How much does an AI agent cost — the five items on the bill no price list adds up

Asked “how much does an AI agent cost”, price lists answer with a range — from a few to several tens of thousands of zloty. The figure is true, and it is the least useful of the five items on the bill, because without the other four there is nothing to compare it with. That range is really the price of integration with permissions and an activity log; alongside it you pay for the model (tokens times the number of tasks), for getting the data in order, for the hours people spend approving results, and for maintenance. Below we itemise those five and add up twelve months at prices read from suppliers’ own pages on 7 September 2026.

This text is for companies budgeting a first agent — for the owner, the finance director or the head of operations, not for a developer comparing APIs. On AI agents for business, ranges get published and arithmetic rarely does; here is the arithmetic, and the answer to when not to build. We have covered how an agent differs from a chatbot separately — here it is enough that an agent acts in systems rather than only answering.

Amounts are quoted in Polish zloty (PLN) and are net, that is excluding VAT.

Start with two numbers no price list has

Before you open any price list, settle two numbers no supplier will work out for you.

First: how many times a month is the task the agent is to take over actually done? Count it from the last three months — from the inbox, from the ticketing system — rather than estimating from memory, which inflates the irritating tasks and undercounts the ones done quietly. This number moves the model line item linearly, but leaves the integration cost alone, which is largely fixed. Hence the rule from our methodology: a process that takes two hours a month is almost certainly not worth an integration project.

Second: what happens after a mistake, and what it costs to undo. A ticket with the wrong order number takes someone a minute to fix; a reply with the wrong price sent to a customer, or a status change that triggered a shipment, takes longer to reverse than the task took to do. This number does not go into the spreadsheet directly, but it sets the permissions design — what the agent does by itself, and what always waits for a person — and through that the oversight hours: the most expensive part of the bill, and the part price lists are silent about.

The five items on the bill for an AI agent

The cost of an AI agent falls into five items. Two of them — the model and the platform — can be read off public price lists, which is why most of what gets written is about those. The other three — integrations, data and oversight — depend on your process, and they are what make the bill.

1. Model: tokens times the number of tasks times the price

Tokens are the fragments of text a model breaks everything it reads and writes into; you pay separately for input (the instruction, the tool descriptions, the content of the ticket, the results read back from your systems) and for output (the model’s answer), five times dearer in the price list below.

A chatbot makes one model call per answer; an agent makes several per task — plan, read from the ERP, read from the CRM, write, draft the reply — so it burns a multiple of a chatbot’s tokens. Every call repeats the same fixed part of the input: the system instruction and the tool descriptions. Providers let you cache that part (prompt caching) and charge a fraction of the price to read it back; writing to the cache is billed separately — check the price list.

We read the prices from the Claude API price list on 7 September 2026 and the example below uses them; other providers publish comparable per-million-token lists.

Three notes. The introductory Claude Sonnet 5 price — $2 and $10 per million tokens — has been made permanent, and the increase to $3 and $15 announced for 1 September 2026 was cancelled (pricing page, 7 September 2026). Billing is in dollars only; we convert at the average rate published by NBP, the National Bank of Poland, for 4 September 2026 — 3.7145 PLN to the dollar (table A no. 172/A/NBP/2026), the last rate before the publication date. Amounts are net; settle the VAT treatment of a service bought abroad with your accountant. The Batch API gives a 50% discount, but only for work that can wait hours for a result — an agent handling a ticket rarely can.

And a rule: measure, do not guess. Every API response returns the number of tokens used, so a two-week pilot gives you the real figure per task on your own tickets; the assumptions in the example below are only a placeholder for it.

2. Integrations, permissions and the activity log: count systems, not features

This is the item that makes an agent cost more than automating the same process. You count it in systems, not features: every system means authentication, data mapping (what the ERP calls a “counterparty” and the CRM a “company”), error handling, rate limits and — usually the longest item in the schedule — waiting for access from whoever owns the system, which in smaller companies is often the external ERP supplier.

Writing costs more than reading: run twice, it creates a duplicate ticket, and a wrong write has to be undone. So a write integration needs protection against repeats (idempotency), a rollback procedure and a test environment — hours of work that a quote for an “agent integrated with the CRM” either includes or does not.

On top of that, four decisions about permissions, taken before the first line of code: the read scope, the write scope (usually a much shorter list), the operations that need a person’s consent (anything that cannot be undone: sending to a customer, a payment, a status change) and the activity log (what the agent did and on what basis). We describe those four decisions, and what happens after a mistake, on our page about AI agent implementation. Training the people who will approve the output is a stage of the implementation, not a separate item.

A rule for you: count separately the systems the agent “only reads” and the ones it “writes to”, and insist on both lists in the quote. “Integration with the company’s systems” without a list is not a quote, only a range.

3. Data and the knowledge base

The test takes one afternoon: try exporting the last hundred cases of this process as a single table. If that works, this item is close to zero. If the tickets sit in inboxes, the agreements in people’s heads and the history in email attachments — you have just seen where the first part of the budget goes, and it is not the agent’s fault.

If the agent is to answer from the company’s own knowledge — procedures, price lists, manuals — it needs a knowledge base (RAG), and getting that in order can be the dominant cost of the project. It is the same work a chatbot needs; which is why the price of a chatbot is mostly the price of its knowledge base, not of the conversation.

A separate decision: which data leaves for the model provider, to whom, and on what legal basis. That is a design decision with a cost of its own, set out on our page about security and GDPR.

4. Oversight: the human hours that appear in no quote

This item is in none of the price lists cited below. Our rule reads: “Anything that goes out to customers or changes the state of a system starts in human-approval mode. Loosening oversight is a decision taken on pilot data, not an assumption in the quote.” If that holds, someone is checking the output — and their hours are a project cost, even though they are on no supplier’s invoice.

The arithmetic is simple. In the pilot, everything gets checked: number of tasks times minutes per result. After the pilot what remains is spot-checking (the share of results someone still looks at, times minutes) and exceptions — cases the agent did not close and handed back to a person (share times minutes, usually several times more). You set both shares from the list of mistakes from the pilot, not from an assumption.

This item decides more than the cost. At a thousand tasks a month, checking everything is thirty-odd hours — carryable. At ten thousand it is over three hundred hours and nobody will do it; at that point the feasibility of the project turns on the oversight design: which categories of task may run unattended after the pilot, and which always wait for a person. Exceptions do not disappear — they change owner.

5. How much an AI agent costs to run: platform, monitoring and retired model versions

Platform. If the agent runs inside an automation tool, you pay its subscription: n8n Cloud Starter is 20 euros a month billed annually for 2,500 executions, Pro 50 euros for 10,000, and the Community edition on your own server is free, but someone has to maintain it (n8n pricing, as at 7 September 2026). One task is one execution, so a thousand tasks fit inside Starter. If the agent loop runs at the model provider, Anthropic bills Managed Agents at $0.08 per session hour (same price list): a thousand tasks at three minutes each is 50 hours, about $4 a month. Hosting the loop itself is cheap; the integrations are not.

Monitoring. Someone looks at the activity log and the exception rate every week, rather than after a customer rings. A few hours a month, but every month.

Retired model versions. Providers retire older versions, and an agent built on a retired version has to be retested from scratch — with its permissions, activity log and exceptions. The provider sets the timing, not you; budget it as a fixed line, not a margin for error.

What Polish agencies publish — and how to read it

Below are four published price lists with dates — price lists, not market data: we know of no independent Polish study of agent prices, and averaging other people’s ranges produces a number that describes no real project. Amounts in PLN net.

How to read it. The bottom and the top of each range are two different projects — they differ in the number of systems and the extent of write access, and which one is yours is settled by the list of systems, not by negotiation. If you are asking what a chatbot costs, look at the first two rows: the difference between them is the cost of the knowledge base. The single-automation row shows how the price of AI automation compares with the price of an agent at the same supplier — several times lower, because it carries no permissions, no activity log and no pilot with approval; the price of an AI agent without those three is the price of something else.

Before you compare figures, ask what is inside: how many systems and which, read or write, whether there is an activity log, whether a pilot with approval is in the price, and who ends up owning the code and the prompts. Test an “agent from 5,000 PLN” with one question: what has to be done when it gets something wrong? If only correcting the answer — that is a chatbot, possibly a good one, but a different thing from what the name suggests.

We have no price list of our own and will not quote a range without your process. These numbers are here so you know what someone left out, not so you can average them.

Worked example: an AI agent handling tickets over twelve months

Illustrative figures — this is not a client implementation; swap the assumptions for your own. Scenario: a B2B wholesaler receives 1,000 service tickets a month by email. The agent reads the ticket, checks the order in the ERP (read) and the warranty status in the CRM (read), creates an entry in the ticketing system (write) and drafts a reply that a person sends (an operation requiring consent).

1. Model (12 months, 1,000 tasks a month): Haiku 4.5 ≈ 3,009 PLN; Sonnet 5 ≈ 6,017 PLN, that is about 501 PLN a month, and with the fixed part read from cache ≈ 3,610 PLN (40% less; we ignore the cache write); Sonnet 4.6 ≈ 9,026 PLN; Opus 5 ≈ 15,044 PLN; Fable 5.1 ≈ 30,087 PLN. The spread is tenfold: at high volume, picking the cheapest model that clears your quality bar shows up in the bill — but even the dearest is not the largest item here.

2. Build: integrations, permissions, activity log. We have no quote of our own without the process, so someone else’s price list stands in for yours: on 18 July 2026 PixelAgency published a range of 15,000–45,000 PLN for an agent integrated with a CRM. We work out both ends.

3. Data: 0 PLN, because in this scenario the last hundred tickets export in a quarter of an hour; at your company it may be a quote of its own.

4. Oversight. Pilot: 2 months × 1,000 tasks × 2 minutes = 33 hours a month. After the pilot: 20% spot checks × 2 minutes + 10% exceptions × 10 minutes = 23 hours a month. About 300 hours in the first year. We price them at an illustrative 80 PLN an hour of employer cost — not a benchmark, just a placeholder for your number — so 24,000 PLN.

5. Maintenance. n8n Cloud Starter: 20 euros × 12 × 4.3179 PLN (NBP, 4 September 2026) ≈ 1,036 PLN. Monitoring at 4 hours a month plus two rounds of retesting after a model version change at 16 hours each = 80 hours a year, at the same illustrative rate 6,400 PLN — or at your supplier’s rate.

First year on Sonnet 5 without caching: about 52,500 PLN at the bottom of the build range and about 82,500 PLN at the top; the model’s share 11% and 7%. On Fable 5.1: about 76,500 and 106,500 PLN, the model’s share 39% and 28%. Second year (no build, Sonnet 5): about 35,900 PLN, of which the largest item is oversight hours — 280 hours.

Two sensitivities. At 100 tasks a month the model item drops to about 600 PLN a year, and the build does not change at all — the integration cost is fixed, which is why a rare process almost never pays for itself. At 10,000 tasks the model item rises tenfold (Sonnet 5 ≈ 60,000 PLN a year), and checking everything in the pilot would need 333 hours a month — undoable; scale only becomes possible by grading autonomy after the pilot.

Compare hours, not money. If one ticket took a person 12 minutes before the implementation, that was 2,400 hours a year; the first year with an agent frees about 2,100 of them after subtracting oversight. Whether those hours are worth the bill is yours to decide — we do not convert them into headcount.

When not to build an agent

• The order of the steps can be written out in advance. Then it is automation, not an agent — cheaper and more predictable.
• The task happens rarely. Two hours a month is too little for an integration project whose cost does not fall with volume.
• The data cannot be exported and nobody plans to change that. An agent will not fix the data, only make its state visible — at greater cost.
• There is no process owner. Someone has to settle exceptions and decide what the agent does alone; without that person, oversight has no addressee.
• An off-the-shelf subscription tool exists. Do not build what you can buy — we will say that even when it means no work for us.
• Regulation rules the scenario out. We do not build candidate scoring for recruitment or cold email machines, whatever the budget.
• The agent is to talk to customers. That does not rule the project out, but it adds the disclosure duties under Article 50 of the EU AI Act, in force since 2 August 2026 — the customer has to know they are talking to a machine; we have collected the deadlines in the AI Act technical compliance sprint.

This describes the technical and organisational side, not legal advice — the legal assessment belongs to your lawyer or data protection officer.

How to work it out for yourself

1. Count the tasks from the last three months — from the system, not from memory.
2. Read the last hundred and split them into repeatable ones and exceptions. The first are in scope; the second set the oversight.
3. List the systems: separately the ones the agent reads from, separately the ones it writes to.
4. Settle which operations always wait for a person, and measure how many minutes it takes to check one result.
5. Ask every supplier for the five items in writing: the model and an estimate of tokens per task; integrations with the list of systems; the work on the data; oversight hours during the pilot and after it; maintenance including retesting after a model version change. And one question: who ends up owning the code and the prompts.
6. Add up twelve months at two volumes — the current one and double it — and set the result against the hours from point one.

The cost of an AI agent is the items specific to agents — tokens, permissions, oversight — plus the factors common to any implementation, not only agents; we itemised the common ones in the piece on how much an AI implementation costs.

What next

We work so that the implementation quote comes from data rather than a range on a price list: the audit (1–2 weeks) and the pilot (2–4 weeks) have a fixed scope, and the implementation (4–12 weeks) is priced on the pilot’s numbers — real tokens per task, the exception rate and the minutes spent checking. A negative pilot result is a result, not a failure: it saves the second and fourth items on the bill. The stages are set out in our implementation methodology, and the scope of an agent project on the page about AI agent implementation for business.

Describe the process to us and the number of tasks a month. We will tell you which item we would start counting from — including when the answer is “not worth it yet”.