What it costs to build an AI agent with an agency
Budget for four separate things: discovery, a pilot, the production build and the cost of running the agent afterwards. With a boutique or mid-size firm, a narrow first agent usually lands in the tens of thousands of euros, and a production system connected to several business tools reaches six figures. Enterprise integrators price well above that. The ranges below are indicative, not quotes, and almost no agency publishes a price list.
The four phases you pay for
Agencies split the work into phases because the risk is different in each. Discovery is cheap and tells you whether the task suits an agent at all. The pilot proves it on real data. Production is where most of the money goes, and it is mostly engineering around the model, not the model itself.
| Phase | What you get | Indicative range (boutique or mid-size firm) | Biggest cost driver |
|---|---|---|---|
| Discovery | Task selection, data and access review, success measures, a build plan | €5,000 to €25,000, often fixed fee | Number of processes and stakeholders reviewed |
| Pilot | One agent on one narrow task with a first evaluation set, used by a small group | €20,000 to €80,000 | Quality of data and how many tools the agent must call |
| Production build | Permissions, guardrails, monitoring, error handling, interface, staff training | €60,000 to €300,000 and up | Number of integrations, compliance needs, action risk |
| Run | Model usage, hosting, monitoring, evaluation upkeep, small changes | Roughly 10 to 20 percent of build cost per year, plus usage | Volume of tasks and how often the process changes |
What moves the build price
- Integrations. Each system the agent reads from or writes to adds authentication, error handling and testing. An agent that only answers from documents is far cheaper than one that issues refunds in an ERP.
- Action risk. Read-only agents need light controls. Agents that send messages, change records or spend money need approvals, limits and audit logs.
- Evaluation depth. A serious test set with a few hundred real cases takes domain experts' time, yours and the agency's.
- Data condition. Scattered, outdated or access-restricted knowledge has to be cleaned up before retrieval works.
- Compliance. Health, finance and public-sector work adds hosting constraints, documentation and security review.
- Interface. A chat box inside an existing tool is quick. A dedicated application with roles and history is a software project of its own.
Run costs after launch
Model usage is billed per token by the provider and scales with volume and with how many steps the agent takes per task. A multi-step agent can use many times the tokens of a single chat reply, so ask the agency to estimate cost per completed task during the pilot, not after.
The second run cost is people. Someone has to read traces, review failed cases, update prompts and tools when your systems change, and re-run evaluations when a model version is retired. Agencies sell this as a monthly retainer or a managed service. If you plan to take it in-house, budget for the handover and training.
Pricing models you will meet
| Model | How it works | Suits |
|---|---|---|
| Fixed-fee phase | A set price for discovery or a scoped pilot | First projects where you want a cap on spend |
| Time and materials | Daily or hourly rates for a named team | Production builds where scope will shift |
| Embedded engineers | A monthly fee per forward-deployed engineer | Companies with their own product team |
| Platform plus services | A licence for the vendor's runtime with services on top | Buyers who accept a proprietary platform |
| Managed service | A monthly fee to operate and improve the agent | Teams without internal AI operations staff |
Keeping a first agent project small
- Pick one task with a clear owner and a measurable outcome, such as time to first reply on support tickets.
- Start read-only or with a human approving every action, and widen permissions once the evaluation results justify it.
- Reuse the interfaces staff already work in instead of commissioning a new application.
- Ask for the evaluation set and the tracing setup as named deliverables, so later changes are cheap to verify.
- Agree a stop point after the pilot with a written decision rule for going to production.
Common questions
Why do quotes for the same agent differ so much?
Because agencies include different things. One quote may cover only a working prototype. Another includes evaluations, monitoring, security review, an interface and three months of support. Line the quotes up against the four phases and ask each bidder what is excluded.
Is model usage a large part of the bill?
Usually not at first. For most business agents the engineering and upkeep cost more than tokens in year one. Usage becomes significant at high volume or with long multi-step tasks, which is why a per-task cost estimate from the pilot matters.