An AI consultancy engagement runs from about $20,000 for a scoped assessment to fixed-fee builds in the low six figures; here is what moves the number. We publish our prices, which is unusual enough that we get asked why, and the honest answer is that a consultancy which will not say its number before it has told you its story is asking you to be sold to before you are allowed to compare. We would rather you had the number first and the story second.
What you are actually buying at each stage
There are three shapes a fee takes and they buy different things. An assessment buys a decision. Two to six weeks, ending in a written answer about whether to build, and what. Ours is Map + POC, about $20,000 for two to four weeks. A build buys a working system in front of real users; for a mid-market business the fixed fees we see and charge sit in the low six figures. A retainer buys a consultancy's continued presence, and it is the fee most likely to outlive its purpose.
The mistake buyers make is comparing builds to builds. Compare assessments to assessments, because the assessment is where the risk is priced. Anyone quoting a build without one has guessed at the scope, and the guess is in the number. Usually as padding, occasionally as a loss-leader for the retainer.
Where the money goes, and where it does not
People assume the model is the expensive part. It is not; the model is a line item that gets cheaper every quarter. The expense in an assessment is time with people: leadership first, then the people who do the work, because the gap between what the first group describes and what the second group does on a Tuesday afternoon is nearly always the finding. The proof of concept itself is the cheap part.
The expense in a build is everything that makes a system fit for a business to trust. Permissions inside your own environment. A review gate on every consequential action. An audit trail on every answer. Rollback. Training the people who will run it once we have gone. None of that is glamorous and all of it is where the hours go. Which brings us to the cheap proposal.
The cheapest proposal is usually the most expensive
A proposal that is markedly cheaper than the others has not found a cheaper way to do the same work. It has left something out, and it is nearly always one of the items above, because those are the ones a buyer cannot see in a demo. You will pay for the missing item later, at a worse rate, either as a change request or as the cost of a system nobody trusts enough to use. The second is more common and much more expensive, because it looks like success until you count who is actually using the thing.
Fixed fee or time and materials
Fixed, for each stage, and we hold this opinion strongly. Time and materials moves the risk of a hard problem from the people who claim to understand it onto the people who are paying to find out. There is one honest exception: genuine research problems where nobody can scope the work, and in those cases the right structure is a short fixed-fee assessment to find out whether it is a research problem at all. If a consultancy says your problem cannot be scoped, ask them what they would need to know to scope it, and buy that.
What should move a fixed number is scope: how many parts of the business we look at. One workflow in one team is the bottom of the range. Four workflows across three functions is not four times the price, but it is not one times either. What should never move the number is how the engagement goes. If a fee can rise because the work turned out harder, it was never fixed and the risk was always yours.
The one clause that tells you whether a fee is real
Ask what happens at week four if the answer is no. A real fixed fee for an assessment costs the same whichever answer it produces, which means the consultancy is paid the same to tell you not to build as to tell you to build. That is the only structure in which "no" can be said honestly. A consultancy that has never told a client not to proceed is not unusually lucky in its clients. It is being paid to not say it, and the fee structure is how.
When to stop paying a retainer
A retainer should shrink. If your own team is running the system and the next workflow cost less than the last because the scaffolding already existed, it is doing its job. If two years on the same people are doing the same work at the same rate, the consultancy has made itself permanent, and we would call that a failure whatever the invoices say. Put a review date in the contract and ask, at each one, what you would lose by stopping. If the answer is "nothing much", stop.
Our numbers, for the record
Map + POC, about $20,000, two to four weeks, ending in a written recommendation you keep. Build, fixed fee scoped during Map, six to twelve weeks. Run, ongoing, each build cheaper than the last. You can stop after any stage and some clients do. The detail is on our pricing page, which is where it should be.