Contents
The mandate most CFOs are actually getting Where chargeback lands on the ledger The self-funding loop The loop at two sizes What companies are capping at The hard part: your invoices don’t say who 3 questions finance teams are asking Where CloudZero fits

On October 1, 2026, Accenture changed how it pays for Claude. The firm now funds $29 a month per employee centrally. Anything above that has to be charged to a specific client project or team, what Accenture calls a Work Breakdown Structure code. Internal messages seen by 404 Media, which broke the story on October 15, showed some employees had been running $500 to $1,000 a month on Claude Code. Accenture later told Forbes it hadn’t put a firm limit on anyone’s access, and that’s accurate. It put a price on the access, and handed the bill to whoever was getting the value.

That’s the most instructive AI budgeting story of the year, and not because of the number.

Accenture sells AI transformation to other companies. When a firm like that decides AI spend above a small allowance belongs on the project that consumes it, it’s showing you where the money for AI is going to come from at every other company too. Not from a new line in the budget. From the work.

The mandate most CFOs are actually getting

The instruction arriving on a CFO’s desk in 2026 isn’t “approve an AI budget.” It’s “find one.”

The FinOps Foundation’s State of FinOps 2026 report, drawn from 1,192 practitioners managing more than $83 billion in cloud spend, found that many organizations are being asked to self-fund AI investment out of optimization savings. The same report found 98% of those teams now manage AI spend, up from 31% in 2024, and that AI cost management is the skill 58% of them most want to add.

The amounts aren’t small. KPMG’s Q1 2026 AI Quarterly Pulse Survey put average planned AI spending among the large US organizations it surveyed at $207 million over the next 12 months, nearly double a year earlier, and the Q2 survey held it at $202 million. The board wants the program, nobody wants a new nine-figure line on the P&L, and the only place the money exists is inside spend the company already has: cloud, SaaS, data platforms, and the API bills that started as experiments.

That’s the self-funded AI mandate: the instruction to build an AI program out of money the company already spends, with no new budget line. It has two sources of money, and they’re different in kind.

The first is savings: ongoing reductions in cloud and SaaS spend, not one-time credits, that get redirected rather than returned to the general budget. These are recurring, which is what makes them usable against a recurring AI bill, but they only count once they’re realized. A forecast saving isn’t a budget.

The second is chargeback. Chargeback doesn’t create money. It moves AI consumption off a central line and onto budgets that already exist.

Accenture’s $29 covers routine use. Above $29, the client projects pay. The money was always there; chargeback makes it visible.

Where chargeback lands on the ledger

Your controller will ask this before anyone else does, so settle it early. AI spend charges back to two different places depending on what it’s for.

Inference behind a customer-facing feature, the model calls a product makes to serve its users, is a cost of revenue. It belongs in that product’s cost of goods sold, next to the compute and storage already there, because without it the product doesn’t work. Booking it there is what keeps gross margin honest as AI features scale. Take a product reporting 80% gross margin with its inference parked in central IT; move the inference where it belongs and it can be 65%. The numbers are made up. The reclass isn’t, and the SaaS gross margin piece walks through it.

AI tools people use to do their jobs, coding assistants, chat seats, research agents, are operating expense. They charge back to the cost center of the person using them, the way a laptop or a software seat does. The team that gets the productivity carries the cost.

The rule in one line: inference that serves customers is cost of goods on the product P&L; AI tools that serve employees are operating expense on the cost center. Separate them because they answer different questions. The first tells you whether a product’s unit economics survive AI. The second tells you whether a team’s AI spend is earning its keep. Mixed together, as most AI spend is today, they answer neither.

The self-funding loop

The self-funding loop is a four-step cycle: find realized savings in existing spend, route them into the AI funding line, meter AI usage with allowances and chargeback, and return the gains to the next cycle. It runs on an annual budget cycle, with the metering step running daily.

StepWhat happensWhat it needs
1. FindIdentify realized savings in existing cloud and SaaS spend: idle resources, overprovisioned capacity, unused seats, and commitment discounts left on the tableSpend allocated to teams and products, so waste has an owner
2. RouteBook the freed spend as the AI program’s funding line, with a named owner in financeA recorded decision, so the savings don’t dissolve back into the general budget
3. MeterGive every team, project, or person an AI allowance; charge consumption above it to the work that benefitsPer-person, per-product AI spend data, daily, not at invoice
4. ReturnFeed the savings AI produces (fewer hours, fewer tools, or retired capacity) back into step 1The same allocation data, with AI spend now in it

The loop only holds if step 3 holds. Without metering, step 2 becomes a blank check and the program spends its own funding by the second quarter. Meta found that out. In an April 2026 memo, CTO Andrew Bosworth wrote that “nobody should be using AI tools just for the sake of using them.” Meta had encouraged the usage. It hadn’t priced it.

The loop at two sizes

The arithmetic is linear, so here it is at a mid-size company and at enterprise scale. The figures are illustrative, not a customer’s. Both assume 8% recoverable waste, realized over two quarters.

Mid-sizeEnterprise
Cloud and SaaS spend$40M a year$400M a year
Realized savings at 8% (step 1, booked as the AI line in step 2)$3.2M$32M
Headcount400 engineers, 1,200 others4,000 engineers, 12,000 others
Central allowance: $150 a month per engineer, $30 for everyone else$1.15M a year$11.5M a year
Seed funding for inference behind new AI features before they earn revenue (allocated to those products as cost of goods once live)$2.05M$20.5M
Chargeback ruleSpend above the allowance bills to the product the person is tagged to; above double the allowance needs a manager’s sign-offSame rule, same data, 10 times the rows

Step 4 is where the CFO earns next year’s budget. At month three, the chargeback report shows two products accounting for most of the overage. One has cut its release cycle by a third; the other has a single engineer running agents overnight with nothing shipped.

Finance raises the first product’s allowance and asks the second a question, and both conversations happen inside the quarter. The faster teams show up in next year’s step 1 as fewer contractors, fewer tools, or capacity that can be retired. The other kind shows up by name in the same report.

What companies are capping at

The allowance figures above aren’t pulled from the air. Companies have converged on the structure even though they haven’t converged on the number.

CompanyAllowance or capAbove the line
Accenture$29 per employee per month, centrally fundedA client project or team code
Tesla$200 per employee per week on third-party AI toolsApproval required
Apple$200 per day for some engineers, cut from $300 in October 2026; $100 for many non-engineering teamsExemption from senior leadership
Top-three US aerospace and defense manufacturer$250 per employee per monthBudget increase on request
One of the world’s largest pharmaceutical companies$500 per employee per monthNot reported
Workday, StripeAbout $2,000 per employee per monthNot reported

Reported per-employee AI spending caps and allowances at named companies, as of November 16, 2026. Caps are stated in the units each company uses. Normalized to a month at 20 working days, the range runs from $29 at Accenture to roughly $4,000 for Apple engineers.

Normalized to a common unit, that spread runs nearly 140 to one. There’s no right number yet, and if you run a bank, an insurer, or a manufacturer, the anonymous aerospace and pharma caps are closer to your peers than Apple is. What every one of these companies shares is the same three-part design: a default allowance that covers routine use, a chargeback or approval path above it that attaches the extra spend to a person, a project, and a reason, and visibility fast enough that the conversation happens while the spend is happening.

Accenture’s version is the cleanest because its “approval” is a budget code, not a manager’s signature.

One detail from the SemiAnalysis note is worth keeping. When the aerospace company first set its $250 limit, some of its power users hit it in four days. The cap didn’t stop them. It made them ask, and the asking is the governance.

None of this stalls teams. Apple employees told Bloomberg few people ever reached the old $300 ceiling. SemiAnalysis found most companies let employees request more. What caps do is move the burden of proof from finance (“prove this isn’t waste”) to the person spending (“this is what it’s for”).

The hard part: your invoices don’t say who

Everything above assumes you can see AI spend by person and by product on the day it happens. At enterprise scale, nothing you’re billed today gives you that.

Microsoft 365 Copilot and ChatGPT Enterprise are seats: a fixed price per user whether they use it daily or never. Claude and OpenAI API usage arrives as one invoice per organization; usage by API key exists in the vendor console, but someone has to map every key to a person and a product, and keep the map current. Bedrock and Azure OpenAI land inside the cloud bill, mixed in with everything else that account runs. GPU reservations were committed months ago and show up whether anyone used them or not. Half a dozen vendors, half a dozen formats, none of which names the engineer or the product.

Clearing that takes more than a policy. A company with one AI vendor and a few hundred people can run the loop in a spreadsheet: export the usage report, map keys to teams, reconcile monthly. Spreadsheet chargeback breaks at three points. More than two vendors, because the mapping has to be rebuilt per format. Daily cadence, because a monthly reconcile means the $1,000 month is already over when you see it. And enforcement, because a spreadsheet can report that a cap was crossed; it can’t stop the next request or route the approval.

So the question to ask of any system, including a homegrown one: how long until I can see AI spend by person and product, across every vendor, on the same day? The answer sets the start date of your funding model. Until then, step 3 is a policy document rather than a control.

For the board deck, the whole model fits on one line: AI is funded from realized cloud and SaaS savings, metered per person and per product, with consumption above a default allowance charged to the work that benefits.

3 questions finance teams are asking

Where CloudZero fits

This is the problem CloudZero’s AI Signals is built around: ingesting every AI vendor’s bill alongside the cloud bill, then categorizing AI spend in real time by person, team, and activity across more than 35 kinds of work, with monitors that name whose spend broke its pattern and alert the right person.

Coinbase, Klaviyo, Miro, Nubank, and Rapid7 are among the leaders that trust CloudZero. CloudZero’s VP of Finance, Dan Carducci, described the starting condition most finance teams are in: “AI was the most evasive line in my forecast.”

The guardrails capability, in early access as of November 2026, is the enforcement layer on top of that visibility: allowances, caps, and approval paths that run on the same allocated data the savings came from, so the loop runs in one system instead of across three spreadsheets.

If you’ve been handed the self-funded mandate, the first step isn’t picking a cap. It’s getting cloud and SaaS spend allocated well enough that the savings are real and the chargeback is credible. When you want to see allowances and chargeback running against your own spend data, book a CloudZero demo and bring the budget you were told to find.