $315,600 unspent, 720 days of capacity lost, +$115,200 in offers

    A report that says 18 of 40 roles are filled hides $315,600 of unspent salary and 720 days of lost capacity. Hear the hosts and Cole on how headcount relates to the hiring plan at 13:07

    What it tells the business

    A headcount report would show 18 of 40 positions filled. The drift analysis tells a different story. Finance will see an underspend this year, but engineering is short roughly two engineers' worth of capacity, which is a delivery risk, not a saving. Meanwhile, the positions that did fill will cost $115,200 a year more than budgeted, and that cost carries into every future year.

    The timing of the salary changes is the diagnostic. Changes made before recruiting point to a planning problem: the budgeted rates were stale. Changes made at the offer stage point to the market: candidates are closing above band. Those call for different fixes, and only a dataset that records when each change happened can tell them apart.

    Edge cases to handle

    • Multiple currencies. Compare salaries within one currency, or convert at a recorded rate. Never sum amounts in different currencies as they stand.

    • Positions on hold. A paused position is still in the plan and the budget. Exclude it from time-to-fill, but keep it in planned headcount.

    • Removed positions. A position taken off the plan is not a hire and not a vacancy. Keep it in the history so the original plan can still be reconstructed.

    When should you open a backfill?

    Open a backfill early enough that the replacement is ramped before the current employee leaves. In practice, that means working back from the expected departure by the time to fill plus the time to ramp. It works best in roles with many people doing the same job, where departures are frequent enough to predict.

    The employee lifetime value curve

    Cole describes the value of an employee over their tenure as a curve. "When a human being starts a role, they're usually at a net negative value add to the organization," consuming resources before producing output. They reach break-even, usually level off somewhere between 18 months and three years in role, climb in steps with growth and promotion, and "eventually [the curve] starts to slide down as a person is exiting the organization."

    Employee lifetime value curve: return starts below cost, passes break-even, steps up at promotion, peaks, then declines

    New hires start below break-even, level off between 18 months and three years, and decline toward exit. Cole walks through the curve at 17:43

    Two ways to backfill the same seat

    • Backfill at the exit. The requisition opens when the employee resigns. The seat produces nothing through recruiting, then the replacement starts the curve again from below cost.

    • Backfill at peak return. The requisition opens when the employee reaches peak value. The replacement is already ramping when the employee leaves. There is a short overlap with two salaries, and the seat never goes dark.

    Backfill timing compared: opening a backfill at exit leaves a vacant seat; opening at peak return leaves no gap

    Backfilling at the exit leaves a vacant seat; backfilling at peak return leaves no gap. Cole on opening requisitions early so they overlap, at 36:57

    A worked example: inside sales

    On the episode, headcount365 founder Eric Guidice described an inside sales team where reps moved through the pay tiers over about 18 months, then typically had six to twelve months before they left. Using illustrative numbers for the rest:

    Inside sales backfill timeline: open the requisition by month 19, 60 days to fill, 90 days to ramp, productive by month 24

    Eric's inside sales example starts at 35:28

    Opening at month 19 means the replacement breaks even by month 24. Opening at the resignation means about five months of lost or reduced output for every departure.

    The overlap has a cost. If the employee stays until month 30, the company pays two salaries for longer than planned. This is why the method works best in multi-incumbent roles. On a team of 40 reps, you don't time each backfill to one person. You forecast how many reps will leave each quarter from past turnover, and keep that many requisitions open five months ahead. Cole's description of the same approach: predictive models "that allow you to predictively open a rec so that it overlaps, so that when a person is sunsetting, another person's being trained up and you don't see this huge dip in productivity."

    When not to use it

    • Single-incumbent roles. With one person in a role, there is no history to predict from. Plan succession instead.

    • Roles with long notice periods. Where notice runs three months or more, the notice period itself covers part of the lead time.

    • Positions being redesigned. If AI or a reorganization may change the role, decide that before reopening it.

    Should a people analytics team build its own headcount data pipeline?

    A capable people analytics team can build a headcount data pipeline. The harder question is whether it should own the system of record that creates the data. Analysis can live anywhere. The plan, its approvals and its change history have to be captured at the moment decisions are made, or they are lost.

    That distinction matters. Most in-house headcount models are reconstructed after the fact from spreadsheets, ATS exports and HRIS snapshots. They can show where the plan is now. They usually cannot show what the plan was in March, who changed a salary, or at which stage of hiring it changed. No amount of modeling recovers history that was never recorded.

    Build vs. buy decision matrix for headcount planning data: when to build in-house and when to use a system of record

    When to build your own headcount pipeline and when to use a system of record. Cole on being a data consumer rather than a data producer, at 31:10

    Be a data consumer, not a data producer

    Cole frames the choice by the analyst's role. A data producer builds and maintains the pipeline: extracting from each system, reconciling definitions and fixing breaks. A data consumer starts from clean data and spends the time on analysis. Starting as a producer, he says, is like "staring at the blank screen of 'we have no data infrastructure, now it's my job to fix that.'"

    This matters more as the "talent engineer" title spreads, often filled by recruiters moving into technical work. Cole's view: "It's just data engineering. It's always been data engineering." It is hard work on hard data, and data engineers are in short supply. A new talent engineer will produce more value consuming a reliable headcount dataset than building one from scratch.

    Team size points the same way. Cole estimates a small organization needs no more than about three people in people analytics, "plus a good technology stack," and that even a very large one is hard to justify above about 20. A small team cannot afford to spend its time producing data.

    The last row is the split we recommend. Let a dedicated system capture the plan and its history. Spend the team's time on the models and recommendations that only it can produce.

    What to ask of any headcount data source

    Whether you build or buy, hold the data to the same standard:

    1. Does it record the initial plan separately from the current plan?

    2. Does it keep a timestamped history of every change, with who made it?

    3. Does it link a backfill to the departure it replaces?

    4. Does it count vacant seats until someone decides to close them?

    5. Can you get the data out, in full, into your own tools?

    If the answer to any of these is no, that gap will show up in your analysis later.

    How headcount365 handles headcount data

    headcount365 is a headcount planning system of record. It captures the plan, approvals and every later change as they happen, so the analysis layer starts from a complete history rather than a reconstruction.

    • The unit is the position. Each planned hire is a requisition on a plan year, typed as a new hire or a backfill. A backfill is linked to the employment record of the person who left.

    • Vacant seats stay counted. A departure synced from the HRIS leaves an open seat that stays in headcount and cost views until someone closes it, with a reason. A closed seat can be reopened.

    • Plan, forecast and actual are separate fields. Each requisition keeps an initial snapshot from when it went on the plan, the current expected values, and the actual values from the accepted offer, synced from the ATS or entered in the app. Once the hire is matched to the HRIS, the employee record sits alongside them.

    • Every change is recorded. Each requisition keeps a full history of who changed what and when, as it moves from draft through approval, recruiting, offer and start.

    • Reports are built on that history. Cycle-time reports give medians for approval time, preparation time, recruiting time and notice period. A risk report flags changes to title, level, salary and start date, and classifies each by the hiring stage it happened in. Spend reports cover existing and forecasted cost for the plan year.

    • Scenarios sit beside the plan. A plan year can have scenarios for testing alternatives without changing the approved plan.

    • Currency is explicit. Salaries are stored in local currency, with the currency code and the conversion rate applied.

    • The data leaves cleanly. It connects to HRIS and ATS systems, syncs to FP&A tools such as Pigment, and supports scheduled exports of requisition, employee and employment action data. [CONFIRM: whether to mention the MCP connector for AI assistants, currently enabled per customer.]

    • Recruiting start dates are worked back from the plan. Each requisition has an expected offer date, and each recruiter or recruiting team has a time to fill. headcount365 uses the two to show when recruiting needs to start, which is the date a backfill has to open to arrive in time.

    • Role Based Access Control. Each user sees only the positions their access allows.

    Next
    Next

    Headcount dashboards that score every seat