For everyday office work, plan on a business subscription to an AI assistant. You pay a fixed monthly fee per person, much like other office software. Pay-per-use, meaning per token, only comes in where AI works on its own inside your processes, and it can be capped. Running a model on your own server rarely pays off for a small business.
The actual amount depends on three things. How many people really use AI every day. Whether AI also works without a person involved, for example when processing orders, invoices or enquiries. And how sensitive the data it handles is. Depending on that, your budget is made up of one, two or all three components.
Two news stories point in opposite directions, but neither is typical for a small business
The Czech business daily Hospodářské noviny ran an interview with the head of the online store GymBeam. The headline alone says that AI agents and tokens cost the company tens of thousands of euros a month. There, AI costs are a real line item in the budget.
Meanwhile, Mistral published the weights of its large model Mistral Large 4, with a trillion parameters. The model can be used free of charge, including commercially, and run on your own infrastructure.
So one story says AI can cost a lot of money every month. The other says you can get a model for free. Both are true, and both can confuse the owner of a smaller business.
But a large online store with a high order volume is a different situation from a company with twenty people. And “free to download” doesn’t mean “free to run”. To make a decision, you need to understand what you are actually paying for with AI.
Distinguish three ways of paying before you start calculating
Each of the three suits different work and carries a different risk.
| Per-user subscription | Pay-per-use (tokens) | Self-hosting the model | |
|---|---|---|---|
| What you pay for | A fixed fee per person per month | The amount of text the AI reads and writes | Server, electricity, administration, backups |
| When it fits | People use AI themselves in their everyday work | AI processes documents or messages automatically | Data must not leave the company and volume is high and steady |
| Main risk | Paid licences nobody uses | An unexpectedly high bill | A large investment sitting idle |
| How to manage the risk | Review once a quarter who uses their licence | Monthly limit and alerts with the provider | First test the same model with a provider |
A token is a chunk of text, roughly part of a word. Providers quote a price per million tokens separately for text the model reads and for text it writes. Writing tends to be more expensive.
Most small businesses start with a subscription. Paying for tokens comes in when you want AI to do work on its own, such as reading incoming invoices. Self-hosting is only the third step, and many companies never need it.
Work out the per-person budget in four steps
The budget doesn’t have to be accurate to the penny. But it does have to reflect how your people actually work.
- Split people into three groups. Who would use AI daily (writing, summaries, spreadsheets, translations), who only occasionally and who not at all. In a company of ten, that might be six, two and two.
- Give subscriptions only to daily users. You’ll find the price of the business plan on the provider’s website. Multiply it by the number of people in the first group. Add licences for occasional users once they ask for one. Providers’ terms usually don’t allow several people to share one account.
- Estimate automated tasks separately. For each task, note how many documents or messages it will process per month and how long they are. Then test on a small sample how many tokens one item actually uses. You’ll find the sample’s usage in the provider’s dashboard. Multiply it by the monthly volume.
- Add a buffer and set a cap. An estimate based on a sample tends to be lower than reality, because real documents vary in length and some get processed twice. Set the cap at a level that won’t give you a shock when you hit it.
The monthly budget per person is then simple to calculate. Add up the subscriptions and the cost of automated tasks and divide by the number of people AI helps. It’s more useful, though, to track the two components separately. Subscriptions change with headcount, tokens with workload.
If you don’t want to deal with token calculations, start with subscriptions only. Add automated tasks once you know which work you want to hand over to AI.
Keep token spending under a cap from day one
Fear of an unexpected bill is justified. Pay-per-use has no natural end. If something goes wrong, you keep paying until someone notices.
Bills most often grow for these reasons:
- The agent is stuck in a loop. It repeats a step because it didn’t get the result it expected, and every attempt costs money.
- The model receives the full context every time. For example, the whole price list or a long email history with every order.
- A large model is used for simple work. Sorting messages or extracting an invoice number can be handled by a smaller, cheaper model.
- A test task was left running. Someone tried it out and forgot to switch it off.
What to set up right at the start:
- a monthly spending limit that stops the service once reached, if the provider offers one (otherwise at least alerts and regular spending checks),
- email alerts when half and most of the limit is reached,
- a separate API key for each task, so you can see what costs how much,
- check usage every day for the first week after launch, then once a week.
A stopped service is annoying. A bill you only find out about at the end of the month is worse.
Your own server rarely pays off for a small business
Published weights mean you don’t pay for a model licence. But you pay for everything else.
A trillion-parameter model won’t run on an ordinary company server. It needs a machine with several powerful GPUs and plenty of memory. On top of the purchase come electricity, cooling, backups, security updates and someone to run it. And the server costs money even when nobody is using it at night or over the weekend.
Smaller open models run on more modest hardware. They may be enough for simpler tasks. For more complex work, however, they usually lag behind the large models offered by providers.
Self-hosting starts to make sense when most of the following apply:
- under your customer contracts or the rules of your industry, data must not leave the company,
- the workload is large and evenly spread throughout the day,
- you have an in-house person who can manage the server,
- the task is clearly defined and a smaller model has been shown to handle it.
If data protection is your main motivation, there is also a middle way. A provider can run an open model for you in a region of your choice, or on a machine dedicated solely to you. Before you decide, find out what rules apply to processing personal and sensitive data in your case, and check with a local specialist.
A sensible approach is to try the same open model with a provider first. If it proves itself and the volume really is large, you’ll have the data to calculate whether your own machine will pay for itself.
After a month of trial use, adjust the budget to reality
No advance estimate can replace a month of real use. Roll out AI on a limited scale and after a month sit down with the figures from the provider’s dashboard.
Checklist after the first month
- Is everyone who got a licence using it? Cancel licences that nobody has used.
- Has any occasional user said they’d need it more often?
- Does token usage match the sample-based estimate? If it’s higher, where does the difference come from?
- Did any task hit its limit? Was it genuine volume or an error?
- Has an automated task taken over work you can name? For example, how many documents nobody has to retype anymore.
- Has any data turned up that shouldn’t be sent to an AI tool?
Then adjust the budget and schedule the next review in three months. Provider prices and the range of models change fairly quickly. What held true at the first calculation may not hold a year later.
Sources
Frequently asked questions
Should I give a business AI subscription to all employees at once?
Probably not. Start with people who have clear work where AI will help, and add others once they ask for a licence themselves. That way you won’t pay for licences nobody uses, and you’ll find out where AI really helps in the company.
Is the free version of an AI assistant enough for a business?
For a first look, yes; for working with company data, probably not. Free versions often have different data-handling terms than business plans and usually lack user management. Read the terms of the specific service and tell people what they must not paste into it.
What if the provider raises prices or changes its terms?
That’s why it pays not to tie all your work to a single model. If your automated tasks are built so the model can be swapped, moving to another provider is a matter of adjustments, not a new project. With per-user subscriptions the change is simpler: just move people to another service.
Is an open model worse than a paid one from the big providers?
It depends on the size of the model and the task. Large open models can come close to paid ones. Small ones that run on cheap hardware tend to be weaker at complex work. A test on your own documents decides, not general comparisons.
How do I know AI is paying off for me?
Compare the monthly costs with the work AI has demonstrably taken over or sped up. It’s easiest to measure for repetitive tasks where you know how many items you process each month. For writing and summaries, ask people whether they would give AI up, and if not, what they use it for.