Budgeting People, Vendors, and AI Consumption

A fictional online freight marketplace approved a \$17.45 million annual budget for its data and AI platform group. The budget covered payroll for 40 employees and 8 approved hires, 10 contractors, a data warehouse commitment, other software vendors, and a customer-facing assistant that summarizes shipment exceptions. Every figure in this case is invented. At the end of month 6, the group’s vice president is looking at three facts. The 8 hires started three months late, which saved money and capacity at the same time. The assistant’s product team added an agent mode, and the model line is running at nearly twice its plan. And with company revenue behind plan, the CFO asks that full-year spending land 8% below the approved budget, at \$16.05 million. Six months remain.

Those facts come from three different kinds of cost. Payroll moves with decisions made months ago, a warehouse commitment moves at renewal, and a token bill moves with every product release. A budget that treats them as one pool of dollars cannot say which can bend, how fast, or what it costs in lost capacity. The Case Said Four Million and the Cost Was Nine covers how a technology budget is built, forecast, and allocated, and The Bill Doubled covers cloud commitments and unit costs. This case works through the lines those articles leave to one side, then returns to the CFO’s request.

People: payroll is paid in months, plans are made in capacity

A headcount line is a number of months on payroll multiplied by a loaded monthly cost. Loaded cost is salary plus employer-paid benefits and taxes, and whatever else the organization attaches to a person, such as equipment and licenses. For scale, the U.S. Bureau of Labor Statistics reports that in June 2026 private-industry employers spent \$46.89 per hour worked on total compensation, of which \$32.82 was wages and salaries and 30.0% was benefits. Total compensation was therefore about 1.43 times wages (46.89 / 32.82). That is an average across all private industry; a company should use its own figure. The case uses a \$150,000 salary and a 1.43 multiplier: \$214,500 a year, or \$17,875 a month, per employee.

The budget assumed that the 8 approved hires would start at the beginning of month 3 and be on payroll for 10 months each: 8 × 10 × \$17,875 = \$1,430,000. They actually started at the beginning of month 6, for 7 months each, which costs \$1,001,000. The line underspent by \$429,000, and capacity moved the other way. New engineers take time to become productive. Assume, hypothetically, that each delivers half a productive person-month in each of the first three months and a full one afterward. A hire starting in month 3 then provides 3 × 0.5 + 7 = 8.5 productive person-months in the year, or 68 for the eight. A hire starting in month 6 provides 3 × 0.5 + 4 = 5.5, or 44 for the eight. The 24 person-months lost equal the three months of delay for each of eight people, and the group now pays for 56 payroll months to receive 44 productive ones. The saving is real cash, bought with capacity. A roadmap planned on 68 person-months from these hires has 44.

Contractors change the unit of payment. An employee is paid every month, while a contractor is paid for hours billed. Suppose a contractor bills \$140 an hour for 1,800 hours a year, \$252,000 in total, and an employee also supplies 1,800 productive hours after vacation, holidays, and training (both figures hypothetical). The employee costs \$214,500 / 1,800 = \$119.17 per productive hour, so the contractor costs 17.5% more per hour. That premium buys the right to stop paying when the work stops. For one person’s worth of recurring work, the contractor is cheaper only if the need is below \$214,500 / \$140, about 1,532 hours a year, or 85% of the employee’s productive hours. Covering the 24 lost person-months with contractors at 150 hours a month costs 24 × 150 × \$140 = \$504,000. That is \$75,000 more than the \$429,000 saved, which is the 17.5% premium applied to the same work. The hiring slip saves money only if the work waits.

The cost per hour leaves out several things a budget owner has to price separately. A contractor’s knowledge leaves with the contract, and an employee usually has to direct the work, so contractors consume management capacity. The distinction is also a legal one. The IRS weighs behavioral control, financial control, and the type of relationship, which means a contractor who is managed like an employee for years creates a classification risk and not a flexible budget line. Finally, when finance caps headcount but not contractor spending, contractors become the release valve. The budget should then show both in the same capacity unit so that nobody mistakes a shifted cost for a saved one.

Fixed, stepped, and usage-based costs

The useful classification of a line is who can change it and how soon, and its label matters less. Some costs are fixed for a period: payroll for people already hired, or the floor of a commitment. Some are stepped and change in blocks, often only at a contract date: software seats, capacity units, a contractor team. Some are usage-based and change with every request: tokens, storage, ingested log volume. The same budget holds all three.

Line	Shape	What moves it	How soon it can change
Employees (48 by month 6)	Fixed per month	Hiring, attrition, layoffs	Months to fill a seat; layoffs add severance and lost capacity
Contractors (10)	Stepped by person	Platform demand and roadmap	30 days’ notice (assumed in the contracts)
Warehouse commitment (\$3.0M a year)	Floor, with overage above it	Query and storage volume	Floor changes only at renewal; overage changes with usage
Observability (priced by data ingested)	Usage-based	Log volume and retention settings	As soon as engineering changes what is logged
BI seats	Stepped	Number of licensed users	Increases anytime; decreases only at the anniversary
AI model usage (priced by tokens)	Usage-based	Tasks, calls per task, context size, model choice	The next release or limit change

The warehouse shows how a floor changes a range. The company committed to \$3.0 million of usage a year at committed rates, and any usage above that is billed at 1.2 times the committed rate (a hypothetical term). If usage valued at committed rates comes to \$2.6 million, the bill is still \$3.0 million, because the floor binds. If it comes to \$3.3 million, the bill is \$3.0 million plus 1.2 × \$0.3 million, or \$3.36 million. Spending cannot fall below the floor but can rise above it, so the line should be budgeted at the floor and carried with an overage range, not at an expected value. The commitment discount entry works through the break-even arithmetic when usage falls short of the commitment, and the earlier FinOps article shows how coverage and utilization are tracked.

Seat-based tools follow the same pattern by date. A contract may let the count rise during the term and fall only at the anniversary, a true-up without a true-down. A reduction in staff in March then reduces nothing until the anniversary, however the budget was written. The Supplier Changed Its Roadmap and Ours Changed With It covers what to look for in those terms; for a budget, the point is to record the date at which each stepped line can change, because that date decides how much of a mid-year cut can arrive in the same year.

AI consumption: a product decision billed by the token

Model providers meter usage in tokens, the units of text that a model reads and writes (the token entry explains why a word is not a fixed number of them). Input and output tokens carry different prices. In Anthropic’s published price list, output tokens cost five times as much as input tokens for every model listed, tokens read from the prompt cache cost a fraction of the base input price, and batch processing is discounted by 50%. Tool use adds tokens too, because tool definitions and tool results are counted as input. The case uses hypothetical prices of \$3 per million input tokens and \$15 per million output tokens. Check actual prices when budgeting; they change.

The cost of a task is calls × (input tokens × input price + output tokens × output price). The original design made one call per task with 6,000 input tokens and 800 output tokens: \$0.018 + \$0.012 = \$0.030 per task. At 2.0 million tasks a month that is \$60,000 a month, or the \$720,000 in the budget. Agent mode makes three calls per task. Each call carries 9,000 input tokens, because the earlier steps and tool results are sent again, and produces 700 output tokens. That is \$0.027 + \$0.0105 = \$0.0375 per call and \$0.1125 per task, 3.75 times the original. If 60% of the input tokens are read from cache at 10% of the input price, the effective input price falls to \$1.38 per million tokens, and the task costs 3 × (\$0.01242 + \$0.0105) = \$0.06876.

Design	Calls per task	Input tokens per call	Output tokens per call	Cost per task
Original, one call	1	6,000	800	\$0.0300
Agent mode, no caching	3	9,000	700	\$0.1125
Agent mode, 60% of input from cache	3	9,000	700	\$0.0688

None of this needed a single additional user. The line grew because of a product design decision, and the platform group’s budget absorbed it. That is the central difference from payroll: the person who changes the cost is usually not the person who owns the budget. A budget for AI usage therefore needs a named consumption owner, a unit that connects spending to outcomes, and a limit or review trigger agreed before launch. Cost per request misleads because retries, agent loops, and human review multiply it; cost per successful task and Controlling AI Cost per Successful Task show how to build a unit that includes them. The FinOps Foundation’s guidance on AI adds that predictability is generally lower, especially early, and that product, sales, and leadership groups contribute to AI spending directly, so showback and usage limits matter more than they do for infrastructure.

Some providers sell dedicated capacity in place of per-token billing. Microsoft Foundry’s provisioned throughput, for example, holds a fixed amount of capacity for a deployment and bills per capacity unit per hour regardless of tokens used, with reservations at a lower rate and optional spillover to pay-per-token deployments when the dedicated capacity is full. That turns a usage line into a fixed one. Capacity has to be sized for peak load, so if peak traffic is 2.5 times the average, average utilization is at most 1 / 2.5 = 40%. On cost alone, the fixed fee must then be below 40% of the on-demand value of the capacity at full use to win. The other benefits, such as predictable latency, may justify paying more, but that is a service-level decision. The provider’s own guidance points provisioned capacity at predictable traffic, and this assistant’s mix changed within a quarter.

Forecast ranges that finance can use

The FinOps Foundation distinguishes a forecast, which is a promise by engineering and product owners to deliver value for an amount of spending, from a budget, which is the organization’s commitment to fund them. The two are reconciled, not merged, and the useful forecast is built from drivers: tasks per month, the share of tasks in agent mode, cost per task in each mode, warehouse usage against the floor, and the contractor roster. Dollars are the result of those drivers and not the thing forecast.

The first six months are actual spending of \$8,280 thousand, so ranges apply to the second half. The three scenarios are coherent stories, not independent extremes. Low: 2.2 million tasks a month, with 20% in agent mode. Base: 2.6 million tasks, 35% in agent mode. High: 3.2 million tasks, 60% in agent mode, and warehouse usage valued at \$3.3 million. The first-half AI spending of \$486,900 comes from 3 months at the planned \$60,000 and 3 months at 2.2 million tasks with 20% in agent mode.

\$ thousands	Low	Base	High
First half, actual	8,280	8,280	8,280
Second half: employees	5,148	5,148	5,148
Second half: contractors	1,260	1,260	1,260
Second half: warehouse	1,500	1,500	1,860
Second half: other vendors	600	600	600
Second half: AI model usage	614	918	1,526
Full year	17,402	17,706	18,674
Against the \$17,450 budget	−48	+256	+1,224

Two things stand out. First, the people lines have no range, because the months are already committed. The entire spread of \$1,272 thousand between low and high comes from AI usage (\$912.6 thousand) and the warehouse overage (\$360 thousand), and AI usage was only 4.1% of the approved budget. Second, the base case is \$256 thousand over budget even after the \$429 thousand hiring saving: the AI line alone is \$685 thousand above its \$720 thousand plan. Quoting the ranges as independent percentiles would be wrong, because the lines are correlated (the growth that raises token spending also raises warehouse queries), and independent percentile values do not add up to the total’s percentile, as the quantile forecasts entry shows.

A forecast needs a review rhythm and measured error. The FinOps Foundation suggests that organizations define acceptable variance between budget and actual spending, with example thresholds of 20%, 15%, and 12% as practice matures. The AI line missed its plan by 95% (1,405 / 720 = 1.95), which no threshold of that kind would accept, and the post-mortem belongs in the driver, not the dollar: nobody forecast agent mode. Track the error and the bias for each line (a line that is always short is a method problem), and agree on a trigger in advance. For example, if the monthly AI run-rate exceeds the high-case path two months in a row, the forecast is reopened and the consumption owner is asked for a decision.

Working with finance

Finance is the partner that decides what counts as a budget change, which measure a target uses, and how an overrun in one line is covered by an underrun in another. A platform leader helps by bringing drivers and dates and not just totals: when the hires slipped, finance should have heard it as a capacity change with a \$429 thousand cash effect and a \$75 thousand cost of covering it with contractors, rather than as a favorable variance. Ask finance for three things up front: whether the target is stated in cash, reported expense, or headcount; whether a saving in one line can be spent in another, and by whom; and when the next planning calendar closes.

The measure matters because the same month of an engineer can appear as an expense or an asset. Under U.S. GAAP guidance for internal-use software (ASC 350-40), employee payroll and benefits for people working directly on a software project during the application development stage can be capitalized, while training, data conversion, and general overhead are expensed. EisnerAmper’s summary describes the stages. The cash paid is the same either way, and the OpEx entry explains how expense recognition and payment timing differ. A cut stated in reported expense can therefore be partly met by reclassification that saves no cash, and the classification must follow what the engineers actually did, not what meets the target. The decision belongs to finance and its auditors.

Allocation is the other conversation. If the platform group carries the assistant’s model bill centrally, every product decision that raises tokens lands on its budget. Showback reports the cost to the product team without moving money, and chargeback posts it to the product team’s budget. Starting with showback and moving to chargeback once the product owner has the authority and the unit cost to act is a reasonable sequence, and the choice of date belongs to finance.

When the target drops: protected commitments and the arithmetic of a late cut

The CFO’s target of \$16,054 thousand (92% of \$17,450 thousand) sounds like an 8% reduction. Measured against the base forecast of \$17,706 thousand, the required cut is \$1,652 thousand, 9.3% of the full year. All of it must come from the second half, which is forecast at \$9,426 thousand, so it is 17.5% of what remains to be spent. About 70% of the second half is hard to move: employee payroll (\$5,148 thousand) and the warehouse floor (\$1,500 thousand) total \$6,648 thousand, and neither can fall quickly without layoffs or a contract breach. That leaves \$2,778 thousand in contractors, AI usage, and other vendors, and the request would take 59% of it.

Before negotiating, the vice president lists the commitments that should not be cut and why. A commitment is protected when cutting it creates a larger cost later, breaks an obligation, or removes a control that others rely on. In the case, those are the warehouse floor, the on-call and security rotations within the employee payroll, and 4 of the 10 contractors who are working through a security remediation backlog and producing audit evidence. Deferring such work does not remove it: the backlog grows, and the cost returns with interest or as an incident. Protection is a decision with an owner and a review date, and a new risk or a new CFO target can reopen it.

The rest is a ladder of actions with different lead times, and each is valued only for the months it is in force. Notice periods and build time matter as much as the rate. Ending a contractor with 30 days’ notice given at the start of month 7 saves from month 8, five months, not six.

Action	Effect begins	Second-half saving	Cost to the group
Do not backfill the 2.4 departures expected in the second half (10% a year of 48 people), each vacant about 3 months	As people leave	\$128,700 (2.4 × 3 × \$17,875)	7.2 person-months of capacity
End 4 non-protected contractors with 30 days’ notice	Month 8	\$420,000 (4 × 5 × \$21,000)	20 person-months
Prompt caching for agent mode (two months to build, live in month 9)	Month 9	\$159,214 (base case: 910,000 agent tasks a month × \$0.04374 saved × 4 months)	Engineering time; savings depend on traffic
Reduce ingested log volume by 20%	Month 9	\$40,000 (\$300,000 ÷ 6 × 4 × 20%)	Less data for debugging
Total of the four		\$747,914	27.2 person-months

The four actions leave a gap of \$904 thousand against the \$1,652 thousand needed, even though they take every lever that does not touch a protected commitment. One-time costs, such as handover work, reduce the first-year saving and are not in the table. Ending the two remaining non-protected contractors adds 2 × 5 × \$21,000 = \$210,000 and 10 person-months, so the total reaches \$957,914 and 37.2 person-months. The full-year forecast becomes \$16,748 thousand, which is 4.0% below the budget and \$694 thousand above the CFO’s target. The lost capacity is 30 contractor person-months plus 7.2 from vacancies, 37.2 of about 348 person-months in the second half (10.7%).

The remaining \$694 thousand has no further painless source. It can come from protected work or from layoffs, which carry severance and lose capacity in the quarters that follow. It can come from limiting how much traffic uses agent mode, which is a product decision with revenue consequences. Or the company can accept a smaller reduction. The vice president’s job is to make that choice visible with its effects: which roadmap items move, and what delay costs the business (the cost of delay entry shows how to price it). Money already spent on a migration in flight is sunk and should not decide whether to finish it; the remaining cost, the delay, and the risk should.

In the case, the CFO accepts a revised target of \$16.75 million for the year, 4.0% below the original budget. The AI model line becomes a separate line with a named owner in the product group, showback begins the next month, and the trigger from the forecast section is written into the budget. The two additional contractors leave in month 8. The next budget carries the hires at their actual start dates, the warehouse at its floor with an overage range, and the AI line as drivers with a range.

What a reviewer should ask

• Is each line classified as fixed, stepped, or usage-based, with the date it can first change?
• Does the people plan state capacity (productive person-months, including ramp-up) as well as cost, and does it show contractors in the same unit?
• For every commitment, what is the floor, what happens above it, and when can it be reduced?
• Which drivers move the AI line, who owns each one, and which unit of outcome connects spending to value?
• Are scenarios built from coherent stories of the drivers, and is forecast error tracked by line?
• Does the target use cash, reported expense, or headcount, and who at finance decided that?
• Which commitments were protected before the cut, who owns that decision, and what would reopen it?

References

• U.S. Bureau of Labor Statistics, Employer Costs for Employee Compensation, June 2026
• IRS, Independent contractor (self-employed) or employee?
• EisnerAmper, Capitalizing internal-use software
• FinOps Foundation, Forecasting
• FinOps Foundation, Budgeting
• FinOps Foundation, FinOps for AI Overview
• Anthropic, Claude API pricing
• Microsoft Learn, Provisioned throughput for Foundry Models