The 2027 AI budget: what to spend, where it goes, and how to prove it paid

Most 2027 AI budgets will be set the way 2026 budgets were: last year's number plus a percentage, with the return to be worked out later. That approach is running out of road.

In September, Gartner grouped its top predictions for 2027 and beyond under three themes, and one of them was "cost to value." Its forecast is that by 2029, 60% of organisations deploying AI will run a dedicated function just to map AI's total cost to value or profit. Large enterprises will staff that function. Mid-market firms will not, so they need a method that does the same job on one page.

This article is the follow-on to our 2027 enterprise AI readiness checklist. The checklist told you what to fix before 2027, and ranked measured value first and token economics fourth. This one tells you how much to spend fixing it, and how to prove the money paid.

Questions this guide answers

How are large enterprises deciding their 2027 AI budgets?

Mostly by moving money rather than adding it. In BCG's May 2026 IT Spending Pulse, 66% of firms plan to increase AI spending, while most other categories, including CRM and ERP, face net cuts to fund it. Gartner expects 60% of organisations deploying AI to run a dedicated cost-to-value function by 2029, and the firms reporting the best returns are narrowing use cases and tracking token spend down to team level. See what firms are actually spending.

What line items belong in an enterprise AI budget for 2027?

Five: seats and licences, usage and tokens, integration and data readiness, governance and security, and people and change. Most budgets only show the first, yet integration and data is usually the largest one-off cost. For a first-year programme we would expect it to take 25% to 35% of the total, with about 10% held centrally in reserve. See the five budget lines.

How do you audit current AI spend before 2027 planning?

Start where AI cost already hides: bundled AI features in software renewals, and token usage running without ceilings or alerts. Baseline every candidate workflow before funding it, then track cost per completed task, counting only tasks finished successfully. See the five traps that sink AI budgets. For an outside view, our AI Gateway & Governance Sprint inventories every AI integration in your stack in two weeks.

How do you present an AI budget request to the CFO?

Lead with realised value, not hours saved: ROI equals realised value minus total cost, divided by total cost, using a realisation rate agreed with finance before the project starts. Show all five cost lines and the payback period, and ask for funding in quarterly tranches released against gates rather than a single annual approval. See how to calculate the return.

1. What firms are actually spending

Three numbers set the context.

Spend per employee is rising fast. The Atlanta Fed's survey of executives puts expected 2026 AI spending at roughly $2,000 per employee, up about 50% on 2025. In professional and business services the figure is around $3,470, a 74% rise. The spread between firms in the same sector is wide.

AI is being paid for by cutting elsewhere. BCG's IT Spending Pulse, a May 2026 survey of 423 directors and above at mid-size and large firms, found 66% plan to increase AI spending. Most other categories, including CRM and ERP, face net cuts to fund it.

Returns exist, but the gap is widening. The same survey puts average measured ROI on GenAI and agents at 13.8%. High-maturity firms report about 19%, against 8% to 9% for low-maturity ones. The difference is not the size of the budget. Mature firms know where their tokens go, often at team and user level.

For the mid-market, there is a further warning. McKinsey found that 40% of large organisations are now scaling AI agents, up from 27%, while smaller organisations were flat at 22%. The firms in between are the ones most at risk of spending like a large company and measuring like a small one. The same pattern shows up further down the supply chain. In India, for example, the BCG-FICCI report estimates AI could unlock more than $500 billion in value across 64 million small firms, yet found a clear gap between interest and actual investment. Many of those firms are mid-market companies' suppliers and customers.

2. Three ways to size the number

There is no correct percentage. There are three useful methods, and the answer is where they roughly agree.

Per-employee benchmark. Multiply headcount by a sector benchmark such as the Atlanta Fed figures. Fast, defensible in a board pack, and blind to what you will actually do with the money.

Share of the IT budget. Start from the overall IT budget and decide how much is being redirected to AI, and from where. This forces the reallocation conversation BCG's respondents are already having.

Bottom-up by workflow. List the workflows you will fund in 2027, cost each across the five lines below, and add them up. This is slower, but it is the only method that produces a number you can later measure against.

Our recommendation. Build the budget bottom-up, then check it against the other two. If bottom-up comes in far below the per-employee benchmark, you are probably under-funding integration and data. If it comes in far above, you are probably funding too many workflows at once. BCG found that the stronger performers are narrowing their use cases, not multiplying them.

3. The five budget lines

Most AI budgets only show licences. The full cost of a working AI workflow sits across five lines.

Seats and licences. Assistant subscriptions, copilots and AI features in existing software. Watch renewals closely. Gartner expects much of 2027's AI spend to arrive as features embedded by incumbent vendors, which means AI cost can rise without anyone approving an AI purchase.

Usage and tokens. API calls, inference and agent runs. This is the line that surprises. Agent workloads consume far more than chat, and Gartner has predicted that AI coding costs alone will overtake the average developer's salary by 2028.

Integration and data readiness. Connecting AI to systems of record, cleaning and classifying data, and building retrieval. This is usually the largest one-off cost and the one most often left out.

Governance and security. Logging, agent identity, access controls, evaluation and red-teaming. Gartner now expects insurers, not regulators, to drive AI governance by 2030 through underwriting standards for AI liability cover. Controls you cannot evidence will cost you in premiums.

Our five-test behaviour due diligence shows how to spend part of it before you sign.

People and change. Training, process redesign and the time of the business owner who runs the workflow. Without this line, hours saved stay as hours saved.

Our working view on allocation. For a first-year programme, we would expect roughly 20% to 30% on seats and licences, 15% to 25% on usage, 25% to 35% on integration and data, 10% to 15% on governance and security, and 10% to 15% on people and change, with a reserve of about 10% held centrally. These ranges are our judgement, not a sourced benchmark. In year two the integration line should fall sharply. If it does not, something is being rebuilt that should have been reused.

4. How to calculate the return

One formula, three supporting measures.

ROI = (realised value − total cost) ÷ total cost

Realised value is hours saved × loaded hourly cost × realisation rate, plus any revenue you can attribute, plus any quantified reduction in losses or penalties.

The realisation rate is the part most business cases leave out. An hour saved only becomes money if it is redeployed: more output from the same team, slower hiring, or less outsourcing. It is the reason McKinsey's respondents could report 80% productivity gains alongside only 37% attributing any EBIT impact, the gap behind the first priority in our checklist, redesign work and measure value. Set the rate with your CFO before the project starts, not after.

Total cost is all five lines, including one-off integration, not just the subscription.

Cost per completed task is total run cost divided by tasks completed successfully. Failed or reworked tasks do not count. This is the number Gartner expects large firms to manage in real time by 2028, and it is the best single indicator of whether a workflow is healthy.

Payback is total cost divided by monthly realised value, in months.

A worked example

Illustrative figures. A 600-person lender uses AI to help 40 analysts draft credit memos. Each analyst prepares 30 memos a month at three hours each. With AI, each memo takes 1 hour 45 minutes.

• Hours saved: 1.25 hours × 1,200 memos a month = 1,500 hours a month, or 18,000 a year.
• Gross value at a loaded cost of $60 an hour: $1.08 million.
• Realisation rate agreed with finance: 50%, because half the time goes into higher memo volume and half is absorbed. Realised value: $540,000.

Year-one cost:

Year-one ROI is ($540,000 − $348,500) ÷ $348,500, about 55%. Payback is about eight months, and the all-in cost per completed memo is roughly $24. In year two, with integration largely done, cost falls to around $233,000 and ROI rises to about 130%.

Now change one assumption. If only a quarter of the saved time is redeployed, realised value halves to $270,000 and year-one ROI turns negative, at about −23%. The model, the vendor and the licence price have not changed. The realisation rate decided the outcome. That is why it belongs in the budget conversation, not the post-project review.

5. Five traps that sink AI budgets

Licences as value. Seats deployed is an input. If it is the headline metric, nobody is measuring the output.

AI hidden in renewals. Price rises tied to bundled AI features land in the software line, not the AI line. Review every major renewal for AI clauses before signing.

Usage without ceilings. Agent workflows can multiply token consumption overnight. Every team and every agent needs a spend ceiling and an alert, which is the core test in the token and unit economics priority of our checklist. We covered the governance side in the AI spend governance gap.

Pilots without baselines. If you did not measure the workflow before AI, you cannot prove the return after it. Baseline first, always.

Public-facing AI with no cost controls. Gartner predicts that by 2030, 80% of organisations with public-facing AI will have suffered a cost exhaustion attack, where attackers deliberately run up usage to inflate the bill. Treat unusual token consumption as a security signal, not just a finance one.

6. Fund in tranches, not in one approval

Release the 2027 budget quarterly against gates, not as a single annual approval.

Quarter 1: baseline and one workflow. Fund baselining for every candidate workflow and full delivery of one or two.

Each gate asks three questions. Is cost per completed task being tracked? Is there evidence of realised value at the agreed rate? Are governance controls working and evidenced?

Three outcomes. Scale if all three are yes. Hold for one quarter if value is close but not yet proven. Stop if the baseline was never set or realised value is far below plan.

Gartner expects more than 40% of agentic projects to be cancelled by the end of 2027. A gated budget does not prevent cancellations. It makes them cheap, early and deliberate, instead of expensive and late.

Your 2027 AI budget on one page

If you cannot fill in the owner and the realisation rate for a row, that workflow is not ready to be funded.

Before you fund anything, score your foundations on the 2027 enterprise AI readiness checklist. A low score there means the budget should go to priorities 1 to 3 first. For a fuller self-assessment, take our AI readiness checklist. To test your budget against your board's expectations, request a strategy conversation.

Much of the integration and governance spend in this guide goes into two layers we cover in our 2027 Agent Stack series: the context layer and the control layer. For how to build both on a mid-market budget, see the 2027 agent stack for mid-market enterprises.

Related: DPDP 2027 and enterprise AI: where the ₹250 crore exposure actually sits