How much does an AI agent cost?

I get this question at almost every first call: how much does an AI agent cost? The honest answer is not a number, and I am wary of anyone who gives you one before seeing your case. Here is what really drives the price, so you can read a quote instead of just accepting it.

Why there is no catalog price

An agent for 19 euros a month exists: it is software that is already built, shared across thousands of customers, that you configure at the edges. Useful, sometimes enough. A custom agent is a different thing: it talks to your tools, follows your rules, acts inside your systems. Nobody built it before you, so nobody can sell it to you at the price of a subscription. The catalog price does not exist because the product does not exist yet. When a custom quote looks like a subscription fee, that is not a bargain: usually it means the scope is vague, or the maintenance vanished from the math.

The two budgets you must not confuse: build and run

• The build, once: scoping, connecting to your tools, business rules, testing, going live. A one-off cost, and the more the agent touches real systems, the heavier it gets.
• The run, every month: the tokens it consumes, hosting, monitoring, fixes. This is the budget people forget to price, and it is the one that decides whether the agent stays worth it over time.

What actually pushes the bill up

• The number of tools plugged in. An agent that reads one database is simple. An agent that orchestrates several tools and several steps is a different job, and a different budget.
• The depth of the integrations. Reading data is easy. Writing into a production system (creating an order, updating a customer record) needs guardrails, which means time.
• The level of reliability required. An agent that is wrong one time in ten is built fast. An agent you trust with customers or money has to be tested, bounded, supervised. That is where most of the budget goes.

Token cost: why it is unpredictable

Every time the agent thinks or answers, it burns tokens, the unit the model bills. The catch: that consumption tracks real usage, not a flat fee. An agent handling ten requests a day and the same agent handling ten thousand do not get the same bill. That is why a run price quoted off the cuff, with no volume assumption, means nothing. The good news: it can be bounded, if you decide to at design time. Pick the right model for each step (a small one to sort, a big one only when needed), cache what repeats, cap the volumes. On a bill already in production, an audit often finds 30 to 70% savings on an audited LLM bill, at equal quality. Ask your provider how they plan to cut the LLM bill rather than just absorb it.

Maintenance: the line everyone underestimates

An agent is not a piece of furniture, it is living software, wired to models and tools that move without warning you. A model gets deprecated, an API changes, an unforeseen case shows up: without upkeep, the agent degrades quietly. Maintenance is not an option you add if it breaks, it is a budget line from day one. An agent shipped then abandoned costs twice: once to build it, once to repair the damage.

An agent with no maintenance budget is not cheaper. It is just more expensive, later.

Realistic ranges, as logic rather than numbers

I will not throw a random price at you: that would be dishonest without seeing your case. The logic, though, is clear. Take the simplest kind of agent as your base unit, and watch what each added capability does to it.

• Read-only agent, answering from your documents without changing anything: the floor. This is the ground for a sourced answer in 2 seconds, with 70% less searching for your team.
• Agent that acts in a system, opens a ticket or updates a record: above that, because you now have to handle errors and edge cases.
• Multi-tool, high-reliability agent, like 24/7 support handling sensitive actions: often an order of magnitude above the first on build, with a monthly run that never stops. This is the price of 24/7 support that escalates 50% fewer tickets.
• The rule: every capability you add (writing, one more tool, more reliability) raises the build and the run at the same time. The price does not follow what the agent says, it follows what it is allowed to do.

The 5 questions to ask a provider before you sign

• Is this budget the build or the run? If the two are not separated, the quote is incomplete.
• Who pays for the tokens, and on what volume assumption? What happens if it is exceeded?
• What happens to the agent when the model changes or is deprecated? Is maintenance included, or billed separately?
• Which actions go through a human check before they fire? (Sending, paying, deleting, writing to a customer.)
• Is this custom work I own, or a repackaged SaaS? If we part ways, what stays mine? An agent built on an open standard like MCP (adopted by Anthropic, OpenAI, Google and Microsoft, handed to the Linux Foundation in late 2025) does not lock you to one provider.

When an agent is not worth its cost

The best advice I can give about the price of an agent is sometimes not to build one. If your need is "when X happens, do Y", with no judgment to apply and no language to interpret, an agent is too expensive and too fragile for the task. A plain automation workflow does the job, costs far less, and breaks far less often. Keep the agent for what genuinely requires understanding, deciding, writing. Paying for an agent to do what a simple "if this, then that" rule handles perfectly is throwing money away.

If you have an agent quote in front of you and you cannot tell what you are paying for, that is exactly where I help: framing the scope, splitting build from run, and pricing the real cost before building. A first call is usually enough to see clearly, with no commitment: let's talk.