Skip to main content

AI spend management: how to see and control what your company spends on AI.

A guide to AI spend management: where AI spend hides beyond the invoices, how to see it by team, project and agent, how to cap it, how to report it to finance and how to choose a platform.

Topic
AI cost
Product
Novie Control
Format
Guide
Reading time
11 min
Published
2 October 2026
Updated
2 November 2026
In short

AI spend management is how an organisation sees, attributes and controls what it spends on AI: model APIs, AI features inside software it already licenses, tools bought on team cards, and agents. In practice it gives finance one view of every AI cost, whether it arrives on an API invoice, inside a software licence or on a team card.

Some teams call the same work AI cost management. The phrase AI spend management is also used for software that applies AI to a company's general procurement spend. This guide covers the first meaning: managing the spend on AI itself.

Seeing what your company spends on AI means recording the cost of every model call at the moment it runs, tagged with the team, project and workflow that caused it, so finance reads a ledger instead of waiting for a provider invoice. Most organisations cannot do this today because providers bill at the credential, and one credential is usually shared by many teams. The fix sits in the path the call takes, before the money is spent.

Your total AI spend versus what the invoices show.

A provider invoice tells you how many tokens an API key consumed last month. It does not tell you which team ran the workflow, which project the spend belongs to, or whether anyone approved it. Seeing AI spend properly means attaching those three facts to every call before the money moves.

AI spend also arrives from more directions than the API. A writing tool on a team card, a vendor's AI feature switched on inside a product you already license, an agent running on a contract signed for something else. Each is sensible alone. Together they have no owner and no single number.

TakeawayThe invoice shows tokens per key; seeing spend means attaching team, project and approval to every call before the money moves.

What is the organisation actually spending on AI in total, and why does it differ from the invoices?

The total is every route added together: provider APIs, cloud marketplace charges, AI features inside licensed software, tools on team cards and agents running under other contracts. Provider invoices cover only the first route, billed per key, so their sum is almost always lower than the real figure. Each route sits in a different ledger, which is why nobody holds the total.

Five kinds of AI spend usually sit outside the provider invoice, or inside it with no owner attached.

  • Seats. Chat and coding plans priced per person, often bought by one team on a card, so they reach the ledger as expense claims or software.
  • Shared API keys. One key used by several teams shows on the invoice as one total, so the spend is there and nobody owns it.
  • Tokens already used. Usage is billed after the month closes, so this month's spend appears on next month's invoice.
  • AI inside software you already license. A vendor adds an AI feature or an AI tier at renewal, and the cost is booked as the software it came with.
  • Shadow AI. Tools staff sign up for on their own and pay for by card, which appear on no list of AI spend until someone looks.

Finding the true number starts with listing these routes, which is the first step of an AI spend analysis.

TakeawayInvoices name the key and arrive late, and they miss every route that never touches the API.

Why it matters this budget season.

In 2026, 98 percent of FinOps practitioners manage AI spend, up from 31 percent two years earlier, according to the FinOps Foundation's State of FinOps 2026 report of 1,192 respondents. The same report names AI cost management as the top skill teams want to add. Spend arrived before the ability to attribute it.

FinOps Foundation, State of FinOps 2026.

The visibility gap is measured, too. In its Q2 2026 AI Quarterly Pulse Survey of 204 US C-suite leaders at companies above $1 billion in revenue, KPMG found that 26 percent had full, real-time visibility into what their AI systems cost to operate. Only 36 percent had direct token or usage controls in place. The same companies planned an average of $202 million in AI spend over the following twelve months.

KPMG, AI Quarterly Pulse Survey, Q2 2026.

Spend is also arriving from more places at once. In September 2026, Gartner forecast worldwide AI spending of $2.7 trillion in 2026, up 49.5 percent on 2025, and noted that software vendors are embedding agentic AI inside the products companies already buy. More of the AI bill now arrives inside contracts signed for something else.

Budgets are growing into that gap. Forrester's 2027 Budget Planning Guides, published in July 2026, report that more than 80 percent of leaders expect their overall budget to increase in 2027. A 2027 AI line written without a 2026 baseline is a guess with a signature on it.

Forrester, 2027 Budget Planning Guides, July 2026.

TakeawaySpend and the teams managing it arrived before attribution, and a 2027 AI line without a 2026 baseline is a guess.

How to run an AI spend analysis.

An AI spend analysis is a one-off count of what the organisation spent on AI over a recent period, by route, team and purpose. It gives the baseline that the tracking steps below keep current, and it can be done in a week with records finance already holds.

  1. Pull three months of records: provider invoices, cloud marketplace charges, card statements and the renewals of software that added AI features.
  2. Mark each line with its route and, where the record allows, the team and the purpose.
  3. Total it by route and by team, and put the share nobody can attribute on its own line.
  4. Set the total against the budget, if an AI line exists, and against last quarter.
  5. Name an owner for each route and a date for the next count.

TakeawayThe unattributed share is the finding: it shows how much of the AI bill has no owner today.

Where attribution breaks.

Attribution breaks in three places, in the order they usually surface.

  1. Shared credentials. One organisation API key serves marketing, engineering and operations. The invoice shows a total and hides who caused it.
  2. Agents that call agents. One workflow triggers three models across two providers. The cost of the business outcome is split across invoices that never meet.
  3. Attribution after the fact. Finance reconstructs who spent what from logs, weeks later. By then the month has moved on and the pattern has repeated.

TakeawayAttribution done after the fact cannot recover what a shared key or a chain of agents hid at the call.

How to track AI spend by team, project and agent: five steps.

To see AI spend by team, project and agent, tag every model call with all three before it runs and keep the cost of each call in one ledger.

  1. Inventory the spend paths. List every route AI spend takes into the organisation: provider APIs, cloud marketplaces, seats, SaaS features and cards. Ask engineering, marketing, operations and finance separately; the lists will not match, and the differences are the finding.
  2. One credential per team, or a layer that tags. Either give each team its own key, or route calls through a layer that stamps team, project and workflow on every request. Providers bill the credential, so make the credential mean something.
  3. Record cost as a property of the action. At the moment a model call runs, write down the team, project, workflow, agent, model and cost. This is the ledger. Everything downstream is a query against it.
  4. Put the cap where the call happens. A budget in a spreadsheet cannot stop a call. A cap the runtime can read holds before spend is committed and routes the increase to a named person. Set caps at organisation, team, project and agent level.
  5. Review attribution monthly, before the invoice. The measure of the programme is how quickly someone can answer what team X spent last week, on what. Aim for minutes.

Agents need a line of their own. Give each agent its own identity, a named owner and a budget, and record which workflow started it. The cost of a chain of agents calling agents then rolls up to the piece of work that caused it, whichever key the last call used.

TakeawayInventory the spend paths, make the credential mean something, record cost at the call and put the cap where the call happens.

How to monitor AI spend as it happens.

Monitor AI spend daily from the record of each call, and alert on pace against budget, because a monthly total arrives after the money has gone. Four alerts cover most of what goes wrong.

  • Pace. Spend to date against an even burn of the month's budget, with a projection to month end.
  • Near limit and over limit. A warning at a set share of each team, project or agent budget, and a second at the cap.
  • Sudden change. One key, agent or workflow spending several times its usual daily amount, which is how a retry loop or a switch to a pricier model shows up.
  • Untagged spend. Calls that carry no team or project, sent to the person who owns the key.

Send each alert to a named person who can act on it, and copy the owner of the budget it touches.

TakeawayWatch pace, limits, sudden change and untagged calls every day, and send each alert to someone who can act on it.

Generative AI spend management: seats, usage and agents.

Generative AI is bought in three ways, and each needs its own control. Seats, such as an enterprise chat plan, cost a fixed fee per person, so the control is the seat count and any member limit on usage beyond the plan. Usage, such as a model API, is billed per token, so the control is a budget on each key, project or workspace. Agents combine both and run without a person in the loop, so the control is a budget per agent and per task.

TakeawaySeats are capped by count, usage by budget and agents by budget per task, so one company needs all three.

How to set AI spending limits that hold.

An AI spending limit holds when the system that makes the call checks the budget first and stops or escalates the call at the cap. An alert is useful, and it is read after the money is spent. A limit that holds needs four decisions written down.

  1. The unit. Set limits in dollars, because each model prices tokens differently and the same token cap is a different amount of money on every model.
  2. The level. Organisation, team, project and agent, each inside the one above, so a single agent cannot spend a team's month.
  3. What happens at the cap. Stop the call, move the work to a cheaper model, or hold it until someone approves more.
  4. Who approves an increase. A named person, for a set amount and a set period, with the approval kept on record.

Providers now cap spend inside their own accounts: OpenAI by organisation and project, Anthropic by organisation and workspace, and the seat products by member and group. Azure OpenAI and Amazon Bedrock rely on budgets that alert or trigger automation you build. Each cap holds only inside its own account, so a company using several providers needs one budget above them for each team and agent. Our comparison of AI spending limits shows what each provider lets you cap, and where each stops.

TakeawaySet limits in dollars, nest them from organisation to agent, decide what happens at the cap and name who approves more.

How to report AI spend to finance.

Report AI spend to finance once a month on one page, built from the record of each call so it is ready before the invoices arrive.

  • Total AI spend across every route, with the share the provider invoices cover and the share they miss.
  • Spend by team and by project, each against its budget.
  • Usage already run and not yet invoiced, so the month can be accrued.
  • The unattributed remainder, with the person who owns it.
  • Every cap or budget raised in the month, by how much, and who approved it.
  • Cost per outcome for the most expensive workflows, where the outcome can be counted.

Many AI providers bill in US dollars. If your budgets are set in New Zealand or Australian dollars, convert at a stated rate each month so a currency movement is never read as a change in usage.

Finance also decides how the numbers land on teams. Showback reports what a team spent; chargeback moves that amount into the team's own budget. The board needs a shorter version once a quarter.

TakeawayOne page a month, from the call-level record: totals, budgets, the unattributed remainder and every increase with its approver.

How to choose an AI spend management platform.

Choose an AI spend management platform by two tests: whether it records the team, project and agent at the moment of each call, and whether its budgets hold before spend is committed.

Tools sold for AI cost do different jobs. Some read provider bills after the fact, some carry the calls as a gateway, some trace calls for debugging, and some govern the spend with budgets and approvals. Many companies run more than one. Settle these questions before a demo.

  • Which fields does it record for each call: team, project, agent, person, model and cost?
  • Which routes can it see: provider APIs, cloud marketplaces, seats, AI features in licensed software and card spend?
  • Where is a budget checked, before the call or after it, and what happens at the cap?
  • Who approves an increase, and is the approval recorded and time-limited?
  • Does it work with the models, clouds and providers you already run, without asking you to change them?
  • Can finance read and export the reports without asking engineering?
  • Where does your data stay, and who can see the prompts and responses?

Ask each vendor to name the routes it cannot see, and plan how those reach the same ledger.

TakeawayAttribution at the call and budgets that hold before spend are the two tests; everything else is reporting.

What good looks like.

Three measures tell you whether it is working. The share of AI spend attributed to a team and project, which should approach 100 percent. The time it takes to answer a cost question from a CFO, which should be minutes while the meeting is still running. And the number of overruns caught before the invoice arrives, which should be all of them.

If you have run cloud infrastructure, this is familiar. Cloud cost became governable when tagging moved into the request and limits moved into the platform. AI spend is earlier on the same curve, and the curve is steeper.

TakeawayAttribution close to 100 percent, cost questions answered in minutes, and every overrun caught before the invoice.

How Novie Control does it.

Novie ControlAvailable now

Full cost visibility.

See spend by team, workload and model, before the invoice arrives.

$5,440

spent this cycle.

Sources

  1. FinOps Foundation, State of FinOps 2026
  2. KPMG, AI Quarterly Pulse Survey Q2 2026 (24 June 2026)
  3. Forrester, Use 2027 Budget Optimism To Drive An AI Reset (14 July 2026)
  4. Gartner, Gartner Forecasts Worldwide AI Spending to Grow 49.5% in 2026 (16 September 2026)

Other articles

Frequently asked questions

What is AI spend management?

How an organisation sees, attributes and controls what it spends on AI: model APIs, AI features in licensed software, tools on team cards and agents. It records the cost of each call when it runs, ties it to a team and project, and caps spend before it is committed.

What is our organisation actually spending on AI tools in total, and why does it differ from the invoices?

The total adds every route: provider APIs, cloud marketplaces, seats on chat and coding plans, AI features in licensed software, tools on team cards and agents under other contracts. Provider invoices cover only the APIs, billed per key and after the month ends, so they almost always show less than the real figure.

How do you run an AI spend analysis?

Pull three months of provider invoices, marketplace charges, card statements and software renewals. Mark each line with its route, team and purpose, total it by route and team, and put the unattributed share on its own line. Then name an owner for each route.

How can we audit AI spending by team or project?

Give each team its own credential, or route calls through a layer that tags team and project on every request, then record the cost of each call as it runs. Spend outside the API, such as card and software spend, is matched to teams by hand until it moves onto the same ledger.

Is AI spend management the same as AI cost management?

In practice, yes. Both phrases describe seeing, attributing and capping what an organisation spends on AI models, tools, seats and agents. AI spend management is also used for software that applies AI to general procurement spend, which is a different subject.

What should AI spend management software do?

Record the team, project, agent, model and cost of every AI call as it runs. Cover spend that arrives outside the API, such as seats and AI features in licensed software. Check budgets before spend is committed and route each increase to a named person. And give finance reports it can read without asking engineering.

Build the intelligenceyour organisation compounds.