ai budget 2027 planning ai investment capex opex

Building an AI department budget for 2027 means turning a list of intentions into five defensible lines: licences and consumption, new projects, maintenance of what is already live, people, and a reserve. For a company of 50 to 200 employees with something already in production, the realistic figure sits between €40,000 and €150,000 a year, and the split that survives contact with reality is roughly 25% licences, 35% projects, 15% maintenance, 15% people and 10% reserve.

The expensive mistake is not getting the total wrong — it is budgeting only for building. Most companies that reach October with a blank line end up asking for a number for "AI projects", then discover in March that API consumption, per-seat licences and the upkeep of last year's builds have already eaten half of it. This guide walks through the five lines, who signs off on each, how the split between capex and opex works, and which grants and tax credits to have in hand before you close the number.

How much should AI weigh in a 2027 budget?

It depends on maturity, but the market context helps calibrate: in 2026 Gartner raised its worldwide IT spending forecast above $6.3 trillion, with double-digit growth driven by AI, and expects AI-linked investment to account for close to half of corporate IT budgets by 2027 (forecasts published and reviewed during 2026). Three in four executives said in 2026 that they would increase their spend on AI tooling.

Those headlines are useful for the board deck, not for your spreadsheet. Translated to a mid-sized European company, these are the three scenarios we work with:

Situation	Indicative annual budget	Goal for the year
Year one, nothing in production	€15,000-40,000	One scoped use case and a measured baseline
Year two, 2-4 processes automated	€40,000-90,000	Extend to adjacent departments, consolidate maintenance
Operations already running on agents	€90,000-200,000	Shared platform, data governance and internal capability

If you are starting from zero, do not anchor on a percentage of the IT budget: anchor on the cost of the processes you intend to attack. The inverse calculation — what doing it by hand costs you today — is far easier to defend in a committee than any sector benchmark, and we develop it in our guide to automation budgets for smaller companies.

What are the five lines and what goes in each?

Five lines, because each behaves differently over the year: two are recurring and predictable, one is discretionary, one grows on its own, and the last exists so the other four do not break.

1. Licences and consumption (≈25%). General-purpose per-seat assistants, tool subscriptions and model API consumption. Enterprise per-seat assistants sit around €20-30 per user per month depending on vendor and commitment (list prices reviewed August 2026); API consumption is variable and must be estimated per use case, never per company.
2. New projects (≈35%). Building agents and integrations. A custom agent for a mid-sized company falls in the €15,000-40,000 range depending on scope; a basic chatbot, €1,500-3,000. This is the discretionary line: it gets cut first and is justified case by case.
3. Maintenance and evolution (≈15%). Between 10% and 20% a year of what you have built, covering API changes in connected systems, model updates, behavioural tuning and fixes. It grows on its own every year you ship something new, and it is the line most people forget.
4. People (≈15%). Team training, allocated hours for the internal owner of the system and, in larger organisations, part of a dedicated hire. Without an internal owner with real hours, the rest of the budget degrades.
5. Reserve (≈10%). Not an accounting cushion: the line that lets you take a mid-year opportunity or absorb a vendor price rise without reopening the budget.

On a €60,000 budget for an 80-person company, that is €15,000 in licences, €21,000 in projects, €9,000 in maintenance, €9,000 in people and €6,000 in reserve. With those numbers on the table, the conversation with management stops being about AI and becomes a conversation about management — which is where things actually get approved.

What counts as capex and what as opex?

The working rule: what you build and keep as an asset tends to be capex; what you consume month by month is opex. Building a custom agent, the integrations and internal development are usually capitalised; per-seat licences, token consumption, hosting and maintenance are operating expense.

This distinction is not a minor accounting detail, because it changes who approves what and how fast. In many companies, opex below a threshold is approved by the department head, while any capex goes through an investment committee on its own calendar. Knowing this before you design the year's plan avoids the classic project that is approved in principle and then sits idle for three months waiting on process.

A reality check: capitalising AI development has genuine nuances — useful life, rapid model obsolescence, spend that mixes build and consumption — and the call belongs to your accountants, not to your technology vendor. What you can do is hand them a breakdown already separated into build, consumption and maintenance. With that, the criteria get set in one meeting instead of three.

Which grants and tax credits should the calculation include?

In Spain there are two routes that meaningfully change the net figure, and both are worth verifying before you close the budget. The figures below are as of August 2026 and programmes change: confirm current terms and deadlines before relying on them.

• Tax credits under Article 35 of the Corporate Income Tax Act. Technological innovation carries a general 12% deduction and R&D 25%, with an enhanced 42% rate applying to spend above the average of the two preceding years. These are compatible with grants, subject to the accumulation rules of each case.
• Kit Digital. The programme remains active in 2026 and its catalogue added a dedicated artificial intelligence category in April 2026, with a voucher of up to €6,000 within a wider scheme reaching €12,000 depending on company segment. It targets small companies and self-employed professionals with tax residence in Spain that are current on their tax obligations.
• Complementary programmes. A separate innovation-voucher scheme aimed at AI and automation projects also appeared in 2026, with amounts of up to roughly €7,000 per company. How it stacks with Kit Digital depends on the terms of each call.

The rule we always give: budget gross and treat any grant as probable income, never as a cost reduction. A plan that only works if the subsidy lands is not a plan. What each line covers is reviewed in AI grants in Spain 2026.

Who signs off on the AI budget?

Ideally three different signatures for three different things — and that split explains why so many AI budgets stall. When one person signs everything, it either gets approved without technical judgement or blocked without business judgement.

• Finance signs the total, the capex/opex split and the disbursement calendar.
• The head of IT or technology signs the platform layer: licences, security, integrations and fit with what already exists.
• The process owner — operations, finance, HR, customer service — signs each individual project and commits to the metric it must move. Without that third name, the project has nobody to answer for it.

In companies above 100 people this triangle usually formalises into an AI committee that meets monthly and decides what enters and leaves the portfolio. In smaller ones it is enough that all three functions sit in the same budget meeting. What does not work is AI being "an IT thing": that is where tools nobody uses come from. The logic for splitting money across people, licences and teams is developed in how companies should invest in AI.

Five mistakes that blow up an AI budget

1. Not budgeting maintenance. Mistake number one. Everything you deploy generates a recurring cost of 10-20% a year; with three accumulated projects, that line no longer fits inside a reserve.
2. Estimating API consumption per company instead of per use case. Consumption follows each process's volume; the same agent can cost €30 or €900 a month depending on how often it runs.
3. Booking savings before they exist. Savings are recognised when they show up as a vacancy you do not backfill or as growth without hiring — not on the day the project is signed.
4. Budgeting tools instead of processes. A list of licences is not a plan; the plan is which process improves, by how much and by when.
5. Leaving training out. Without training hours and an internal owner, licences get paid for and never used — the quietest way to waste a budget.

On the third point, be especially candid in committee: the return does arrive, but it arrives later than the spend, and presenting it any other way leaves you exposed at the mid-year review. The recurring side of the maths is in AI agent maintenance cost, and if what you need is to understand what an external partner will invoice you, we break it down in what an AI consulting budget includes.

Frequently asked questions

Do we actually need an "AI department"?

In most mid-sized companies, no: you need a budget line, an owner with allocated hours and a committee that decides. Creating a department before you have a portfolio that justifies it usually produces a structure that explores rather than delivers. The department comes later, when volume demands it.

How do I estimate API consumption if I have never paid for one?

With a scoped pilot: build the use case, run it for two weeks at real volume and extrapolate. Any estimate made before that measurement is guesswork, and the spread between vendors and models is far too wide to budget blind. That is one more reason to make your first project of the year a small one — besides delivering value, it calibrates your budget.

What if my budget gets approved at a lower figure?

Cut projects, never maintenance or training. One fewer project means one improvement that does not arrive; trimmed maintenance means what already works starts failing, and that destroys the internal trust you need for next year. The correct cutting order is: reserve, new projects, underused licences and only then anything else.

Should we budget annually or quarterly?

Annual total, quarterly allocation. The technology moves fast enough that committing all twelve months in October is unwise, and slowly enough that you cannot work without a reference figure. We review the portfolio every quarter and move money between lines without touching the total.

How do I defend the line if last year produced no clear results?

With a baseline and process metrics, not anecdotes. If nothing was measured last year, that becomes the first deliverable of the new budget: define two or three metrics per process and establish the starting point. A budget that opens by admitting measurement needs fixing is far more credible than one promising a 300% ROI.