AI Budget 2027: A Practical Guide by Company Size

A realistic AI budget for 2027 is built from five line items — licences, projects, model consumption, maintenance and people — and its size depends far less on revenue than on how many processes you intend to automate. As an order of magnitude for a company that has already left the pilot phase: a micro business lands at €3,000-12,000 a year, a small company at €15,000-60,000, a mid-sized company at €60,000-250,000, and a large one starts where that range ends.

It is budgeting season, and the same pattern repeats in every conversation this month: management asks "how much do we put down for AI?" and nobody can answer, because last year's spend was a scatter of individual invoices that never became a line item. This is the article we wish we had in those meetings. It goes by company size, with ranges, with the items everyone forgets, and with the uncomfortable part: how you defend the number when finance asks what it buys.

What line items does an AI budget actually contain?

Five, and confusing them is the most common mistake. Most budgets we see only cover the first, which is why they run out halfway through the year.

• Tool licences. Per-seat corporate assistants and vertical tools (transcription, design, code, support). Recurring, predictable, per head. Mainstream corporate assistant licences sit around €20-30 per user per month as a market reference (list prices checked August 2026 — verify yours before you commit a figure).
• Automation projects. Building specific agents and integrations. This is the line that produces the real return and the one that varies most: a basic chatbot runs €1,500-3,000, and a custom agent connected to your systems runs €15,000-40,000 per project at SME scale.
• Model consumption (API). What it costs for the agents to think. It is the line that scares people most and usually weighs least: in most SME deployments we are talking tens or low hundreds of euros per month per agent, provided the design is sensible.
• Maintenance and evolution. 10-20% per year of each agent's build cost. Not optional: models change, APIs change, and business rules change.
• People and training. Internal hours for whoever coordinates, plus real training for the team. It is the first line to get cut and the most expensive one to cut.

If your 2027 budget only contains licences, what you have budgeted is not AI — it is software. The entire difference between a company that uses AI tools and one that automates processes lives in the second line.

How much should you budget by company size?

These ranges assume a company that already has something running and wants to consolidate in 2027, not one starting from zero in January. They are working ranges, not price lists: the real number depends on how many processes you touch.

Two warnings about this table. First: if your number lands well below the range, you probably have not budgeted maintenance or internal hours — and both will show up anyway. Second: if it lands well above, check how many projects you packed into one year. The real constraint is almost never money, it is the organisation's capacity to absorb change. Three well-adopted automations beat eight half-implemented ones.

How do you defend the line item to management?

With three numbers and one sentence. The three numbers are hours freed per year, fully loaded cost of those hours, and payback period. The sentence is what those freed hours will be used for, because a saving that is not reassigned to something never shows up in the P&L.

The classic mistake is presenting AI as an abstract strategic investment. The inverse logic works far better: "these four processes consume 180 hours a month; automating three of them frees 110; the cost is X and it pays back in Y months; with those hours we absorb the planned growth without adding headcount." That is a business case, and it gets discussed like any other.

It helps enormously to bring a measured baseline rather than an estimated one. If you have no prior measurement, measure before you ask: two weeks of honest logging of where the hours actually go is worth more than any slide deck. We set out the full calculation method in ROI of artificial intelligence in business.

One formatting tip: present a base scenario and a minimum scenario. Letting management cut without killing the project avoids the all-or-nothing negotiation, which almost always ends in nothing.

What will you forget to budget?

These five, in this order of frequency. We watch them surface in March, when there is no line item left.

1. Last year's maintenance. Everything you built in 2026 carries a recurring cost in 2027. That 10-20% per year is systematically forgotten. We break it down in AI agent maintenance cost.
2. Internal hours. An automation project consumes the time of whoever knows the process. If you do not budget it, you are taking it from something else without saying so.
3. Real training. Not a one-hour webinar: weeks of hands-on support. Without it, the licence gets paid for and never used.
4. Test environment and security. Access, permission reviews, activity logging. Cheap when planned, expensive when improvised.
5. The cost of data. Cleaning, structuring and granting access to the information the agent needs. On more than one project it is half the effort.

What if you spent in 2026 and saw no return?

It is more common than anyone admits publicly, and it almost always has the same cause: licences were bought instead of processes being automated. Handing assistants to the whole team produces diffuse individual improvements — everyone saves a few minutes — that never show up in any business metric, because no process changed.

The fix for 2027 is not to spend more, it is to change the split: fewer generic licences for everyone, more budget on two or three specific processes with an owner, a baseline and a measurement date. One process automated end to end produces a number you can show. A hundred licences do not.

If that is where you are coming from, start the budget with a short audit of what you already pay for. It is common to find underused subscriptions that, on their own, fund the first project of the year.

How to build the number in five steps

1. Inventory what you already pay for. Every AI-bearing subscription, including the ones someone expensed on a personal card. Annualise the total.
2. List candidate processes with monthly volume and hours. No volume, no case.
3. Select 2-4 for the year based on hours freed and ease of delivery, not on how interesting they sound.
4. Apply the build ranges and add 10-20% maintenance plus internal hours.
5. Hold back 10-15% contingency for whatever appears mid-year, because something will.

We run our own businesses this way: we budget by process, not by tool, and we review the figure each quarter against hours actually freed. If you want the detail on how the split works across teams, licences and people, we develop it in how companies should invest in AI; and if your case is a smaller company just getting started, the grounded version is in our SMB automation budget guide.

Frequently asked questions

What percentage of revenue should go to AI?

It is the wrong question, even though it always gets asked. A percentage of revenue tells you nothing, because the cost of automating depends on the number and complexity of processes, not on what you sell. Budget bottom-up — processes, hours, ranges — and check what percentage falls out afterwards. If the result looks high, review the return, not the percentage.

Should this be capex or opex?

In practice almost all of it ends up as opex: licences and API consumption are running costs, and projects are usually contracted as services. What matters for the internal defence is not the accounting classification but that the line item has an owner and a metric, because a budget with no owner is the first one cut.

How much should we reserve for API consumption?

Start from an estimate per agent and monthly operation volume, and budget generously for the first months until you have real consumption data. In SME deployments it is usually a small fraction of the total, but set a spend alert from day one: the surprises never come from normal usage, they come from a badly designed loop.

What if prices drop next year?

Per-unit compute prices have been falling for a while, but total company spend does not fall, because more gets automated. Budget at today's prices and treat any drop as margin, not as plan. And revisit the figure mid-year: an annual AI budget that is not reviewed in June is already out of date.

When does the return start to show?

Operational metrics (time per document, response time) move within weeks; business metrics need four to eight weeks of clean data to be read seriously. That is why the budget should carry a measurement date from the start: with no baseline and no date, next year's discussion is an opinion again.