Enterprise AI Budget 2027: How to Forecast and Allocate

Most budget lines have a history behind them. The AI line does not: no stable unit price, no single owner, and last year's figures say little about next year's. In a McKinsey survey run in May 2026, 93% of qualified respondents said they had gone over their AI budget.

This comes from the AI Partners podcast episode where Thomas Spitz, CEO of AI Partners, hosted Théo Martin, Ecosystem Lead at Dust, and from building these plans with clients. Our companion piece on controlling agentic AI costs covers money you already spend. This one covers next year's number.

What a 2027 AI budget should cover

Four things. Most plans include the first two and forget the rest.

• Licences per user. A fixed price per person per month, easy to plan, and the part everyone remembers.
• Usage-based model costs. What you pay each time an AI does a piece of work, billed by volume. It moves with adoption, so it belongs in the plan as a range rather than a single figure.
• Running the agents. Someone sets them up, keeps them working and checks what they produce. That time is real and almost never budgeted.
• Training. A budget nobody knows how to use produces no return, which is why upskilling teams belongs inside the AI budget rather than next to it.

One more cost reaches your accounts without appearing anywhere. Your CRM, ERP and collaboration tools now sell AI features in a higher pricing tier, usually with no separate line to show it.

Why AI spend is hard to forecast

The same task does not cost the same twice. McKinsey found that the computing needed for an identical task can vary by up to 30 times.

The reason is in how agents work. With a chat assistant, someone asks a question and gets an answer. When an agent is given a goal, it plans, tries, hits a dead end, tries again, and calls other tools along the way.

Each run takes a different path, and you pay for every step. A licence costs the same whether someone opens the tool monthly or works in it all day. Agent costs do not behave that way, so a budget based on last quarter will be wrong, not just tight.

There is a second trap, and it catches the finance teams who did their research. The price of each unit of AI has fallen sharply, with analysis of Stanford HAI data showing the cost of a fixed level of performance dropped roughly 280-fold in two years. Bills went up anyway, because usage grew faster. Cheaper per unit does not mean cheaper in total.

Who should build the AI budget

Take a company of 1,000 people starting from nothing. The instinct is to ask finance or IT for a number. Both will give you one, and neither can defend it, because neither runs the work that creates the spending.

Build it the way you build headcount. Finance sets the rules and the calendar, then asks each department head what they expect to need and why. In a company of a hundred people this is already the only workable method, and at a thousand there is no alternative.

Many companies have already shifted this way without deciding to. Among those that went over budget, 23.1% moved the spending onto a business-unit budget outside IT. If marketing and customer service already pay part of the bill, they belong in the room when the number is set.

How to build an AI budget without historical data

The instinct is to take last year and add a percentage. There is no last year worth using, so the number has to be built from behaviour instead.

Start from the people who already use agents well. Measure what they consume and what they produce, then work out what happens if the rest of their team reaches that level. Scale it up, and not in a straight line: McKinsey found AI spending rises by close to four times when a company moves from isolated use cases to company-wide use.

Set a ceiling per function on two criteria at once: how much of the work can realistically be handed to an AI, and how ready the team is to do it. One without the other gives the wrong answer. A function with plenty of repetitive work but no experience cannot absorb a large budget yet. An experienced team with little repetitive work does not need one.

Then look at the real spread of usage. On a recent engagement, the ten heaviest users accounted for close to 10% of the budget, a large group in the middle used far less than they were allowed, and around a fifth of the people with access used almost nothing. The average describes none of them.

Be careful projecting from the top of that curve. Heavy users are often enthusiasts testing things with little connection to their job, so their consumption is not what the function needs.

Then sanity-check the result against the market. Gartner expects worldwide AI spending to grow from $2.59 trillion in 2026 to roughly $3.49 trillion in 2027, around 35% in a single year. A number landing far outside that trend needs an explanation, in either direction.

Can AI spend still be reduced after the budget is set?

That depends on decisions taken long before the budget meeting. A number you cannot act on is a forecast, not a budget, and two things decide which one you have.

Whether agents are shared or built individually. When a handful of shared agents carry most of the value, moving one to a cheaper model is a single change. When the same work is spread across dozens of personal agents, it is dozens of changes, and nobody makes them.

Whether you depend on a single model provider. If switching means rebuilding, you have no lever at all, whatever the market does. Switching should be a setting.

This matters most at contract time, when providers offer better pricing in exchange for multi-year commitments. Decide how much flexibility you are willing to trade before you sign, not when prices move.

The risk of setting the AI budget too low

Gartner's reading of the market is direct: companies show limited appetite for using AI to drive disruptive change and prefer tactical initiatives with incremental gains. Caution is the market norm, which is precisely why it deserves examining.

Setting the 2027 ceiling at last year's subscription price, and asking for proof of return before every increase, looks responsible. In practice it slows adoption to the pace of your most cautious department, because proof of return is easy to ask for and slow to produce.

The real question for a leadership team is what the company looks like in five years, not what the line looks like next quarter. Changing how people work is a transformation on the scale of digitisation, and nobody yet knows whether it settles at one, five, ten or twenty percent of payroll.

If a competitor ends up with a sales team twice as effective as yours, the money you saved will be worth less than the business you did not win. Both risks are real, and only one of them appears on an invoice.

What to do once the AI budget is approved

A number set in October will be wrong by February. Not because the estimate was bad, but because usage grows faster than planning cycles turn, which argues for reviewing quarterly rather than annually.

Three habits keep the plan usable. Model a low, expected and high scenario, and plan against the high one. Buy the extra capacity centrally at the start of the year, so a team that hits its limit triggers a reallocation instead of an emergency request. And compare the plan to real consumption per team each quarter, which only works if you can see spending per team, as described in our article on controlling agentic AI costs.

Watch the full episode

Thomas Spitz, CEO of AI Partners, hosts Théo Martin, Ecosystem Lead at Dust, on controlling agentic AI costs, governing a fleet of agents, and staying flexible when model providers change their pricing. The conversation is in French.

Frequently asked questions

How much does enterprise AI cost?
There is no reliable absolute benchmark, because published per-employee figures vary enormously between sectors. Build your own from observed usage, then compare it to what comparable companies spend as a share of their software budget.

Should the AI budget sit with IT or with the business units?
With the business units, with IT or finance providing the measurement. A central team cannot judge the return on work it does not run, and a department cannot manage a budget it cannot see.

How often should an AI budget be reviewed?
Every quarter, against real consumption per team. Yearly is too slow for a cost that follows adoption, and monthly turns planning into reporting.

How do you budget for agents you have not built yet?
Through readiness rather than a list of use cases. Give departments that are not ready a small fixed amount to experiment with, and keep the real budget for those already running something.

Your 2027 budget depends on something few companies have measured: how ready each function actually is to use AI. That assessment is the starting point of our AI First Organization engagement, which turns it into a costed roadmap. Book 30 minutes to talk it through.