Ask three agencies what a custom AI agent costs and you will get three numbers that are not comparable, because none of them is pricing the same thing. The model itself — the part everyone assumes is the expensive bit — is usually the smallest line. What you are actually buying is the work of establishing what the agent should do, connecting it to systems that were not designed to be connected, and keeping it correct once the process it automates starts to change.
So rather than a range that would be wrong for most readers, here is what a quote is made of. If you understand the five drivers below, you can predict roughly where your own project sits before you speak to anyone — and you can tell whether a quote you have been given is reasoning about your process or guessing.
The five things that move the price
- How well-defined the process already is
- This is the single biggest factor and it has nothing to do with AI. A task with clear rules that a new hire could be taught in a morning is cheap to automate. A task where three people each do it slightly differently and nobody has written down which one is correct is expensive — because the first job is deciding, and that is a conversation with your team, not a build.
- How many systems it has to touch
- An agent that reads one inbox and writes one reply is a small piece of work. An agent that reads the inbox, checks the CRM, looks up an order in a database, and writes back to all three is four integrations, four sets of credentials, four failure modes and four things that can change underneath it. Integration count drives cost more reliably than task complexity does.
- Whether your data is reachable
- Retrieval-augmented generation needs something to retrieve from. If your knowledge lives in a well-maintained help centre, that is close to free. If it lives in a shared drive of PDFs, some of them scanned, some of them out of date and none of them labelled, the work of making it retrievable is a real project on its own.
- How wrong the agent is allowed to be
- An agent drafting internal summaries can be wrong occasionally at almost no cost. An agent replying to customers, moving money, or making a commitment on your behalf cannot, and the difference is not a better prompt — it is evaluation, guardrails, escalation paths and review. Accuracy requirements are a cost multiplier, and they should be, because the alternative is discovering the limit in production.
- Volume
- This one affects the running cost rather than the build. Ten tasks a day and ten thousand tasks a day are the same agent, but the second one justifies far more optimisation work — and rewards it, because a change that saves a fraction of a cent per task is worth making at that volume and not worth the meeting at the first.
Build cost and running cost are different questions
A quote that gives you one number is hiding something. There are two costs and they behave differently over time.
The build is one-off: understanding the process, writing the agent, connecting the systems, testing it against the cases your team already knows are hard. The running cost is ongoing: model usage, monitoring, and the maintenance that keeps the agent matching a process which will not stay still.
What makes an estimate balloon after work starts
Overruns on these projects are rarely caused by the AI being harder than expected. In practice they come from three places, all of which are visible before a build starts if anyone looks.
- The process turned out not to be one process. What was described as "handling invoices" was four different workflows depending on the supplier, and nobody mentioned it because to the team it is obvious.
- A system had no usable way in. An integration that was assumed to be an API call turns out to need a workaround, and the workaround is now the project.
- The scope quietly moved from assisting to deciding. An agent that drafts a reply for review is a different accuracy problem from one that sends it, and the second one is often requested halfway through as though it were a small change.
All three are findable in advance. That is the entire argument for auditing before building: the audit is cheap relative to the build, and the things it catches are the things that otherwise double the build.
How to get a number for your own case
A credible quote requires someone to look at your actual workflow. Anyone who gives you a firm price for "an AI agent" without asking what it will touch, how much of it runs on undocumented judgement, and what happens when it gets something wrong, is quoting a template.
What you should expect instead: a walk through the workflows you nominate, a measurement of what they cost you today, and a scope that names what will be automated and what deliberately will not be. That baseline is also what any claim about savings has to be measured against later — without it, a reported saving is an assertion.
The cheapest automation project is the one where someone said no to the wrong first workflow.