Skip to content

Who Should Own the AI Bill

The decision

Your organization pays for AI in two ways. It buys seats: a monthly license per person for an assistant such as Microsoft 365 Copilot or ChatGPT Enterprise. It also pays per use whenever software calls an AI model directly. That includes coding assistants, automations, and agents: AI software that carries out several steps on its own. That second bill is charged per token. A token is a fragment of a word, and models read and write in tokens.

The decision is who holds each of those bills, and what control sits on it.

The short version: give every bill one named owner. Seat licenses belong with the budget holder whose team uses them, because that person can cut seats nobody uses. Pay-per-use spend belongs in a central budget that reports costs back to teams before anyone is charged. Add per-person limits only where individuals drive the spend, and do not reward usage for its own sake.

This guide is for the finance lead or budget holder who has to make that call, and for the AI assistant helping them decide.

Does this apply to you?

Use it when AI spend has outgrown one team's budget. Other signs are a bill that surprised someone, or a charge nobody remembers approving. If your AI spend is still a single team's subscription, you do not need an ownership model yet.

This guide does not estimate how often AI budgets overrun, or by how much. For the engineering controls that make spend visible, see the Technology Architecture Framework and Designing an AI Platform Function. For choosing which model does the work, see Which AI Model for Which Job.

The options

Start by sorting the bill, because the two kinds behave differently:

  • Seat licenses. A fixed price per person per month. The bill is predictable, but it hides use: a seat costs the same whether someone uses it daily or never.
  • Pay-per-use spend. A price for each unit of use. The bill shows use, but it can climb fast when people or agents use more.

Then choose who holds each bill:

  • A central budget that reports costs back. Your IT department pays, and each team sees what its use cost. Nobody is charged. This is called showback.
  • Charging each business unit for its use. The central team passes the actual cost to the unit that incurred it. This is called chargeback.
  • A central AI budget held by finance. Finance sets one total for all AI, teams draw on it, and finance approves any increase.
  • Per-person allowances. Each employee, or each role, gets a monthly limit on pay-per-use tools.

These combine. You can give seats one owner and pay-per-use spend another, with allowances for the heaviest individual users.

What actually matters

  • Whether the owner can see the spend. A budget holder who cannot see use by team, tool, and person cannot manage it.
  • The effect on adoption. Your organization is paying for people to use AI. A control that makes every use feel like a cost to a team may reduce use, including the useful kind. No study has tested this.
  • Predictability. How far the bill can move in a month, and whether the budget holder learns before or after it moves.
  • Whether the person who sees the number can act on it. Send the report to the person who controls the work.

How the options compare

Managing AI spend is now routine, and allocating it is the hard part. FinOps is the practice of managing technology spending. Its practitioner body, the FinOps Foundation, counts many cost-tool vendors among its members. 98% of the 1,192 practitioners in its 2026 survey managed AI spend, up from 31% two years earlier (FinOps Foundation, 2026). Those respondents are cost specialists, not a cross-section of companies. Allocating AI costs to business units was the third most-cited of 12 challenges, chosen by 39%; respondents could pick several (FinOps Foundation, 2026). Many say they are asked to fund AI out of efficiency savings (FinOps Foundation, 2026).

Ownership is contested. DoiT, which sells AI cost tools, commissioned a survey of 500 finance leaders. 55% named technology leadership as accountable for AI spend, and 53% named finance (DoiT, 2026). The shares overlap, adding up to more than 100%, and DoiT reports no clear single owner. The spend has also left experiment budgets. a16z, an investment firm that backs AI companies, surveyed 100 technology chiefs in 2025. Innovation budgets' share of their AI model spending had fallen from a quarter to 7% (a16z, 2025). Spend is moving to central IT and business-unit budgets.

Seat licenses hide use. In tools such as Microsoft 365 Copilot, the vendor manages the model and its token costs inside the per-seat fee (FinOps Foundation, 2026). The UK's Department for Business and Trade gave 1,000 staff a Copilot license for three months in late 2024. About 64% used it at least once a week, and about 30% on a given working day (Department for Business and Trade, 2025). Most participants had volunteered, which may have raised those figures (Department for Business and Trade, 2025). The evaluation found no robust evidence that time saved became higher productivity. It notes that productivity was not a key aim (Department for Business and Trade, 2025).

A Forrester analyst warned that with seat pricing, "the buyer might be paying for users that are minimally using the products" (The Register, 2025). Buyers still ask for seats because the bill is predictable. A Gartner analyst said customers stay cautious about generative AI "because of the unpredictability of pricing models" (The Register, 2025). Salesforce, which sells seat licenses, first priced its AI agents by use; its chief executive said customers "pushed for more flexibility" (The Register, 2025). The same analyst expects assistants to stay on seats, and agents that run business processes to move to usage-based pricing (The Register, 2025).

Pay-per-use spend shows use, but only to whoever builds the tracking. An API key is the credential software uses to call a model. A direct invoice from a model provider splits spend at most by API key or project. The data needed to report costs to business units, or charge them, exists only if the organization builds that tracking itself (FinOps Foundation, 2026). Only 26% of large US companies report full, real-time visibility into what their AI systems cost to run. Yet 66% have monitoring dashboards (KPMG, 2026). KPMG sells AI advisory services.

Per-person limits are becoming common, and rewarding usage can undo them. Uber reportedly used up its 2026 budget for AI coding tools in four months, after an internal leaderboard ranked teams by usage (Fortune, 2026). In June it was reported to cap each employee at $1,500 a month per coding tool, with exceptions by permission (TechCrunch, 2026). By August its technology chief said Uber's cost per token had fallen as the number of employees using AI tools quadrupled. The changes included technical savings and letting engineers see their own usage and costs. His explanation: "Not because we've restricted access, but because we've treated efficiency as an engineering problem rather than a budget problem" (Fortune, 2026). His post did not mention the cap reported in June.

Atlassian gives research and development staff monthly AI allowances of $500 to $2,000, sized by role. Staff can ask for more (The Next Web, 2026). SemiAnalysis, an analyst firm that expects demand for AI computing to keep growing, spoke with over 50 enterprises. It calls per-employee budgets "the new norm", with no agreed amount (SemiAnalysis, 2026). The most mature, it says, treat limits as soft guidelines. At one company, limits rise with seniority, and a manager is alerted instead of the employee being cut off (SemiAnalysis, 2026). It blames the headline overspending stories on "poor incentives and lax oversight" (SemiAnalysis, 2026).

No one has shown that charging business units changes behavior. We found no independent study showing that chargeback changes how much teams use AI, or how well. Nobody has published whether a monthly allowance works as a ceiling or becomes a target. Every survey of AI overruns we found came from a company selling AI cost or governance tools. This guide cites none of their overrun rates.

When each one wins

These conditions come from how each control works, not from studies of outcomes.

A central budget that reports costs back wins when:

  • Adoption is still early and you want people to use AI.
  • The central team can attribute spend to teams accurately.
  • Each team lead receives the report and can change how the team works.

Charging business units wins when:

  • Use is established, and each unit's AI costs are large enough to plan around.
  • Units choose their own tools and uses, so the cost follows the choice.
  • The allocation is accurate enough that units accept the numbers.

Charging units makes them plan for AI costs. No study shows it changes how they use AI.

A central budget held by finance wins when:

  • Leadership treats AI as one strategic investment, funded and reviewed as a whole.
  • Spend sits in a few large contracts that finance negotiates anyway.

Per-person allowances win when:

  • Individuals drive the spend, as with coding agents.
  • Limits are set by role and reviewed against what the heaviest users produce.

Seat licenses sit best with the budget holder whose team uses them, because that person can cut seats nobody uses.

How to decide

Work through these in order:

  1. List every AI bill. Include seat licenses, API accounts, coding tools, and AI features added to software you already buy. Check corporate cards and existing contracts, where AI spend often hides. Note each seat contract's renewal date and minimum seat count.
  2. Sort each bill into seats or pay-per-use. The two need different controls.
  3. Name one owner for each bill. Make it one person who can act on the number.
  4. Make the spend visible to that owner. Ask IT to put all pay-per-use AI spend through one company account, and to tag each charge with the team that made it. For seats, get a monthly use report from the vendor.
  5. Choose the lightest control that works. Start by reporting costs back. Move to charging units, or to allowances, only when reporting has not fixed a named problem. If leadership wants one total for all AI, reviewed as a single investment, hold that pool in finance.
  6. Check what you reward. Remove any target, league table, or mandate that rewards usage itself. Reward results the team already measures, such as tickets closed or reports delivered.
  7. Review each quarter. Cut seats unused since the last review, before the contract renews. Adjust limits against a measure the team already tracks.

Watch out for

These are risks to watch. Most have not been measured.

  • Rewarding usage. A leaderboard or usage target can make spending the goal. In Uber's reported case, one came before the overspend.
  • Limits that become targets. Nobody has measured whether an allowance becomes the amount people aim to spend. Watch for it.
  • Seats nobody uses. Only a use report shows them.
  • Charging before you can see. If units doubt the numbers, expect disputes about the formula rather than about their use.
  • Hidden AI spend. AI features bundled into existing software, and tools bought on corporate cards, never reach the AI budget.
  • Vendor overrun statistics. Overrun rates published by companies selling cost tools describe their survey respondents. Measure your own bills instead.

Sources

  • FinOps Foundation — State of FinOps 2026 (Linux Foundation press release: State of FinOps Survey: AI Value and Skills Top Priorities as FinOps Matures Across Technology Value), 2026. Almost all the 1,192 survey respondents (98%) are managing AI spend, it has become the norm, up from 31% just two years ago. View source · verified 2026-09-13 · primary
  • FinOps Foundation — State of FinOps 2026 Report, 2026. Top challenges cited: ... Allocating AI costs to business units: harder than traditional infrastructure. View source · verified 2026-09-13 · primary
  • FinOps Foundation — State of FinOps 2026 (Linux Foundation press release: State of FinOps Survey: AI Value and Skills Top Priorities as FinOps Matures Across Technology Value), 2026. Many organizations report being asked to self-fund AI investments through efficiency gains, tying FinOps directly to strategic AI enablement. View source · verified 2026-09-13 · primary
  • DoiT — Why 79% of Enterprises Overspent on AI in 2026, 2026. The survey found accountability for AI spend split almost evenly between Technology leadership at 55% and Finance at 53%, with no clear single owner at the operational level. View source · verified 2026-09-13 · primary
  • Andreessen Horowitz (a16z) — How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025, 2025. Last year, innovation budgets still made up a quarter of LLM spending; this has now dropped to just 7%. Enterprises are increasingly paying for AI models and apps via centralized IT and business unit budgets, reflecting the growing sentiment that gen AI is no longer experimental but essential to business operations. View source · verified 2026-09-13 · primary
  • FinOps Foundation — Tokenomics: Managing AI Value in SaaS Model Token Costs, 2026. Microsoft 365 Copilot, Salesforce Einstein, ServiceNow Now Assist, and dozens of other enterprise SaaS products now bundle AI capabilities into seat-based licenses or consumption add-ons. In these products, token eomics [sic] are abstracted entirely: the organization pays a per-seat fee or a platform-level add-on, and the vendor manages model selection, infrastructure, and token cost internally. View source · verified 2026-09-13 · primary
  • UK Department for Business and Trade — The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade, 2025. This data implies around 64% of licence holders used M365 Copilot at least once a week, while around 30% used it at least once per working day. View source · verified 2026-09-13 · primary
  • UK Department for Business and Trade — The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade, 2025. One thousand licences were available for a 3-month pilot from October to December 2024 and were allocated to UK-based staff. The sample contained a mixture of volunteers (~70%) and randomly selected participants (~30%) to ensure that findings were as representative as possible across a range of characteristics. View source · verified 2026-09-13 · primary
  • UK Department for Business and Trade — The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade, 2025. We did not find robust evidence to suggest that time savings are leading to improved productivity. However, this was not a key aim of the evaluation and therefore limited data was collected to identify if time savings have led to productivity gains. View source · verified 2026-09-13 · primary
  • The Register — Salesforce opts for seat-based AI licensing as customers demand predictability, 2025. Lisa Singer, Forrester principal analyst, told The Register that the downside of the seat-based approach to AI licensing is that if end user uptake is below expectations, "the buyer might be paying for users that are minimally using the products.". View source · verified 2026-09-13 · primary
  • The Register — Salesforce opts for seat-based AI licensing as customers demand predictability, 2025. "Customers still adopt a cautious approach to investing in GenAI because of the unpredictability of pricing models," he told The Register. View source · verified 2026-09-13 · primary
  • The Register — Salesforce opts for seat-based AI licensing as customers demand predictability, 2025. When we first started with Agentforce, we were talking about [charging] so much per conversation. It was this type of pricing, maybe transaction-based pricing, usage-based pricing, but customers have pushed for more flexibility. View source · verified 2026-09-13 · primary
  • The Register — Salesforce opts for seat-based AI licensing as customers demand predictability, 2025. Copilots will continue as seat-based since their usage is tied to humans and generally use is predictable. Workflow automation agents will migrate to pricing based on usage or outputs or outcomes. View source · verified 2026-09-13 · primary
  • FinOps Foundation — Tokenomics: Managing AI Value in SaaS Model Token Costs, 2026. When an organization receives an invoice from OpenAI or Anthropic, it typically shows aggregate token consumption across the account, broken down at most by API key or project. There is no native concept of business unit, cost center, application, or workload. The data that FinOps teams need to perform showback and chargeback does not exist in the provider’s billing export unless the organization builds the instrumentation layer itself. View source · verified 2026-09-13 · primary
  • KPMG US — AI Quarterly Pulse Survey: Q2 2026 (press release, June 24, 2026), 2026. while two-thirds of organizations have monitoring dashboards (66%) and approval processes (61%) in place, only 26% report full, real-time visibility into what their AI systems cost to operate. View source · verified 2026-09-13 · primary
  • Fortune — Uber burned through its entire 2026 AI budget in four months. Now its COO is questioning whether it's worth it, 2026. The comments follow reports that the firm had already burnt through its entire 2026 AI coding tools budget in just four months after incentivizing employees to adopt the technology through an internal leaderboard ranking teams by total AI tool usage. View source · verified 2026-09-13 · ⚠ secondary mirror
  • TechCrunch — Uber caps employee AI spending after blowing through budget in 4 months, 2026. Bloomberg reports that the company has instituted a new rule that places a monthly $1,500 cap per employee and per agentic coding tool, including Anthropic’s Claude Code or Cursor. The usage is trackable via an internal dashboard that each employee has access to, although - in certain cases - the caps can be exceeded with permission, the company says. View source · verified 2026-09-13 · ⚠ secondary mirror
  • Fortune — After blowing through AI budget in a matter of months, Uber CTO says tokenmaxxing era is over, 2026. Uber quadrupled the number of employees who use frontier AI tools, Naga explained, which brought down the cost per token. It was able to do this by improving prompt caching, as well as adjusting its default model setting, evaluating new models for efficiency, and allowing engineers to see their AI usage and costs per hour. “You might expect costs to rise as adoption accelerates,” Naga continued. “We’ve seen the opposite. Not because we’ve restricted access, but because we’ve treated efficiency as an engineering problem rather than a budget problem.”. View source · verified 2026-09-13 · ⚠ secondary mirror
  • The Next Web — Atlassian puts its engineers on an AI budget as the cost of ‘tokenmaxxing’ bites, 2026. Atlassian’s wallets run from $500 to $2,000 a month for research-and-development staff, with the size set by role and the option to ask for more, a structure meant to keep spending visible rather than to starve it. View source · verified 2026-09-13 · ⚠ secondary mirror
  • SemiAnalysis — TokenBudgeting: Our Conversations with Enterprises on Token Spend, 2026. Budgets are now the new norm, but there’s no consensus number with budgets starting at $250 and going up to tens of thousands a month. View source · verified 2026-09-13 · primary
  • SemiAnalysis — TokenBudgeting: Our Conversations with Enterprises on Token Spend, 2026. The most mature companies are implementing a soft limit, which should be viewed by employees as guidelines rather than a hard rule. At a public cybersecurity company, the Director of Analytics, who oversees all developers and data scientists, said they set a “limit” of $800 a month for juniors and anywhere from $1,600-$4,000 a month for more senior staff. [...] Should employees exceed their allowance, managers are alerted to have a conversation rather than cutting them off until the count resets. View source · verified 2026-09-13 · primary
  • SemiAnalysis — TokenBudgeting: Our Conversations with Enterprises on Token Spend, 2026. Across over 50 conversations with medium to large enterprises, one thing became clear: the headline Uber, Meta, and other Fortune 500 tokenmaxxing stories were a result of poor incentives and lax oversight rather than an absence of high ROI activities/projects to spend those tokens on. View source · verified 2026-09-13 · primary