What Does AI Cost a Small Business? The Cost Types Explained

Key Takeaways

• What AI costs depends mostly on how you use it: licences for an assistant are a very different thing from a solution connected to your systems.
• Plan for six cost types: licences, usage-based fees, project work, training, running and maintenance, and your team's time.
• Estimate each cost type from its own driver, such as the number of users or cases per month.
• Separate one-off from recurring costs and add them up over two to three years.
• A pilot with a fixed budget shows the usage costs in your own business before you commit.

What AI costs an SME depends less on the technology than on how you use it. A few licences for an assistant cost something very different from an AI agent connected to your systems. The costs fall into six types: licences per user, usage-based fees, project and integration work, training, running and maintenance, and your own team's time. You estimate each type from its own driver and add them up over two or three years.

Why is there no single price for AI?

AI isn't one product. Three set-ups are typical for an SME, and each has its own cost profile:

• An AI assistant for the team, as a chat assistant or as a feature in your office software. Here you mostly pay for licences, and costs are easy to plan.
• A ready-made tool with AI for a specific job, such as meeting notes, reading data out of documents or an AI feature in your industry software. Here licences and often usage-based fees come together.
• A custom solution, such as an agent that reads inquiries from your inbox and drafts replies from your own documents. Here one-off project costs come first, followed by usage fees and maintenance. Our page on custom AI agents and assistants shows what such solutions look like.

Sometimes you don't need AI at all. Classic rule-based automation is cheaper and more predictable for some tasks. AI or rule-based automation explains which fits when.

Which cost types do you need to plan for?

1. Licences per user.A monthly or annual fee per person. The driver is the number of people who use the tool regularly, not your total headcount. Watch for minimum seat numbers and annual contracts, and check whether the AI feature is already included in software you have or is billed as an add-on. Business versions with contractual commitments on data use are usually priced differently from consumer versions.
2. Usage-based fees.When a solution reaches an AI model through an interface (API), you pay for the amount processed. Billing is usually in tokens, which are pieces of words, often with input and output priced separately. Longer documents and larger models cost more. Many automation platforms also charge per run. The driver is volume: cases per month times the size of each case.
3. Project and integration work.One-off costs for analysis, design, building, connecting to your systems such as the inbox, file storage or industry software, preparing data and testing. The drivers are the number of systems, the state of your data and the number of exceptions the solution has to handle.
4. Training.The trainer's time, materials and, above all, the working time of the people attending. Add the deeper training for AI champions. The driver is the number of people and roles.
5. Running and maintenance.Ongoing costs for hosting, monitoring, support and adjustments. When a provider changes its model or interface, the solution has to be retested and sometimes adapted. The drivers are how complex the solution is and how critical it is to the business.
6. Internal time.Leadership's time for decisions, and the team's time during the pilot, for checking results, maintaining templates and measuring. These costs appear on no invoice, which is why they're easily forgotten.

How do you estimate the costs before you start?

1. Describe the use case.Write down in one sentence what the solution should do, and match it to one of the three set-ups: assistant, ready-made tool or custom solution.
2. Identify the drivers.For each cost type, note what it depends on: number of users, cases per month, length of documents, number of systems, number of people to train.
3. Get current figures.Licence prices change often. Check them directly with the provider, ask for a quote for the project work and work out usage costs with a test run.
4. Separate one-off and recurring.List one-off and recurring costs separately and add both up over two to three years. That shows you the total cost as well as the entry price.
5. Use a range, not a single figure.Calculate with low, expected and high volume, and add a contingency for the unknown. An honest range is more useful than one number.
6. Set it against the benefit.Compare the costs with the time the solution saves. Our method for measuring time saved by AI shows how, and how to build a business case for AI explains how to present both to leadership.

How do you estimate usage-based costs?

This cost type is the hardest to predict, because it grows with volume. The most reliable approach is a small test run. Take a sample of typical cases, for instance two dozen inquiries or documents, and let the solution process them. Read the usage from the provider's dashboard, divide it by the number of cases and multiply by your monthly volume. Then add a margin for longer cases and for retries when a run fails.

Two levers reduce these costs without putting quality at risk. The first is the choice of model: for simple tasks such as sorting emails, a smaller and cheaper model is often enough, while demanding drafts need a larger one. The second is how much you send: passing on only the relevant sections of a document instead of the whole file costs less. Also set a budget limit with an alert, where the provider offers one, so that a fault in the automation doesn't turn into a surprise on the invoice.

An example: a joinery firm wants email requests for quotes to be read automatically and turned into draft quotes with line items from its own price list. Project costs arise for connecting the inbox and the price list and for testing with past quotes. Usage fees arise for every inquiry, depending on the length of the emails and attachments. The two people who write quotes need training. Running and maintenance cover adjustments when the price list or the model changes. And the owner's time, checking every draft during the pilot, belongs in the calculation too. Faster quotes in the trades shows what such a process looks like step by step.

Which costs are easy to overlook?

• Tidying up data: outdated templates, duplicate folders, price lists in several versions.
• Licences for people who hardly use the tool.
• Price and plan changes by providers, and exchange rates if you're billed in a foreign currency.
• Retesting when a provider introduces a new model or retires an old one.
• Reviewing data protection and data location, especially for client or personal data. Where your data goes explains what to look for.
• The cost of leaving: exporting data, moving templates, switching to another provider.

The invoice shows the licence. Your calendar shows the rest.

How do you keep AI costs under control?

• Start with a pilot that has a fixed budget and a clear end date.
• Give licences only to people who need them, and review usage every quarter.
• Set limits and alerts on usage-based fees.
• Clarify notice periods, price adjustments and what you own in the contract: prompts, templates, code and data.
• Name one person who keeps track of all AI costs, including subscriptions scattered across individual teams.

A reliable cost estimate is part of any good AI strategy and roadmap. If you'd like to go through the costs for a specific use case, get in touch. We'll work out with you which drivers matter in your case and where you need a quote.