What AI adoption actually costs — the bill, item by item

Fifty-two percent of small and mid-sized business owners cite cost as one of the two main barriers to adopting AI (OECD/Eurostat, Dec 2025). The trouble is that most conversations about cost stop at the licence price list — the item that is usually not the largest.

Below is the breakdown into four items, ordered from the most discussed to the most overlooked.

Item 1: licences#

Easiest to calculate, which is why it dominates the conversation.

Things worth looking at beyond the headline rate:

Whether price depends on role. A flat per-person rate is a problem in companies where part of the team uses the tool intensively and part checks it once a day. A model that varies price by scope of use can halve the bill — and it determines whether you can afford to cover the whole company or only the office.

Whether AI is included. This is where price lists diverge. Some include model usage in the subscription, some sell credits, some bill actual consumption. Three different risk profiles as adoption grows.

What the annual discount is. Typically 15–20%. Worth asking, because it is rarely on the price list.

Item 2: rollout#

Organising the data, configuration, migrating from whatever came before.

One question decides this item: does the tool require a developer? A platform where a process is configured by describing it rather than coding it moves this cost from an external invoice to your own team's time. That is usually an order-of-magnitude difference.

A second question: after the rollout, can anyone in the company change a process on their own? If not, the rollout cost never ends — it returns with every change in the business.

Item 3: your own people's time#

Usually the largest item and the one least often written down.

It consists of:

• Learning the tool — a few hours per person, multiplied by headcount.
• Reduced output in the first weeks. Realistically: two to four weeks of working slower than before.
• The time of whoever shepherds the rollout. Someone has to, and it is usually someone whose time is expensive.

At twenty people and four hours of learning each, that is eighty hours. Plus two weeks of reduced output across the team. This item routinely exceeds the annual licence cost.

Practical conclusion: a tool that takes two days to learn is more expensive than one understood in an hour — even if its licence costs half as much.

Item 4: model usage#

An item that does not exist in classic software, which is why it is easy to forget.

Language models bill by usage. The key thing to understand: an agent costs many times more than a chat. A single question is one call. An agent running a six-step task is six calls, each carrying accumulated context.

Three questions for your vendor:

1. How do you bill AI usage? Included, credits, or pay-as-you-go.
2. Can I see consumption broken down by user and process? Without that you will not control the bill once the team genuinely adopts the tool.
3. Can I set a limit? An agent triggered on every record change can generate cost out of all proportion to its value.

How to calculate the return#

Three numbers are enough.

Number one: the baseline. How many hours a week the activity you want to improve consumes today. Ask three people who do the work and take the average. Record it before the rollout — nobody remembers honestly afterwards.

Number two: the state after a month. Same activity, same measurement.

Number three: cost per hour. Gross salary plus overheads, divided by working hours.

The return is (one minus two) × three × 52 weeks, minus the sum of the four cost items.

If the result is negative in year one, that does not necessarily mean a bad decision — items 2 and 3 are one-off, while the saving recurs annually. But it is better to know that in advance than to explain it to the board in November.

The cost nobody mentions: doing nothing#

One last item, this time on the other side of the ledger.

What another year in the current state costs. Hours spent retyping data between systems, phone calls chasing status, hunting for documents, decisions made on stale data.

It is a number easy to ignore because it appears on no invoice. In the companies we talk to, it is often larger than the total cost of adoption.

Summary in one paragraph#

Licences are usually the smallest item. The largest is your own people's time, and the most unpredictable is model usage. If you are comparing offers, compare four items, not one. And measure the baseline before you start anything, because without that number none of the others has a reference point.

The wider context of the whole process is in the guide on AI adoption for companies with no IT department.