Service — from $15k
One workflow that works, rather than five that demo.
Most failed AI MVPs did not fail because the model was not good enough. They failed because the scope was five features wide and none of them were finished, so there was nothing to put in front of a user and nothing to measure — which meant the next decision got made on opinion.
An MVP’s only real job is to convert an argument into evidence. Whether users want it. Whether the accuracy is good enough for the thing it has to do. Whether the unit economics survive contact with real volume. All three questions need one path working end to end, and none of them need five.
So the scope gets cut hard, and the evals get written before the feature does. That second part is unusual and it is the whole reason version two is a measurement rather than a debate.
The short answer
AI MVP development from Code and Proof starts at $15,000 and ships a first working version of an AI product — narrow enough to build honestly, real enough to put in front of users. The scope is cut to one workflow that works end to end rather than five that demo, and the evals are written first so the second version can be measured against the first instead of argued about.
How it runs
Six steps, in this order. Step two is the one that gets skipped everywhere else and the one that pays for itself fastest.
Cut to one workflow
Not the product — one path through it, from the input a real user actually has to the output they actually need. Five half-working paths cannot be evaluated, cannot be demoed honestly, and cannot be charged for.
Write the evals first
Twenty to fifty real examples with known-correct answers, before the feature exists. This is the single highest-leverage hour in the project: without it, version two is an argument, and with it, it is a number.
Build the boring path
The case that happens 80% of the time, working properly, with the failures visible. Edge cases are a v2 decision made with data, not a v1 guess made with optimism.
Make failure legible
An MVP that fails quietly teaches you nothing and will mislead your first users. This one refuses, flags, and says which field it was unsure about — because the point of a first version is to find out what is actually hard.
Ship it to real inputs
Your documents, your customers' phrasing, your edge cases — not the clean sample set. Almost everything worth learning shows up in the first week of real input and none of it shows up before.
Hand it over
The repository, the eval set, the written reasoning. If the next step is an in-house hire rather than more of me, that should be a cheap move to make.
What is in and what is not
Written down before the work starts rather than negotiated during it. A first version whose scope is still moving is not a first version.
When this is the wrong page
If the scope genuinely cannot be cut to one workflow — it needs three integrations on day one, or the agent has to act on its own before a human sees it — then it is not an MVP, and pricing it as one would be setting up a fixed-price project to run over. That is AI application development or AI agent development.
And if you are not yet sure the thing should be built at all, the cost guide is free, includes the arithmetic, and is honest about when the answer is to not build it. Read that first — it is a better use of an afternoon than a call with me.
Common questions
- How long does an AI MVP take to build?
- An AI MVP from Code and Proof starts at $15,000 and is scoped to one workflow that works end to end rather than five that demo. The evals are written before the feature, so the second version can be measured against the first instead of argued about. If the scope cannot be cut to one workflow, it is not an MVP and it is priced as an application build.
Related
and I will tell you what fits inside it.