A quote for an AI project is usually a quote to build a specific thing. That's the part worth noticing, because the thing has generally been chosen before anyone looked closely at your business. The scope arrives fully formed: a chatbot, an assistant, a custom model. The price follows from the scope, and the scope followed from what the agency builds.
That doesn't make it dishonest. It makes it unexamined, which costs you the same.
Four questions that price the problem instead
Ask these of any proposal, including ours. They're cheap to ask and expensive to skip.
1. What does this task cost us today?
In hours a week, on the actual clock, counted by the person who does it. If nobody has counted, that's the first job, and it takes an afternoon. A build that saves six hours a week is a different proposition from one that saves six hours a month, and the proposal usually doesn't say which it is.
2. Which part is the slow part?
Work that feels like one job is usually four, and they rarely cost the same. Pulling set fields out of a document, checking them, filing them, and answering questions about them are four different problems with four different answers — and only one of them is chat-shaped. Automate the wrong one and you've bought a faster version of the fast part.
3. Can the software we already pay for do this?
Ask it plainly, and ask someone with no stake in the answer. Microsoft 365, Google Workspace, your CRM and your accounting package have all shipped AI features in the last two years, and most businesses are paying for several of them already. Sometimes the honest answer is a licence upgrade and two days of setup. An agency that only builds has no reason to reach that conclusion, and no mechanism to.
4. What happens if it's wrong?
AI features produce confident output that is sometimes incorrect. That's fine when a person checks the result before it matters, and it is not fine when the output goes straight to a customer, an invoice or a compliance record. This single question decides the shape of the whole build — where a human sits in the loop, what gets logged, what the fallback is — and it's the question most likely to be missing from a proposal.
A proposal that answers all four is worth its number, whatever the number is. One that answers none is a price for a guess.
Why the estimate came in high
Large numbers on AI work usually come from three places, and only one of them is the AI:
- Integration. Getting the thing to read and write to systems you already run is almost always the bulk of the work. It's unglamorous, it's real, and it's where the hours genuinely go.
- Uncertainty. When nobody has mapped the workflow, the estimate has to cover the worst case. You're paying for the not-knowing.
- The model itself. Usually the smallest line, and often a per-use cost measured in cents.
Which means the reliable way to bring the number down isn't negotiating the build. It's removing the second one before anybody quotes.
What to do with a quote in front of you
Don't reject it and don't sign it. Spend a few days and a few hundred dollars establishing what the work costs you now, which part is slow, and whether anything you already own solves it. Then take that to the same agency. Either their number survives contact with the facts — in which case you buy with confidence — or it doesn't, and you've found that out for a fraction of the price.