How Much Should a Mid-Sized Business Spend on AI?

Why 'what percentage of revenue' is the wrong question

Finance directors keep asking us for the benchmark. What do comparable companies spend on AI as a share of revenue? We understand the instinct. A benchmark makes a budget easy to defend.

The trouble is that the number is meaningless. A 150-person logistics firm with a painful manual scheduling process and a 150-person design agency with no obvious repetitive work have completely different sensible AI budgets. Any figure averaging them tells you nothing about either.

So we budget the other way round. Start with the specific problems worth solving, cost each properly, and add the smaller standing costs that every business now carries. The total is what it is.

The four buckets an AI budget actually contains

Most first-year AI budgets we see include the first two and miss the last two. That is how a project that was on budget becomes a project that is costing more than expected in month nine.

Illustrative ranges for a 150-person business

Treat these as an example of the arithmetic, not a benchmark. They assume a UK or European services or distribution business, in 2026 prices, with no prior AI build.

• Staff tools. Rolling out a business AI assistant to, say, 60 of the 150 staff who would genuinely use it. Budget the per-seat cost your chosen vendor quotes, then add a small allowance for training. Do not give licences to people who will not use them; check usage after three months and reclaim seats.
• One well-chosen project. A document extraction or classification system integrated with an existing ERP or CRM is often a low-to-mid five-figure build. A forecasting model with data cleaning and integration can sit higher. A multi-step agent with several system integrations is more again. Our software project budgeting guide explains why integration, not the model, drives the number.
• Running costs. A reasonable planning assumption is that year-two running and maintenance cost a meaningful fraction of the original build each year. Model usage can be small or large depending entirely on volume and design.
• Internal time. Allow a subject-matter expert for a day or two a week during a project, and ongoing review time afterwards. At fully loaded salary costs that is real money.

Put together, a business at this size doing one serious project and a sensible tools rollout often ends up somewhere between a modest five-figure and a low six-figure annual commitment. A business doing three builds in parallel spends more and, in our experience, frequently finishes fewer of them.

The running costs that surprise people

Build costs are quoted. Running costs arrive later, in pieces.

• Model usage that grows with volume, especially agent-style systems that call a model many times per task
• Human review time for low-confidence outputs, which does not disappear once the system is live
• Re-evaluation when a vendor updates the underlying model and behaviour changes
• Monitoring and logging storage, particularly where regulation expects retained logs
• Retraining or re-tuning as your data drifts from what the model learned on
• Price changes from AI software vendors, which have not all moved in the buyer's favour

Ask any supplier, including us, for an estimated monthly running cost at your expected volume and at three times that volume. If they cannot give you both, the design has not been thought through.

How to structure the budget so it can change

AI budgets set once a year tend to be wrong by March. Structure matters more than the headline figure.

1. Fund discovery separately and cheaply. A small, fixed budget to test whether a problem is solvable with your real data, before committing the build.
2. Release build money in stages. Tie each tranche to evidence: accuracy on real cases, user adoption, measured time saved.
3. Ring-fence running costs. Put them in operating budgets owned by the team that benefits, so the cost sits beside the saving.
4. Keep a small contingency for opportunities. Something useful will emerge mid-year. Better to have room than to raid another project.

This is how we scope work at SpiderHunts: a short fixed-price discovery, then a build priced against a clear definition of done. It makes the budget conversation with a finance director far shorter.

When to spend less than you planned

Some businesses should budget very little for AI builds this year, and it is worth saying so plainly.

• Your core data lives in spreadsheets and email, and nobody trusts the numbers in your main system
• You cannot name a specific process where time or error rate is measurably hurting you
• The last software project is still not finished
• There is nobody internally with time to own the project

In those cases the best AI investment is often plain software and data work: getting orders, customers and stock into one reliable system. It makes every later AI project cheaper and more likely to work. Our custom software practice does a lot of that preparatory work, and it is not glamorous, but it is where AI returns start.

A budget request that gets approved

The requests that get through a sceptical finance director share a structure: the problem in business terms, the current cost of that problem, the proposed spend split into the four buckets, the measurement plan, and the point at which the project would be stopped. If you want a reference for how to size the return side, see calculating automation ROI before you buy.

Frequently asked questions