How much does an AI agent cost? The real numbers, and why they mean almost nothing

Corentin Pinelco-founderJune 10, 20268 min read

TL;DR

A custom AI agent runs 6,000 to 12,000 USD to build and 400 to 1,200 USD a month to run. But the number alone means almost nothing. What decides whether it is a great spend or wasted money is your use: your goal, your stack, your volume.

You came for a price. Here it is, no detour.

From experience, in France and in the United States, a custom AI agent is built for 6,000 to 12,000 USD. Count 400 to 1,200 USD a month to run it. At the high end of the US market, I have seen complex agents billed at 75,000 USD and up.

Now, the truth few providers will tell you: that number, alone, means almost nothing.

The same price can be the best spend of your year. Or money thrown out the window. It does not depend on the agency. It depends on your use.

The real numbers, to place the market

I will not drown you. These are the orders of magnitude I see in practice.

Custom conversational agent: 6,000 to 12,000 USD to build, 400 to 1,200 USD a month to run.
Simple, single-use agent: 1,500 to 5,000 USD (figure from the 2026 Destilabs guide).
Serious support build with document retrieval: 8,000 to 25,000 USD (2026 Kellton benchmark).
Complex agent wired to CRM and analytics: from 75,000 USD at the high end of the US market.

Before the price: quantity or quality?

This is the biggest cost lever. Not the tech.

Ask yourself what you are really after.

Chasing short-term ROI? You are playing quantity. Absorb volume, qualify leads, deflect repetitive tickets. You optimize cost per interaction. You start simple. It pays back in weeks.

Chasing experience? You are playing quality. An agent that carries your brand voice, handles delicate cases, never lets a good client down. There you invest in care, guardrails, testing. And the price goes up, for good reasons.

Two goals. Two budgets. The expensive mistake: paying for quality while thinking you are buying quantity. Or the reverse.

The stack: no backhoe for a single flower

You will want the latest model. The most powerful one. The one everyone talks about. That is human.

But digging the hole for a single flower with a backhoe is expensive and slow to set up. The result is no better than with a spade.

To qualify a lead, answer a FAQ, book a meeting: a smaller, cheaper model does the job. Fast and well.

The frontier model earns its place when the task truly deserves it: complex reasoning, nuance, high stakes.

Choosing your stack means taking the tool just big enough for the hole you are digging. Not the most impressive one.

Worth knowing: the infrastructure (model tokens, hosting, database) is billed at cost, separately. With prompt caching, a qualification conversation costs a few cents. The model is almost never the line that hurts.

The three other variables that move the price

Once the goal and the stack are set, the rest is quick to price.

Scope of action. An agent that only answers costs little. One that qualifies, books, writes to your CRM and escalates to a human is another system. Every action in another tool is real integration work.

The state of your integrations. A clean, documented API connects fast. A legacy CRM, WhatsApp, a calendar and a data warehouse, each with its own auth and traps: that is where the budget goes.

The quality of your data. Clean base, direct retrieval. Base scattered across PDFs, spreadsheets and inboxes: someone has to consolidate it first. That work is in the quote, not in a surprise invoice.

That is why a 5,000 USD build and a 200,000 USD build can both be honest. It is not the same product.

The trap: the hidden costs of no-code

No-code tools are a legitimate way to start. They are also where budgets leak in silence.

Rent that climbs with your success. The per-conversation bill rises exactly when the tool starts working.
Lock-in. The flows almost never export. Leaving means rebuilding everything.
The stack pile-up. A chatbot tool, plus an automation tool, plus a database. Each with its subscription, its limits, its outages.
The last mile. The demo takes a day. Writing reliably to your CRM, handling edge cases, monitoring in production: that is where no-code runs out of road.

The real question

So before asking me for a number, ask yourself: what is my real use? My goal? My volume?

Without those answers, any quote is just a number in the air.

If you want to place your case, our comparator puts the real cost of a conversational agent from more than ten vendors side by side. Handy for an order of magnitude.

But it will not answer the only question that matters for you: do you really have the use for this, and why?