How to set your company's AI budget for 2027.

01How much should a company budget for AI.

There is no reliable percentage of revenue or of the IT budget to copy. The right figure comes from your own run-rate, multiplied by the drivers you expect to grow, with a reserve for the high case.

Published figures help with context and set expectations of growth. They are the estimates of the firms that publish them, mostly from large organisations, and none is a benchmark for one company.

• Gartner forecasts worldwide AI spending of about US$2.7 trillion in 2026 and about US$3.6 trillion in 2027. Most of it is infrastructure, such as AI-optimised servers and chips, so the total says little about a single company's bill.
• In the same September 2026 forecast, Gartner expects spending on AI agents and assistants to more than double from 2026 to 2027, and spending on generative AI models to rise by more than 80 percent.
• KPMG's global Q3 2026 AI Pulse survey of 2,131 senior leaders found average planned AI investment of US$210 million over the next 12 months, up from US$186 million in the first quarter.
• Forrester's 2027 budget planning research, published in July 2026, found that more than 80 percent of leaders expect their overall budget to increase over the next 12 months.

Read these as direction. AI spend is growing quickly, and much of the growth is in agents, which are priced by use. A company that budgets for 2027 by adding a percentage to 2026 will usually be wrong in the lines that matter.

TakeawayBuild the number from your own run-rate and drivers, and use analyst forecasts as context only.

02Where to start: what you spend on AI today.

Start with an inventory of every route AI spend takes today, with an owner for each line. Provider invoices cover only part of it, so the first job is finding the rest.

• Model APIs, from each provider's billing and usage pages, by project or key.
• Seat plans for chat and coding tools, from the admin consoles and accounts payable.
• Cloud marketplace charges and AI services billed through your cloud accounts.
• AI features added to software you already license, often switched on at renewal.
• Tools on team cards and in expense claims.
• Agents running in production, with what each costs per task.

Use the average of the last three months as the run-rate, and mark how much of it has no team or project attached. That unattributed share is a finding in its own right. Our AI spend management guide sets out how to run the count in a week.

TakeawayThe budget starts from a three-month run-rate across every route, with an owner on each line.

03How to split the budget: seats, API usage and agents.

Split the AI budget by how AI is bought, because each kind grows for its own reasons and needs its own cap. One line called AI hides the fastest-growing part.

Gartner noted in September 2026 that software vendors are embedding agentic AI inside the products companies already buy. That fourth line is easy to miss, because it is usually booked as software.

TakeawayEach way of buying AI gets its own line and its own cap.

04How to forecast AI usage growth for 2027.

Forecast each line from its drivers, as an expected case and a high case. Fund the expected case, and hold the gap to the high case as a reserve that needs approval to spend.

1. Take each line's monthly run-rate from the inventory.
2. Write down its drivers: users or tasks per month, calls per task, tokens per call and price per model.
3. Project each driver to December 2027, using adoption plans from the teams that own the line.
4. Add new work: each agent planned for production next year, priced from its pilot's cost per task at production volume.
5. Allow for price movement in both directions, and check current rates on the provider's own pricing page before you use them.
6. Set the high case for the drivers you are least sure of, such as agent adoption and calls per task.

Agents need the widest range. Gartner expects AI inference costs per agentic workflow to increase more than fivefold through 2028, even as the price per token falls, because agentic workflows use far more tokens than a simple chat. In June 2026 it also predicted that by 2028 AI coding costs will exceed the average developer's salary, as coding tools move to consumption-based pricing.

The FinOps Foundation adds a note on method: top-down budgets are less accurate for AI, and pilots need flexible funding. Our essay on corporate AI spending limits shows what happened in 2026 when large companies budgeted coding tools like licences.

TakeawayFund the expected case from drivers, and hold the high case as a reserve that needs approval.

05How to set team and agent budgets with caps.

Allocate the organisation ceiling to teams, then set caps for each agent and each person, counted across every provider. Give every cap a default action and a named person who can approve an increase.

1. Set the organisation ceiling for the year, and release it month by month.
2. Allocate it to teams, each with a named budget owner.
3. Cap each production agent, and each person on usage-priced tools.
4. Limit tokens or steps per task, so a loop stops early.
5. Decide what happens at each cap: stop, slow, step down to a cheaper model, or ask for approval.
6. Name one approver per team, and give every increase a reason and an expiry.

The exception path matters as much as the cap. If an increase takes days, people route around the cap with personal cards and shared keys. Aim to answer requests within hours. Our guide to token budgets and spending limits for AI agents covers what should happen at each cap.

Many large companies are already setting these controls. In KPMG's Q3 2026 AI Pulse survey of 314 US leaders at companies with revenue of US$1 billion or more, 43 percent had implemented usage or token budgets and 74 percent included cost reviews in AI approval processes. For how to report team budgets back and when to charge them, see our guide to AI cost budgeting by department.

TakeawayGive every team an owner, and every increase an approver and an expiry.

06How to review the AI budget every month.

Review the AI budget monthly against actual spend by team and by line, and re-forecast quarterly. Usage-priced spend can use a quarter's budget in weeks, so an annual review finds the problem after the money is gone.

• Spend against budget by team and by line, with the pace of burn against an even month.
• The projected year-end figure at the current pace.
• Every increase approved, with its reason and expiry.
• The unattributed share, which should fall each month.
• Cost per outcome for each production agent.
• Any line growing faster than its forecast driver.

Each quarter, re-forecast from the drivers, release or hold the reserve, and move money from lines that missed their outcome to lines that met it. Our article on what a CFO should see about AI spend lists the numbers finance should be able to read on one page.

TakeawayMonthly reviews catch overruns while the money can still be saved, and quarterly re-forecasts move it to what works.

07How to tie the AI budget to outcomes.

Give every budget line an outcome it pays for and a cost per outcome target, and report both together. A line that cannot name its outcome is the first candidate to cut.

Few organisations do this today. KPMG's global Q3 2026 AI Pulse found that 12 percent consistently assess AI value against cost across the organisation. Forrester's 2027 planning guidance makes a similar point, advising leaders to stop funding AI efforts that lack structure, and to agree strategy, governance, decision ownership and shared success metrics before investing more.

• Seats: active use per seat, and the work it replaces or speeds up.
• API usage: cost per request for each use case, against the value of the request.
• Agents: cost per completed task, including human review, against the cost of the task before the agent.

Our ROI guide gives finance a method and a worked example for agents.

TakeawayFund outcomes with a cost per outcome target, and cut the lines that cannot name one.

08What to put in an AI budget worksheet.

Put the inventory, the drivers, the two cases, the allocations with their caps, the outcomes and the review rhythm on one sheet. Copy the worksheet below into a spreadsheet or a planning document and fill in one row per line.

AI budget worksheet for 2027.

Part 1. What we spend today (average of the last three months).
Seats: tool / seats / price per seat / monthly cost / owner.
API usage: provider / project or key / owning team / monthly cost.
Agents in production: agent / owner / tasks per month / cost per task / monthly cost.
AI inside other software: product / AI add-on / renewal date / annual cost.
Unattributed spend: amount / named owner to resolve it.
Total monthly run-rate: ______.

Part 2. Drivers for 2027 (one row per line above).
Users or tasks per month, now and in December 2027: ______.
Calls per task and tokens per call, input and output: ______.
Price per model, checked on (date): ______.
New agents planned for production, with pilot volumes: ______.

Part 3. The budget.
Expected case for the year: ______.
High case for the year: ______.
Reserve (high case minus expected case), released only on approval: ______.
Organisation ceiling: ______.

Part 4. Allocation and caps.
Team / annual allocation / monthly cap / agent caps / person caps / budget owner.
Default at a cap: stop, slow, step down to a cheaper model, or ask.
Exception path: request to (named approver), reason, amount, expiry, answer within (hours).

Part 5. Outcomes.
Line / outcome it pays for / cost per outcome target / value per outcome / reviewed by.

Part 6. Review.
Monthly: spend against budget by team, burn pace, increases approved, unattributed share.
Quarterly: re-forecast, move reserve, retire lines that missed their outcome.

TakeawayOne worksheet holds the run-rate, the drivers, the budget, the caps, the outcomes and the review.

09Common mistakes when setting an AI budget.

• Starting from a market figure. Analyst forecasts describe a market that is mostly infrastructure, so start from your own run-rate.
• Budgeting AI as one line. Seats, API usage and agents grow for different reasons and need different caps.
• Budgeting usage like a licence. Usage-priced tools run out early when adoption rises, so forecast from drivers.
• Caps with no exception path. People route around a cap they cannot raise in time.
• Missing AI inside renewals. AI add-ons arrive as software spend unless someone checks each renewal.
• Reviewing once a year. A monthly review finds an overrun while it can still be stopped.
• Funding lines with no outcome. Every line needs an outcome and a cost per outcome target.

TakeawayStart from your own run-rate, and review it monthly against outcomes.