AI agent cost: what you pay for the build, operation and model usage

AI agent cost is set by the workflow the agent works in, far more than by the agent itself. Five factors drive it: how many systems the agent reads from and writes to, how varied the incoming documents are, which controls the process requires, how many cases arrive each year, and who runs the workflow after go-live.

Two agents on the same model can therefore cost very different amounts to build. An agent that classifies service requests and writes a draft reply into the ticket system needs one write integration and one release step. An agent that prepares quotes reads the article master, price lists and stock levels in the ERP, creates drafts there and enforces the discount floors that sales and finance have set.

It follows that any estimate measures volumes, handling times and the systems involved first and prices the build afterwards. These are the numbers a Workflow Assessment measures over four weeks before it ends with a fixed-price build scope.

The cost of AI implementation splits into one-off blocks for the build and running blocks for operation. The build covers the assessment, integrations, control set, test set, shadow mode and support for the people affected; operation covers model usage, hosting, monitoring and refinement.

Most weeks of a build go into integration, permissions, a test set made of real cases and agreeing with IT and compliance what gets logged, and only a few into work on the model. Pilots skip exactly this work, which is why they are fast and why they stall before production as soon as someone has to settle write access and release rights.

Internal time belongs in the estimate even though nobody invoices it. In a Workflow Build we plan for about two days a week from the workflow owner, plus an IT counterpart for access and architecture and a contact from security or compliance.

Model providers charge per token, a unit of text a few characters long, and price input and output separately. Input is everything the model reads, from instructions and rules to the email, its attachments and the context pulled from the ERP or CRM. Output is everything it writes, and for the models in this example it costs five times as much per token.

A worked example at list prices checked on 11 October 2026: OpenAI lists GPT-6.1 Sol and Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens. Take a hypothetical request for quote that arrives as an email with a PDF, about 4,000 tokens in total. Across all steps, from reading and matching against the article master to drafting, the agent reads about 40,000 tokens including instructions and ERP context and writes about 4,000. That comes to 8 US cents for input and 4 US cents for output, about 12 US cents per case, or around $120 a month at 1,000 requests.

Token prices change often, so every calculation needs a date. Google's price list already shows that the list price of Gemini 3.8 Flash doubles on 1 January 2027. The same price per token does not mean the same cost per case either, because Anthropic states that the tokeniser of its newer models produces roughly 30 per cent more tokens for the same text than the previous one. The figure becomes reliable only once real cases run through the workflow and the tokens are counted.

Breadth and risk make a workflow expensive; repetition makes it cheap. Every additional system brings an integration, every write permission brings status logic and tests, and every new document type adds cases to the test set.

Low volumes make each case expensive, because integration, control set and shadow mode are spread over few cases and shadow mode takes longer to collect enough comparisons per category. Regulation works the same way. A process that decides about people, such as job applications, can count as high-risk under Article 6(2) and Annex III of the EU AI Act, which brings in the requirements of Articles 8 to 15, from risk management to human oversight. Article 6(3) sets out exceptions, and the classification of a specific case stays with your compliance function or counsel.

Both routes lead to an agent in production, but they distribute cost and control differently. A platform brings operation, updates and ready-made components, and its price is tied to a usage unit. Microsoft Copilot Studio bills in Copilot Credits, where the number per response or action depends on the complexity of the task, and n8n prices its plans by monthly workflow executions.

A build in your own cloud tenant needs no platform licence. You pay for compute and model usage on the terms of your cloud contract, code, prompts and logs sit in your repositories, and your IT chooses the model. In return the workflow needs a team that runs and refines it, in-house or as a service such as Managed Run.

Which route costs less depends on volume, integration depth and audit duties. If the price rises with every case, the volume of the coming years matters more than that of the pilot year. If the available connectors do not reach the system of record, the integration work falls due on both routes, and auditors and data protection need a log per case that your organisation can read and retain either way. Cortaris builds in the client's tenant and sells no platform; the trade-off still depends on your volume and your IT strategy.

Set the budget from the value of the workflow, not from the first quote. The AI ROI calculator turns cases, handling time and fully loaded cost into hours saved, subtracts the running costs, and shows the payback in months and the build budget that pays back within twelve months.

Take a hypothetical quote desk that handles 1,000 requests a month at 30 minutes each, where the agent prepares 60 per cent of them so that handling time on those halves. That saves 1,800 hours a year, worth €108,000 gross at a fully loaded rate of €60 per hour. After assumed running costs of €2,000 a month, of which the model usage from the example above is a small part, €84,000 net remains. A build that should pay back within twelve months can therefore cost at most €84,000 in this example; the figure is a ceiling derived from value, not a market price.

The €60 rate is an assumption. Eurostat estimates average hourly labour costs for 2025 at €34.9 in the EU and €38.2 in the euro area, with industry the highest-paying sector in the euro area at €40.3, and a fully loaded rate adds workplace, IT and management on top. The two most uncertain inputs are the share the agent prepares and the time saved per case, so set both cautiously and decide on the cautious variant.