Service — from $15k
An AI product, built end to end by the engineer who has to keep it running.
There is no shortage of people who will build you an AI demo. The demo is a solved problem and has been for two years. What is scarce is someone who builds the version that is still correct six months later, with real customers in it, when the inputs stop looking like the ones in the test set.
The difference is not the model and it is not the prompt. It is what the system does when it is wrong — whether it says so, whether you can see why, and whether anything irreversible happened before a person looked. Those are architectural decisions, they get made in the first week, and they are expensive to retrofit.
So they get made deliberately here, and they get written down. Every output traces to the evidence it came from, a human approves anything that cannot be undone, and the pipeline is expected to notice its own gaps rather than go quiet.
The short answer
AI application development from Code and Proof starts at $15,000 and covers building an AI product end to end — the pipeline, the evals, the review surface a human actually uses, and the operational spine that keeps it running. The difference from a general development shop is what happens when the model is wrong: every output traces to its evidence, anything irreversible waits behind a human gate, and the pipeline notices its own gaps rather than failing silently.
The six layers
An AI application is not a model with a form in front of it. It is six things, and the model is the one that takes the least time.
The pipeline
Input in whatever state it actually arrives — PDFs, photographs, transcripts, spreadsheets someone made by hand — and structured, checkable records out. This is the part everyone imagines when they say “AI feature”, and it is about a fifth of the work.
The evals
A golden set, and a score the pipeline is measured against before anyone argues about whether a prompt change helped. Written first, so version two can be compared to version one rather than debated.
The review surface
Somewhere a person sees what the system produced, next to the evidence it came from, and accepts or corrects it. The queue design decides whether the feature saves time or just relocates the work.
Provenance
Every record carries the page, timestamp, or frame it came from. A number you cannot point at is not an answer — and the first time a customer disputes one, this is the only thing that settles it.
The operational spine
Retries that do not compound, idempotency on anything with a side effect, cost ceilings enforced in the runtime, and alerting that goes somewhere a person reads. The unglamorous half, and the half that decides whether it is still running in six months.
Handover
Written documentation, the eval set, and the reasoning behind the decisions — yours, in your repository. A system only I can operate is a liability I have sold you, not an asset.
Which shape is yours
“Build us an AI application” covers at least six different engagements with different prices and different risks. This is the table that sorts them, and it is worth reading before an intro call rather than during one.
AI inside an existing SaaS product
The most common version of this request is not a new product at all. It is a SaaS company with paying customers, a roadmap, and a board asking what the AI story is — and the constraint is everything already in the database.
That work is real and it is welcome, but it is a different job from greenfield and it is priced differently, because the hard part moves. Greenfield risk is whether the thing works. Brownfield risk is whether shipping it breaks what already pays you. The relevant page is AI integration.
What is not offered here: general SaaS development. If what you need is authentication, billing, a dashboard and an infrastructure setup with no AI in the interesting part, there are hundreds of shops that do that well and more cheaply than I would. Saying so costs me a proposal occasionally and it is still the right answer.
What it costs
Application builds start at $15,000, scoped from what the thing actually has to do. The bands are published in full in what an AI feature costs to build, including the case for not building it — a thin feature runs $8k–$20k, a real pipeline $25k–$60k.
If your budget is well under that, the honest move is to cut the scope rather than shop the price. A cheaper build of the same surface is not a saving; it is the same work done once badly and then again properly, and you pay for both.
Common questions
- What does it cost to build an AI application?
- AI application development from Code and Proof starts at $15,000. That covers an AI product built end to end — the pipeline, the evals, the review surface a person actually uses, and the operational spine that keeps it running. The figure matches the published cost guide on this site rather than undercutting it: a thin AI feature runs $8k-$20k, a real pipeline $25k-$60k.
Related
and I will tell you what it takes to make it survive.