AI agent cost has 3 parts: the vendor's fee, the model usage behind it, and your own team's time to set the agent up and handle its exceptions. Fees are charged per process, per seat or per use. For back-office work the token bill is usually the smallest of the 3, and the business case rests on hours removed, counted in FTE.
At Lleverage we think every AI agent price should be judged against one number your finance team already trusts: the full cost of one employee. A fixed monthly fee that costs about as much as 1 person is a good deal when the agent removes clearly more than 1 person's hours. Otherwise it is a poor one. Token prices, seat counts and credit bundles are details inside that comparison.
We build AI agents for the back office of companies that make, move and sell physical products. Buyers ask us 2 money questions most: whether a fixed fee that costs about 1 employee pays for itself, and whether token costs run out of control. This article answers both, compares the common pricing models and sets out the FTE method we use. If you want that method applied to one of your own processes, book a demo.
What does an AI agent cost a mid-sized company?
An AI agent costs a mid-sized company a recurring fee to the vendor, a usage bill for the language model, a one-off setup and the hours its own people spend reviewing exceptions. The fee is the visible number. The setup and the review hours decide whether the agent earns its fee back.
It helps to list the 4 lines separately, because quotes from vendors rarely show all of them:
- The fee. What the vendor charges to run the agent: per process, per seat, per document or as a usage bundle.
- Model usage. The tokens the language model reads and writes. Sometimes inside the fee, sometimes billed on top, sometimes paid by you directly to the model provider.
- Setup. Connecting the agent to your ERP and inbox, teaching it your customers' habits and setting the line where it stops and asks. Someone does this work: the vendor's engineers, a partner or your own team.
- Your own time. The people who review what the agent hands over, correct it and own the process. This line never reaches zero, and it should not.
Compare on the first line alone and you favour the vendor with the lowest fee and the most work left for your team. The build or buy question has the same trap: building looks cheap until you count the engineers who keep the agent running.
Which pricing models do AI agent vendors use?
AI agent vendors use 4 main pricing models: a fixed fee per process, a fee per seat, a usage price per document or per credit, and a build-it-yourself route where you pay the model provider and your own engineers. Each one moves a different risk to a different side of the table.
| Pricing model | What you pay for | Who carries the usage risk | Fits best when |
|---|---|---|---|
| Fixed fee per process | One agent running one process, such as order entry, at an agreed monthly fee | The vendor | Volume is steady and the process is clearly scoped |
| Per seat | Each person who uses the agent | You, through headcount | People work with the agent all day, as with a writing assistant |
| Usage: per document, per run or per credit | Each order, invoice or action the agent handles | You | Volume is low or very seasonal |
| Build it yourself | Model tokens, hosting and your own engineers | You | You have an AI engineering team and a process nobody sells |
Per-seat pricing sits awkwardly with back-office agents. The point of an order intake agent is that fewer people touch the orders, so a price that grows with the number of users rewards the wrong outcome. Per-seat pricing suits assistants that a person works with directly, and less so agents that work while nobody watches.
Usage pricing looks fair, and for low volumes it is. Its weakness is forecasting. A busy season, a customer who sends 40-page PDFs or an agent that takes more steps, and the bill moves. A fixed fee per process moves that risk to the vendor, which then has every reason to keep the agent efficient.
The number to ask every vendor for is the same: the total monthly cost of one process at your volume, with setup, model usage and support included. If they cannot give it, the price is not finished yet.
Are token costs for AI agents unpredictable?
Token costs vary per document, because a long PDF or a tricky exception makes the agent read and reason more. At back-office volumes the monthly total is still small and fairly stable. The real risk is an agent design that loops or rereads, which a cost per processed item exposes quickly.
A token is a piece of a word, and model providers charge per million of them, read and written. On Anthropic's published price list, read on 19 October 2026, its mid-range model Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, with a 50% discount for batch processing. Prices are falling fast. Stanford's 2025 AI Index report found that the cost of running a model at GPT-3.5 level fell more than 280x between November 2022 and October 2024.
Agents do use more tokens than a chat. In a June 2025 engineering report, Anthropic's own researchers wrote that agents typically use about 4x more tokens than chat interactions, and multi-agent systems about 15x more. That is where the worry about unpredictable bills comes from, and it is a fair one for open-ended research tasks.
Back-office work is narrower. Take an order intake agent. It reads an email and a PDF, checks the lines against your item master and prices, and creates the sales order. Assume a generous 40,000 input tokens and 4,000 output tokens per order. At the prices above, that is $0.08 plus $0.04: about $0.12 per order. At 2,000 orders a month the model bill is around $240. If our assumption were 5 times too low, it would be about $1,200 a month, still a fraction of 1 employee.
Two habits keep token costs in hand. Track the cost per processed item, per order or per invoice, rather than the monthly total, so a spike points to a cause. And ask the vendor who pays when usage rises: under a fixed fee that is their problem, under usage pricing it is yours.
How do you build a one-FTE business case for an AI agent?
You build it by measuring the hours one process takes today, pricing those hours at the full cost of an employee, and subtracting the hours that remain after the agent takes over. If the difference is clearly larger than the agent's total cost, the business case holds. One FTE is a useful threshold to test against.
The method takes an afternoon with the people who do the work:
- Pick one process. Order entry, supplier invoices or customer questions. One process with its own inbox and its own ERP step.
- Count the volume. Items per month, from the ERP or the inbox, not from memory.
- Time the work. Minutes per item, including lookups, corrections and the emails back and forth. Exceptions take longer, so time a few of them separately.
- Turn minutes into FTE. Monthly hours times 12, divided by the hours one person works in a year. We use 1,600 as a round figure; your HR team has the real one.
- Price the hour. Use your own full labour cost: salary, holiday pay, pension and employer contributions, not the gross salary alone.
- Set a realistic automated share. The share of items the agent finishes correctly without a person. Count the review time that stays on the rest. Our post on touchless rate and correct rate explains why the correct share is the one to plan with.
- Compare with the total cost. Fee, model usage, setup spread over the contract and your own team's time on the agent.
A worked example with invented but typical figures: 2,000 orders a month at 8 minutes each is about 267 hours a month, or 3,200 hours a year. At 1,600 hours per person that is 2 FTE. If the agent finishes 80% of orders correctly on its own and the rest still take 8 minutes, the desk saves 2,560 hours a year, about 1.6 FTE. At the Dutch average labour cost of €47.9 an hour that is about €122,600 a year of capacity, before the agent's cost.
The last step matters as much as the sum. Hours saved only become money when they go somewhere: a hire you do not make, a retirement you do not replace, or people moved to work that sells more. Decide that before the project starts, and write it down. Our earlier piece on measuring the ROI of AI automation covers the softer gains, such as speed and fewer errors, that this method leaves out.
What does the FTE business case look like in practice?
In practice the business case shows up as hours moved, not people dismissed. The published cases among our customers follow the same pattern: one process with several people's worth of manual entry, an agent that takes most of it, and the team moved to work the agent cannot do.
At Topa Bathroom Products, a Dutch bathroom wholesaler, about 3.8 full-time equivalents went into manually entering orders into Business Central. Today more than 90% of Topa's incoming orders reach Business Central without manual input. The 4 FTEs that did the entry moved to after-sales support and service planning. That is the order management agent doing what the business case assumed.
At Xpol, a Dutch flower wholesaler, the agent saves about 20 minutes per large order across around 150 orders a week. The company avoided 1 to 2 planned hires while volume grew and a senior specialist retired. That is the other way the hours turn into money: headcount stays flat while the work grows.
Koninklijke Dekker, a timber company with more than 140 years behind it, started with order intake and now has a roadmap for more of its back office:
"We started with automating our order intake process, we're saving about 4 FTE with that solution. Now we've got a full roadmap towards automating our full back-office. All with Lleverage." Mart Kruidenier, Continuous Improvement Specialist, Koninklijke Dekker
All 3 started with a process large enough to carry the cost on its own. The second process then runs on the same connections and the same company brain, so it costs less to start than the first.
How should a Dutch company manage AI agent costs?
A Dutch company should manage AI agent costs per process: one owner, one measured cost per item, one target share of work done without a person. High Dutch labour costs make the business case easier to clear, and a monthly review of cost per item keeps it honest after go-live.
The labour side of the sum is heavier in the Netherlands than in most of Europe. According to Eurostat's 2025 figures, average hourly labour costs were €47.9 in the Netherlands, against €34.9 across the EU and €38.2 in the euro area, the third highest after Luxembourg and Denmark. The same release puts non-wage costs at 24.8% of total labour costs across the EU. At 1,600 hours a year, €47.9 an hour comes to about €76,600 per FTE. An hour of manual order entry in Eindhoven costs more than in most of Europe, so the same agent clears the one-FTE threshold sooner.
Managing the cost after go-live comes down to 4 habits:
- Give every agent an owner. One person who answers for its volume, its accuracy and its cost.
- Report cost per item. Fee plus usage divided by the items handled each month. A rising figure is the early warning.
- Track the automated share and the review time. If people spend more time correcting the agent than planned, the FTE saving shrinks even when the fee stays flat.
- Keep the line where the agent stops in your hands. With thresholds and a full audit trail, you decide what the agent finishes alone and what waits for a person, and you can see what each decision cost.
Most of this does not need a finance project. A dashboard per process that shows volume, automated share and open exceptions gives the owner the figures for the monthly review.
Frequently asked questions
Is an AI agent cheaper than hiring an employee?
It is when the agent removes more hours than its total cost buys. Price those hours at your full labour cost: in the Netherlands, Eurostat puts the 2025 average at €47.9 an hour. If one process holds less than 1 FTE of manual work, a single agent rarely pays for itself unless further processes follow on the same setup.
How much do tokens cost per processed document?
For a typical order or invoice, cents. At $2 per million input tokens and $10 per million output tokens, a document that needs 40,000 tokens read and 4,000 written costs about $0.12. Long documents and complex exceptions cost more. Track cost per item each month so a change shows up early.
Is per-seat pricing a good fit for AI agents?
Rarely, for back-office agents. Their purpose is that fewer people handle the work, so a price that grows with the number of users rewards the opposite outcome. Per-seat pricing fits assistants a person works with all day. For an agent that runs a process on its own, a price per process is easier to compare.
How quickly does an AI agent pay for itself?
That depends on setup time and on how much of the work the agent finishes correctly. In our setup, forward deployed engineers work alongside your team, so work comes off the desk in the first weeks rather than at the end of a project. Count the payback from the day the agent handles real volume.
What should an AI agent quote include?
A complete quote states the monthly fee, whether model usage is included, the setup work and who does it, support after go-live, and what happens when volume rises. Ask for one total monthly figure for one process at your volume. That figure is what you hold against the cost of one employee.
Test the business case on one process
Start with the process that costs your team the most hours: count its monthly volume, time 20 items including the awkward ones, and price the hours at your full labour cost. If it comes to clearly more than 1 FTE, it is worth an agent. Bring those figures and a sample of real documents, and we will show the agent working on them, with the exceptions it hands to your team and the share it finishes on its own. Book a demo and we will work through the numbers with your operations lead.