What an AI agent implementation costs: the four line items
"What does it cost" is the first question on every call, and the one whose honest answer is "it depends on your process." True, but useless. Below we break the bill into the parts no implementation can skip, and show where budgets usually crack. You won't find specific figures for our work here — we set those after a free consultation, once we know how many processes and systems are involved. What you will find is a structure you can hold up against any quote, including a competitor's.
- The bill has four line items: audit, pilot, implementation and maintenance — everything else is a variant of those four.
- Model usage is usually the smallest item; integrations and cleaning up data are the expensive parts.
- If a process takes less than 20 hours a month, the implementation rarely pays for itself.
Four line items nobody gets to skip
Whatever the vendor and whatever the technology, the budget is built from the same four elements. If a quote is missing one of them, it isn't free — it will show up later as a change order.
Process audit. One-off, before anyone decides to build anything. A process map, the numbers run on your data, a feasibility call and the scope of the pilot. You are paying for a verdict on whether further spending makes sense — and for a document that stays with you even when the answer is no.
Pilot on a single process. A working agent on one narrow scope, usually in two to three weeks. The largest single item at the start, and the only one that gives you proof instead of a presentation.
Implementation and integrations. Connecting to the ERP, the CRM and the inbox, permissions, logging, team training. This item swings the most — two processes of similar complexity can differ several times over on the number of systems alone.
Maintenance and development. The recurring cost: hosting or licences, model usage, monitoring answer quality, rule fixes. It grows with the number of cases, not the number of users.
Where the budget quietly balloons
Three places account for most of the overruns we see in projects we take over from other vendors. None of them has anything to do with the model itself.
Data that has to be cleaned up first
An agent inherits the quality of the data it is given. If order statuses in the ERP are sometimes stale and the same company appears in the database under three different names, somebody has to sort that out before automation. That work sits on your side, rarely appears in quotes, and can add more to the schedule than building the agent does.
Systems with no API
An older ERP with no programming interface doesn't block an implementation, but it changes the price: instead of an off-the-shelf connection, the integration has to be built through the database, through exports or through an intermediate layer. The difference can be twofold, which is why we settle the list of systems during the audit rather than after the contract is signed.
Scope that grows along the way
The most common scenario: the pilot works, so halfway through the implementation three more case types get added. That is good news, but it is new scope, not a "small tweak." This is why we close the pilot on one process and only then decide about the next ones — each with its own bill.
Line items ordered by their share of a typical implementation budget. Model usage, the thing everyone worries about most, is usually the smallest item on the bill — measured in pennies per case handled. We confirm the order of magnitude by measuring it on the pilot, before you commit to an ongoing contract.
When you are better off not spending this money
A simple threshold: multiply the number of cases per month by the average time each one takes. If it comes to less than 20 hours a month, the implementation probably won't pay for itself in a reasonable time — even if it is technically doable. The same goes for processes where half the cases need a commercial decision or a negotiation: you automate 40% of the work at the full cost of the integration.
The signal in the other direction is just as clear. A process that takes several hundred hours a year, has clear rules and keeps its data in systems usually repays the implementation cost within a few months — and then simply stops costing anything. Which is why the first question in an audit isn't "can this be automated" but "what does it cost today."
Frequently asked questions
Why isn't there a price list on your site?
Because the effort depends on the number of processes, the number of systems and the state of the data, and we only learn those three things during a consultation. We quote the audit price up front, and that amount is deducted from the price of the pilot, so you never pay twice for the same analysis.
Does the cost rise as the number of cases grows?
Part of the maintenance cost does, because model usage is billed per case handled. But it grows more slowly than the manual work it replaces, and it doesn't mean hiring extra people for the busy season.
Can we start with something cheaper than a full implementation?
Yes, and that is usually how we do it. A free consultation gives you a rough estimate without any access to your systems, and a paid audit ends with the numbers run on your data. Only then do you decide about a pilot — and you are free to decide against it.