What does AI really cost to run in a finance or support team?

For a finance or support team, the AI model itself is the cheapest part: reading an invoice or answering a routine customer question costs a few cents, and that price keeps falling. What AI really costs a finance or support team is everything around the model: the people who check what it did and handle the cases it passes on, keeping it working as your systems change, and above all how well it was built. For a team of eight, the model can be a thousand euros a year while checking and upkeep come to forty to sixty times that. You can run the numbers for your own team with our AI cost calculator.

How fast are AI model prices falling?

For the same quality of answer, model prices have been falling faster than almost anything else businesses buy.

• The Stanford AI Index found that the cost of running a model at GPT-3.5's level fell from 20 dollars per million tokens in November 2022 to 7 cents by October 2024, a fall of more than 280 times in about two years (Stanford HAI, 2025).
• Epoch AI measured prices for reaching a fixed level of performance falling between 9 and 900 times a year, depending on the task, and warned that the fastest drops may not last (Epoch AI, 2025).
• Andreessen Horowitz put it more simply: for a model of the same quality, the price falls about ten times a year (a16z, 2024).

A token is roughly three quarters of a word. At September 2026 prices, a fast model such as Claude Haiku 4.5 costs 1 dollar per million tokens read and 5 dollars per million written; the strongest, Claude Opus 5.5, costs 4 and 20 (Anthropic pricing). Reading an invoice, matching it and writing the result takes perhaps twenty thousand tokens across several steps, so about 12 cents on the strongest model and about 3 on a fast one.

Why might our AI bill still go up?

Cheaper tokens do not always mean a smaller bill. Newer models think before they answer, and AI that works through a task on its own (an agent) takes many steps to finish it, so each task uses far more tokens than a chatbot reply did. Ethan Ding described the pattern in 2025: people move to the newest, most capable model as soon as it arrives, and the work they give it grows faster than the price per token falls (Ding, 2025).

For a business, the lesson is to be intentional:

• Match the model to the task. Most routine work runs well on a fast, cheap model. Save the strongest for the hard cases. In our calculator, putting 70 percent of items on a fast model, 25 percent on a mid model and 5 percent on the best one costs about a third of running everything on the top model.
• Design the steps. A well-built flow reads a document once and passes only what is needed to each step. A careless one sends the whole document to the most expensive model five times.
• Measure per task, not per token. The number that matters is what one processed invoice or one resolved conversation costs, all in.

Where does the real cost of AI sit?

For a finance team of eight, a realistic picture looks like this:

• The model: about 1,000 euros a year with the model matched to the task, about 3,000 if everything runs on the top model.
• Checking and exceptions: someone reviews what the AI did, handles what it passes on and improves it. Plan for 15 to 25 percent of the time the AI saves. For a team of eight that is roughly 35,000 to 40,000 euros a year, usually the largest running cost, and it is also what makes people trust the system.
• Upkeep: hosting, monitoring, and changes when a supplier changes a format or a system is upgraded. Typically 8,000 to 25,000 euros a year.
• The build: a one-off, and the part no calculator can price honestly, because it depends on your systems and how the work really runs.

For comparison, Ardent Partners' 2025 benchmark puts the full cost of processing one supplier invoice at 9.40 dollars on average and 2.78 dollars for the best teams. Parseur, which sells invoice automation, attributes most of the gap to removing manual steps across the whole flow rather than speeding up one step (Ardent Partners, 2025, as summarised by Parseur).

Does how it is built matter more than the model?

The same work can be built two ways. Built well: the model matched to each task, exceptions designed in, checks only where they matter, running inside the systems people already use. Built quickly: the top model for everything, every output checked by hand, bolted on beside the systems and patched each time they change.

In our calculator, for a finance team of eight at a fully loaded cost of 55,000 euros a person (salary, taxes and benefits), the first costs about 110,000 euros to run over three years and the second about 290,000. That is a difference of about 60,000 euros a year, every year it runs.

Does AI mean fewer people?

The time AI takes on does not have to mean fewer people. In the owner-operated companies we work with, it usually goes into work that was not getting done: a faster month-end, overdue invoices chased, customers answered the same day, growth handled without hiring. A team of eight that gets back two or three people's worth of time can handle about 1.3 to 1.7 times the work it does today.

Our view

When someone tells you AI is almost free, they are pricing the model and nothing else. When someone tells you it is expensive, they are usually pricing a build that was never designed for your work. The honest answer sits in between, and it depends mostly on how well the thing is built. That is why we start by finding out how the work really runs before we build anything: it decides which model each step needs, where the checks go, and what it will cost to keep running.

Questions people ask next

How much does AI cost per month for a mid-sized business?

For one team, the model is usually tens to a few hundred euros a month. The larger monthly costs are the time spent checking the AI and keeping it running. Use the calculator for your own team.

Is it cheaper to use AI than to hire?

It is usually cheaper than hiring for the same routine volume, but the better comparison is capacity: what the team you have can take on once AI handles the routine part.

Will AI get cheaper?

Yes, for the model. For the same quality, prices have fallen roughly tenfold a year, though the newest, strongest models tend to hold their price. The people, the checking and the build will not, which is why how you build matters more than when you start.

Do we need the most expensive model?

For most routine work, no. Match the model to the task and keep the strongest for the hard cases.

What does it cost to keep AI running once it is built?

For a team of eight, mostly people's time: checking the AI's work and handling what it passes on, often 35,000 to 40,000 euros a year, plus 8,000 to 25,000 for hosting, monitoring and changes. The model is the smallest line.

References

• Stanford HAI, The 2025 AI Index Report (2025), https://hai.stanford.edu/ai-index/2025-ai-index-report
• Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks (March 2025), https://epoch.ai/data-insights/llm-inference-price-trends
• Andreessen Horowitz, Welcome to LLMflation: LLM inference cost is going down fast (2024), https://a16z.com/llmflation-llm-inference-cost/
• Ethan Ding, Tokens are getting more expensive (31 July 2025), https://ethanding.substack.com/p/ai-subscriptions-get-short-squeezed
• Anthropic, Pricing, checked 23 September 2026, https://claude.com/pricing
• Parseur, AI Invoice Processing Benchmarks 2026 (September 2026), summarising Ardent Partners, Accounts Payable Metrics That Matter in 2025, https://parseur.com/blog/ai-invoice-processing-benchmarks