AI implementation cost: what you are actually paying for

A 40-person company asks three vendors about AI implementation and gets back $9,000, $38,000 and $95,000. All three quotes describe the same thing in one line: an agent for customer service. The difference lives in what the quotes leave out.

Published guides are no help here. One puts AI implementation cost at $5,000 to $500,000. Another says $10,000 to $10 million. A third says $50,000 to $500,000. Those are honest numbers describing different products, and none of them tells you where you land. Below are the three variables that explain the spread, plus the costs that only surface once you are a quarter in.

Why the published range runs from $5,000 to $3 million

The ranges disagree because the word "implementation" covers three different purchases.

The first is a subscription. Somewhere between $30 and $500 a month for a product that serves every customer the same way. You configure it yourself in a few days. You will not connect it to your CRM.

The second is a build on an existing engine, shaped to your process. Setup in the low tens of thousands, a monthly retainer, two to four weeks of work. Most projects for 10–200 person companies live here.

The third is a system built from scratch. Six figures, three to nine months, your own codebase. It earns its price when the process is genuinely unusual, or when the data cannot leave your infrastructure.

A vendor quoting "from $99/month" and a vendor quoting "from $60,000" are both telling the truth about different things. Settle which one you are buying before you compare any numbers.

The three things that actually move your price

Headcount barely matters. A 15-person company can pay more than a 150-person one if its process has more exceptions in it.

Three other things decide the number.

How many systems it touches. Every connection to a CRM, an inbox, a calendar or an order system is separate work and separate testing. How much authority it has. An agent that answers questions costs a fraction of an agent that issues a document or writes to your database, because the second one needs guardrails, an audit trail and a rollback path. What state your data is in. Cleaning up a knowledge base regularly takes longer than building the agent that reads it.

That third one is the one buyers underestimate, and the SERP agrees: published breakdowns put data preparation at 30–50% of the initial budget. If a quote says nothing about it, the vendor either has not planned for it or intends to bill it later.

The costs that surface in month three

The quote shows setup and a monthly fee. The invoice after a quarter looks different.

You will add model usage, billed in tokens. For a small company that is usually tens to a few hundred dollars a month, and the number tracks conversation volume, not headcount. You will add hosting if the system runs on your own infrastructure. You will add technical care, which industry breakdowns put at 15–30% of the build price per year. And you will add your own time on calibration, because the first four weeks are spent correcting answers rather than admiring results.

Ask about those four lines in the first call. A vendor who does not raise them either has not thought about them or would rather you did not.

Who owns the code when the engagement ends

Almost nobody asks this, and it decides your costs for years.

There are three possible answers. The code is yours and you receive the repository. The code belongs to the vendor and you hold a licence for the term of the contract. Or the system lives on somebody else's platform and there is nothing to hand over.

None of the three is wrong on its own. What is wrong is finding out which one applies while you are terminating. Two years out, the difference between the first and the third is the difference between a migration costing a few thousand dollars and a rebuild from zero.

A four-question estimate you can do in fifteen minutes

Take a sheet of paper and answer four things.

1. How many systems does the agent have to touch? Count each one separately, mailbox included.
2. Does it only answer, or does it also change something in your systems?
3. Is the process written down and are the documents in order, or does the knowledge live in people's heads?
4. How many requests or events per month does it need to handle?

One system, no write access, a documented process, under a hundred events a month: you are in subscription territory. Two to four systems, the agent takes actions, documents need work: you are in a five-figure build. More than four systems, or data that cannot leave the building: you are looking at a custom project.

This arithmetic will not replace a quote. It is enough to recognise a proposal that was written without reference to your situation.

What happens when you get in touch

The owner replies, not a salesperson. The call runs 45 minutes and ends with a price band, not a proposal. If it turns out your case closes inside an off-the-shelf tool for a few hundred dollars a month, we will say so.

The quote comes after the call and breaks into line items: integrations, data cleanup, build, testing, maintenance. You can run the calculator on this site before you speak to anyone.

The exit is stated up front too. The first scope is one process, not your whole operation. After it you have data and you decide. If the numbers do not add up, there is nothing to extend. The warranty covers fixes at no extra charge.