How to use this
Headcount is the one expense line where Finance, Talent, and HR each hold a piece of the decision, and none of you holds the whole thing. This guide works the same way, and unlike the PDF, it's built to work through live: jump to any section from the contents below, read start to finish, or send a teammate straight to the part they need. Every section ends with a short breakdown for Finance, Talent, and HR, so you can read your part fast or read it all for the full picture.
Share this with your Finance, Talent, and HR partners before you sit down to build the plan. Run the readiness score in Part 11 out loud, together, before you write a single line. If the three of you disagree on the answers, that disagreement is the most useful thing that happens in the room.
Part 11 isn't just something to read, it's a live scorecard. Check off each question as your team answers it and watch your headcount and AI readiness scores calculate in real time. Jump to the readiness score →
Why the 2027 plan is the first that has to price people and AI against one budget.
Five ways a headcount plan breaks, including the one hiding in your AI bill. A twelve week planning sequence with AI built into every phase, not bolted on after. A sixteen question readiness score for Finance, Talent, and HR to run together. Real salary benchmarks, an AI coworker built for planning, and Token Spend Management, the ledger that connects AI spend to your budget.
Your 2027 plan has to price people and AI spend as one budget
2027 planning asks you to redesign work, not forecast headcount
Every planning cycle before this one asked the same question: how many people, and at what cost. 2027 asks something harder. Which work needs a person, which work needs a person with better tools, and which work needs neither. Nobody in your company has answered that at scale before, and the answer changes org charts, job descriptions, and career paths. Treat this as the first step of a workforce redesign that runs across the company, because that is what it is.
Boards now expect an approval trail for every headcount decision
That tolerance disappeared with cheap money. Boards want to know which roles were approved, by whom, at what cost, and what happened to the ones that were funded and never filled. Those are controls questions, and most companies cannot answer them from a system.
AI spend is growing too fast to track after the fact. It has to be budgeted upfront.
Enterprise AI spend is growing fast enough that it can no longer sit outside the budget conversation. AI-native application spend rose 108 percent year over year across all company sizes, and 400 percent at organizations above 10,000 employees. Vendors are already converting seat pricing into token and action based pricing. Gartner recorded the demand side response: CFO headcount growth expectations falling from 6 percent to 2 percent between 2025 and 2026, which its analysts describe as a structural pivot from labor expansion to automation, not a side experiment.
Other companies have already named one owner for AI and headcount together
This is not only a TeamOhana argument. Atlassian recently gave its Chief People Officer responsibility for AI enablement, expanding her remit from a 700 person HR organization to a 3,500 person function. Their stated reason: cultural and technological transformation cannot be separated if the change is going to stick. Name one executive who owns the mixed workforce decision at your company, and do it before planning opens. Atlassian also reports that 85 percent of knowledge workers now use AI, but only 29 percent have actually changed how they work. That gap, access without redesigned work, is exactly where planning has to do the heavy lifting. It will not close itself.
Contractors and AI agents are already absorbing headcount work
A real share of 2027 work will not be done by an employee. Contractors, outsourced functions, and AI agents are absorbing work that used to become a req. Companies that cannot control their own headcount today have little chance of controlling a mixed human and AI workforce in 2027. The control layer has to exist for people before it can extend to agents, which is the practical argument for fixing this now rather than after the plan locks.
The choice is not which planning tool to use. It is whether your two largest, fastest growing expense lines, people and AI, are managed by one system, or reassembled by hand every month in two different places. Load and lock is week 12 of the sequence in Part 8, not something to figure out after the plan is already approved.
A hire, an AI upgrade, or a digital worker now compete for the same dollar
Planning conversations still start the way they have for a decade: what do we want to spend, and on how many people. That question is no longer complete. A hiring plan used to have one shape: hire, or do not hire. It now has three shapes, and all three come out of the same budget.
- Fund a new hire, who arrives with a salary and, increasingly, a real AI cost attached.
- Give more AI capability to someone already on the team, raising what they cost without raising headcount.
- Fund a digital worker that does the work instead, with its own budget and, in most companies, no approval process built for it yet.
KPMG frames the same underlying choice as buy, build, borrow, or bot: hire talent, develop it internally, use contingent labor, or use AI. Whatever you call the four options, the discipline is the same: price them consistently, or someone will guess.
The third option needs a name before you can budget it
An agent is a technical thing. It is an API key, a model, a registration in some system. You cannot budget an API key against a role for the same reason you cannot budget a laptop against a role. It is the wrong unit.
What you can budget is a digital worker: capacity you fund to do defined work, with a charter, a spend envelope, a cost center, and a human who owns it. One digital worker may run on several agents, models, and API keys, and those can change underneath it without changing the plan.
This is the same invention that made headcount plannable in the first place. A position is not a person. It is a funded seat that exists before anyone fills it, carrying a level, a budget, and a reporting line, which is why you can plan hiring a year before anyone is hired. The digital worker is the position object for the second workforce, and without it, your third option has nothing for a dollar figure to attach to.
Finance approves the hire. Someone in engineering or IT approves the AI seats. Nobody approves the digital worker, because there is no form for that yet. The 2027 planning problem is not forecast accuracy, but whether your plan can price a hire, a capability upgrade, and a digital worker against the same dollar and let someone choose.
Take your heaviest AI user in engineering. Pull twelve months of their actual tool and API spend and add it to their base pay. That total is what the role costs you. It is not the number sitting in your plan. That gap does not show up as a line item. It shows up eight months from now, in a cost center nobody was watching, and gets explained after the fact instead of planned for.
Not every AI dollar is a digital worker
Copilot licenses and developer tooling are software spend, and they should stay software spend. The test is substitution. If the spend funds work that would otherwise be a role, it belongs in the workforce plan. If it makes existing work faster, it belongs in the software budget and shows up in your plan inside a person's fully loaded cost.
Run that test once, deliberately, before departments start building. Companies that skip it end up with an IT budget quietly absorbing labor decisions, or a headcount plan cluttered with seat licenses that were never a hiring tradeoff.
Five ways the 2027 plan breaks
The first four are the ones that have always broken a headcount plan. The fifth is new, and it is growing faster than the other four combined.
1. Drift: roles get opened that nobody approved
A hiring manager has leftover budget and a recruiter has capacity, so a req gets posted with no gate between intent and action. The org finds out when an offer needs signing, or when Finance closes the quarter.
"There's an enormous backlog of headcount that no one had ever accounted for. That's ridiculous."
A Talent leader at an advanced manufacturing companyThe good news: a TeamOhana customer does not run into this blind. Docker stopped six figures of unapproved monthly spend once every request had to map to an approved position, and SeatGeek prevented 200,000 dollars of overspend in a single cycle.
2. The reconciliation tax: Finance rebuilds the forecast by hand
Every month, someone exports the HRIS, exports the ATS, opens the plan, and manually rebuilds the number so all three agree. Mid market Finance and Talent teams burn 60 to 120 hours a month on this. Greenhouse cut that to roughly two hours a week. Metronet saw a 7x return, the equivalent of at least three full time roles reclaimed from reconciliation work.
3. The approval black hole: no trail, no defensibility
A role was approved, probably. The evidence is a memory and a thumbs up in a Slack channel.
"Finance ends up actually manually rebuilding the whole headcount forecast, and you're pointing back to a thumbs up in a Slack channel."
A Finance leader at a life sciences companyThis is a slow cost until it is a sudden one. It shows up hardest when a board asks who approved a position, or when a company heads toward an audit or a financing event. The good news: SeatGeek replaced exactly this kind of operational chaos with one system of record Finance and Talent both trust.
4. Timing blindness: the variance nobody models
The plan is on track in headcount and off track in dollars, because a handful of start dates slipped a quarter and nobody repriced the plan when they did. IonQ identified 594,000 dollars in annual savings once timing was visible in dollars rather than headcount.
5. The token blind spot: AI spend nobody can attribute
This failure mode did not exist three years ago. It is now often the fastest growing number in the company, and it shows up the same six ways almost everywhere we have looked closely.
- Finance owns the AI number by hand, in a spreadsheet reconciled weeks after the money is already gone.
- The heaviest individual spenders are invisible, because their usage sits inside an enterprise plan while tracking only covers the API console.
- Contractor and intern API keys map to nobody, because contingent workers are frequently missing from the headcount system entirely.
- Autonomous agent spend hides on a cloud bill nobody in Finance owns or sees.
- The same dollars get counted twice across two different reporting sources.
- There is no baseline for normal, so a healthy engineering team's heavy usage and one runaway account look identical in a raw total.
Deloitte treats the token as the basic unit of AI cost and value, and recommends real time monitoring and forecasting discipline, the same rigor Finance already applies to headcount. McKinsey goes further: the unit that should be governed is the completed business outcome, not the raw token count. The good news: this is exactly what Token Spend Management is built to close, every AI dollar attributed to a person, team, or agent, inside the same system that holds the headcount plan.
Find out what the token blind spot is costing your team.
Twenty minutes, your own data. See fully loaded cost, comp plus AI spend, on one line.
What planning season actually costs
To size these in your own numbers: count unapproved reqs from last year, tally reconciliation hours across Finance and Talent Ops, time how long it takes to answer who approved a role, compare planned to actual start dates, and compare per-head AI spend to what is currently budgeted. The table below separates the general benchmark from what a specific customer did about it, since those are two different kinds of evidence.
| Failure mode | Industry benchmark | How a customer solved it |
|---|---|---|
| Drift | Unapproved spend commonly reaches six figures a month before a gate exists. | Docker stopped six figures of unapproved monthly spend. SeatGeek prevented $200K of overspend in one cycle. |
| Reconciliation tax | 60 to 120 hours a month at mid market organizations. | Greenhouse cut it to about two hours a week. Metronet saw 7x ROI on the same problem. |
| Approval gap | Two to four week approval cycles are common without a system of record. | SeatGeek ended operational chaos and scaled without adding headcount to manage it. |
| Timing drift | Timing and level drift often outweigh headcount misses as a variance driver. | IonQ identified $594,000 in annual savings once timing was priced in dollars. |
| Token blind spot | AI-native app spend grew 400% year over year at large enterprises. | Token Spend Management attributes every AI dollar to a person, team, or agent, inside the headcount plan. |
The line that should worry Finance most: business units, not IT, now control 81 percent of software spend, against IT's 15 percent. AI is landing on someone's expense report inside a business unit before any central function sees it. By the time it reaches a budget review, it is already a run rate, not a decision.
How do you manage AI token spend alongside your headcount budget?
Annual planning has always been about deciding what to fund. What changed is that funding now splits three ways, and all three draw from the same pool of money.
| Option | What it does | Who usually decides today | Where it is approved today |
|---|---|---|---|
| Fund a hire | Adds a person and, increasingly, a real AI cost attached to that person | Hiring manager, Finance | The headcount plan |
| Add capability | Raises how much an existing person can do, without adding headcount | Whoever owns the tool budget | Usually nowhere, or a procurement form |
| Deploy a digital worker | Produces output with no one at the keyboard, consuming its own budget | Engineering, IT, or nobody | Usually nowhere at all |
That last column is the whole problem. One of the three options runs through an approval workflow. The other two do not. If your planning process cannot price all three against the same budget, your leaders are not making a tradeoff between them. They are guessing, and Finance reconciles the guess after the money is spent.
The data already shows it. Gartner's CFO survey found headcount growth expectations falling from 6 percent to 2 percent between 2025 and 2026, a pivot their analysts describe as structural. KPMG makes the same point from the workforce planning side: the useful classification is not "will AI take this job," it is whether a task is human led, AI assisted, AI automated with oversight, or not appropriate for AI at all. That classification is worth running task by task inside the Build phase of your own sequence, in Part 8, rather than deciding it once at the role level.
TeamOhana Computer: your AI coworker for planning
A chatbot waits to be asked. TeamOhana Computer connects directly to your ATS, HRIS, and finance data, already inside your access controls, and reasons through multi-step planning requests instead of one-shotting them. Ask it a question, drop in your own file, or assign it a task, and it shows its work, asks when it is unsure, and remembers your corrections.
Why a generic AI agent is not the same tool
Most teams have already tried wiring workforce data into a general purpose AI agent. It usually works, for one person. One team built their own agent so capable that, in their own words, only one person could use it, it was too powerful. A raw agent does not know your org chart, your approval chain, or who is allowed to see what, so the choice becomes either lock it away or hand out access nobody can safely audit.
| Generic AI agent + MCP | TeamOhana Computer | |
|---|---|---|
| Connects to your systems | Yes, if you build and maintain the connectors yourself | Yes, already built and normalized |
| Understands your data model | No, sees raw fields from each system separately | Yes, one reconciled layer across ATS, HRIS, and finance |
| Knows who can see what | No, access is usually all or nothing | Yes, permission aware by role, no back door to salaries |
| Stays current | Depends on whoever maintains the sync jobs | Near real time, always |
| Safe to hand to every hiring manager | Usually not, one power user holds all the access | Yes, by design |
Atlassian's own research is a useful check on this: 85 percent of knowledge workers already use AI, but only 29 percent have changed how they actually work. Access to a chat window is not the same as a coworker embedded in the workflow with permission to act. That gap is exactly what a generic agent leaves open and what TeamOhana Computer is built to close.
The trust model
For any action that changes a real record, the agent runs a dry run first and asks for confirmation before anything actually changes. Nothing updates by accident, and every action leaves a record of who, or what, made the change. Teams hand it real work because of that dry run, not only questions.
What Finance, Talent, HR, and executives are already asking it to do
"I need to find $1M in savings within our hiring plan with the least impact to our goals. Model these scenarios and rank them by savings versus business impact: pushing target start dates later, deprioritizing non-revenue-generating roles, leveling down open roles, and moving roles to lower cost regions."
"Generate a dashboard that compares my May 1 headcount forecast snapshot to my July 1 forecast snapshot, and include recommendations."
"Show me all open headcount with a target start date in the next 90 days that don't have a recruiter assigned."
"Which recruiters should I assign to the open headcount based on their capacity and specialization?"
"Generate an org design dashboard I can filter by division that assesses span of control, org layers, and key ratios including AE to Sales Ops and AE to Solutions Engineer."
"Show me our average attrition by department and tenure, and what are the top five causes?"
"Build a workforce health dashboard including average tenure, year to date attrition, headcount growth versus the same period last year, and average comp ratio. Then analyze the hiring plan: expected growth by department, the biggest risks to hitting it, and 3 to 5 concrete recommendations."
See TeamOhana Computer on your own headcount data.
Try the scenario modeling use case first. It turns "find savings" from a week of spreadsheet work into a same day conversation, with the tradeoffs laid out instead of buried.
Introducing the ledger for AI spend
Token data shows what was spent, but only the employee record explains who spent it and why. Most companies hold those two facts in different systems, so AI spend gets reported weeks late instead of planned in advance.
The idea worth taking into 2027 planning, regardless of which tool gets you there: one ledger for both. Every AI source, from Anthropic and OpenAI to Cursor and AWS Bedrock to seat based tools, consolidated with your HRIS and ATS data in one place. Pairing them makes five things possible.
- Every dollar attributed to a person, a team, or an agent, with an owner to contact when something spikes.
- No double counting, so when two sources report the same spend, you keep one clean number instead of two.
- Outliers measured against a team's own baseline, not a raw total, so one person spending ten times their team's average surfaces in days instead of at year end.
- Fully loaded cost per employee: compensation plus AI spend on one line, rolled up by team and by org.
- Forecasting at the point of the headcount request itself, so a planned hire carries expected token spend alongside comp, budgeted before the offer goes out, not discovered after.
Visibility and planning are not the same thing. McKinsey argues the unit that matters is cost per completed business outcome, not cost per model call. A report tells you what already happened. A plan has to price what has not happened yet.
"The big thing is having all token spend in one place, so we can actually analyze what's going on. A new engineer next year adds incremental payroll and incremental AI spend. Budgeting for that is exactly where we're going."
Elisa Folden, Sr. Manager, Business Systems, SeatGeekThe one prerequisite that determines whether any of this works
None of the above works if your contractors and interns are not in your headcount system with a consistent worker type. They are typically the heaviest unattributed AI spenders and the population most likely to be missing entirely, because a sync filter written years ago to exclude consultants often catches fixed term and intern workers by the same rule. You cannot attribute a dollar to someone your systems never loaded. Fixing this is the actual first step, before any ledger or dashboard.
See fully loaded cost on your own numbers.
Comp plus AI spend, one line, before the offer goes out.
The planning sequence
Twelve weeks, six phases, three owners, with an AI checkpoint built into every phase. Twelve weeks is the pace shown here; a smaller company can compress it to six to eight weeks, a larger one may need sixteen. The order and the ownership matter more than the exact calendar. On a calendar fiscal year, a 12 week sequence starting in early September has you board ready by the end of November. On a non-calendar fiscal year, count back twelve weeks from your board date and the sequence holds.
The critical design choice is in the last phase. Most companies stop at approval. This sequence ends at load and lock, where the approved plan, including its AI spend, becomes the operating constraint in the systems where work actually happens.
| Phase | Wks | Finance, Talent, HR own | AI checkpoint | Exit criteria |
|---|---|---|---|---|
| 1. Baseline | 1–2 | Actuals, open reqs, active roster, all reconciled to one starting number | Confirm contractors and interns are in the headcount system with a consistent worker type | One agreed headcount number and one agreed dollar figure, no side spreadsheets |
| 2. Envelope | 3–4 | Total dollar envelope by department and quarter | Attach a per-role AI spend baseline to the envelope, built from your own 3 to 6 month usage data | Departments receive a dollar envelope that already includes expected AI cost, not a body count |
| 3. Build | 5–7 | Cost per role by start date; org design and feasibility | For each requested role, ask whether it could be a digital worker instead, and price that option alongside the hire | Every requested role has a level, a cost including AI spend, a start date, and a stated reason |
| 4. Pressure test | 8–9 | Downside and upside cases; ranked cut and add lists | Model what changes if AI capability is added to existing teams instead of new hires in the downside case | A ranked cut list and add list that include the tools versus hire tradeoff, not just headcount |
| 5. Approve | 10–11 | Board and exec approval of the full dollar plan | Any digital worker deployment is on the same approval as a hire would be, with a named owner and a cost | Every approved position and every approved digital worker has an approver, a timestamp, and a cost of record |
| 6. Load and lock | 12 | Budget loaded; variance thresholds set | AI spend baselines and digital worker budgets are live in the same system as headcount, not a separate tracker | The approved plan, people and AI together, is the operating constraint, not a document people reference |
Pricing a role by level is only as good as the comp data behind it. TeamOhana's benchmark is built from more than 8,600 real, signed offers, seed to pre-IPO, not surveys. Enter a role and level and get the real range back in seconds at benchmark.teamohana.com. Upload your own recent offers and see exactly where you are under market, over market, or an outlier, before your bands are set.
Three rules that make it work
- Give departments dollars, not headcount. A body count invites level inflation. A dollar envelope, inclusive of expected AI spend, forces the requester to make the tradeoff themselves.
- Never let AI spend live in a separate system from the headcount plan. If Finance is reconciling two forecasts by hand, the plan is already out of date the day it is approved.
- Never end at approval. If phase six does not happen, phases one through five were a forecasting exercise, and next spring will prove it.
See it on your own numbers.
Rule two only works if you can actually see fully loaded cost, comp plus AI spend, on one line. That is what Token Spend Management does.
What happens after the plan locks
Mapped against the full lifecycle, the guide so far covers define, budget, approve, and meter. Review, change, and retirement are still missing, and retirement is the one that decides whether an AI budget compounds or ratchets.
A hire has a life after the offer. Onboarding, payroll, a review cycle, a promotion or a transfer, and eventually an exit. Every one of those events has an owner, a form, and a system, and every one of them updates what the person costs you.
A digital worker approved in your 2027 plan usually has none of that. It gets deployed, and then it runs. The only event anyone notices is the invoice, and the invoice arrives after the money is gone. Three things have to happen after approval, and they are the difference between a plan and a document.
Review it against its charter
You defined what the digital worker was funded to do. Review it against that, on the same cadence you review anything else that costs six figures. The question is not how many tokens it consumed. It is whether the work it was funded to do got done, and what it cost to do it.
This is also the honest answer to a measurement problem most companies are stuck on. Counting hours saved does not work, because saved hours get absorbed by other work and never show up in the P&L. What does show up is the budget. A digital worker with a charter and an envelope is measurable the same way a funded role is measurable, and that is the only measurement your CFO can act on.
Change it deliberately
Digital workers grow. Scope widens, a companion tool starts operating on its own, spend climbs past what anyone approved. Every one of those is a change to a funded position, and it should route through the same approval as a comp change or a level change. The alternative is autonomy expanding by drift, which is how a nine month old deployment ends up doing work nobody authorized at a cost nobody approved.
Decommission it and take the money back
Turning an agent off is a keystroke. Recovering its budget into the plan, where it can fund the next worker, human or digital, requires a decision and a system that holds the plan.
This is the step almost everyone will skip in 2027, and it is the one that compounds. A dollar you do not reclaim becomes permanent. Do this for two cycles and your envelope is structurally larger every year for capacity you stopped using, and nobody can point to when it happened.
Where this sits in your stack
Your HRIS records human workers. Vendors are now shipping registries that record agents: identity, credentials, permissions, and an audit trail of what is running. Workday's Agent System of Record became generally available this year and does exactly that.
Registries answer a real question, which is what is running and who has access to what. They do not answer whether it should be running, what it is worth, or what it cost against the alternative you did not choose. Those are planning questions, and they get settled before anything gets registered.
Keep the two straight as you evaluate tools this season. A registry is where a digital worker gets provisioned and deactivated. A plan is where it gets defined, funded, approved, measured, and retired. You need both. Only one of them is where the money gets decided.
The readiness score
Sixteen questions. Check each one only if the answer comes from a system in under five minutes, and Finance, Talent, and HR would all give the same answer. Score the two sections separately, headcount out of 8 and AI out of 8, since almost everyone is strong on one and weak on the other. Run this out loud, together, before you build a single line of the plan. Disagreement about the answers is the most useful thing that happens in the meeting.
Headcount authorization
AI and agent readiness
Answer all sixteen to see where you stand.
Score under 9? Let's talk through where the gap is.
Run the score with Finance, Talent, and HR in the room, then bring us the answers.
What changes when this is one system
| Running on reconciliation | Running on authorization | |
|---|---|---|
| The headcount number | Produced monthly by an analyst from four exports | Held continuously by one system all three teams read |
| The AI spend number | Assembled by hand from vendor portals, weeks after the money left | Consolidated daily, attributed to a person, team, or agent |
| A new hire's true cost | Comp only, AI spend discovered later | Comp plus expected AI spend, one number, before the offer goes out |
| A hire versus digital worker decision | Made informally, in a hallway or a Slack thread | Priced against the same budget, with a named owner either way |
| A slipped start date | Found at quarter close | Reprices the forecast the day it changes |
| An AI spend outlier | Invisible in a raw monthly total | Flagged against the team's own baseline within days |
| Finance in the process | Rebuilds two forecasts by hand and arrives late to both decisions | Sets one envelope covering both, and is present at the moment of decision |
| Audit readiness | A project, for headcount and for AI spend both | A query |
What this looks like in practice
These are outcomes from companies that brought headcount and AI spend into one system. Full stories are linked below.
The AI spend and token attribution examples in this document are drawn from early customer work, current as of this planning season, not the multi-year track record behind the headcount numbers above. That is exactly why the readiness score in Part 11 scores headcount and AI spend separately. Most companies are further along on one than the other, and 2027 is when that gap needs to close.
Next step
Run the readiness score in Part 11 with Finance, Talent, and HR in the same room. If you score under 9 overall, the gap is not in your plan. It is between approval and execution, for people or for AI spend or both, and that is where 2027's variance will come from.
When you want to see this against your own data instead of a template, each of these is a short conversation away.
Planning season is weeks away.
Bring us your readiness score and we will walk your own numbers, twenty minutes, no template.