There are two AI cost questions people routinely mix together: what it costs to build a feature, and what it costs to run it. This article is about the first; run economics get their own page. The honest anatomy of AI feature build cost, from a capability we built and reuse across products: about one week of architecture, about one week for the first working implementation, and four to five days for each subsequent product integration. One client integration of the same capability was scoped at roughly 40 hours. For the kind of business AI features we build, the API call is rarely the dominant engineering cost. The cost is the engineering around it, and the first version of anything costs multiples of every version after.
On this page
The anatomy, from a real build
We built a reusable AI assistant capability inside one of our field-operations products. Week one was architecture: how the AI would interpret requests, what would stay deterministic (data retrieval runs as structured queries with permission checks, never left to the model), how context would be scoped, and where the boundaries between interpretation and execution live. Week two produced the first working implementation. After that, integrating the same capability into another product runs four to five days, and a client-specific integration was scoped at about 40 hours. These are our numbers from our builds, not industry benchmarks, and the pattern is the whole lesson: architecture and the first implementation carry the cost; reuse is where AI features get cheap.
Where the money actually goes
Not the model. In the kinds of AI features we build, the engineering line items that can dominate the budget are:
- Data preparation: getting your actual records into a shape a model can use well, which for messy operational data can become one of the largest line items.
- Evaluation: testing outputs against real examples until “it seems to work” becomes “it works on our cases,” which is engineering, not vibes; when we added AI-driven capability to our recruitment platform’s core workflow, accuracy dipped with early changes and returned with iteration, and the iteration was budgeted work.
- Guardrails and boundaries: deciding what the AI may not do and enforcing it deterministically.
- Integration: wiring interpretation into workflows that already exist. The API usage itself, for business-volume features, runs tens of dollars a month; the guide to that arithmetic is its own page.
Inside the 40-hour integration quote
The client-specific integration scoped at roughly 40 hours is worth dissecting, because its shape is the reusable lesson. Roughly speaking, the hours went to: understanding the client’s data and workflow context (the largest share, and the part founders expect to be smallest), configuring the interpretation layer to their vocabulary and permissions, wiring outputs into their existing screens, and evaluation against their real cases before anything went live. What the quote did not include is equally instructive: no architecture week, because that was paid once when the capability was first built, and no model training, because none was needed. This is why the first AI feature and the fifth carry completely different prices, and why a vendor quoting your first feature at reuse prices is either reusing someone’s architecture or planning to skip the parts that make it work.
Retrofit vs greenfield, honestly
Adding AI to a live system costs more than building it into a new one, for unglamorous reasons: the live system’s data wasn’t shaped for models, its workflows have users who notice changes, and testing must not disturb production. We’ve done both, including AI document-processing added to a live transaction platform and an AI-modernized workflow inside our own recruitment product (that one: about fifteen days of planning, two months of staged execution on a running system). Retrofits are entirely feasible; they are simply projects, not features.
How to budget an AI feature without getting burned
Ask a vendor four questions. What’s the architecture week producing, in writing? What does evaluation against our real examples look like, and how many examples? Which parts stay deterministic? And what does the second integration cost versus the first? A vendor with real AI delivery history answers all four with numbers. A vendor selling API calls in a trench coat answers with adjectives.
FAQs
What does a typical AI feature cost?
For the class we build: a first capability carries architecture plus first implementation (roughly two weeks of engineering in our anatomy), subsequent integrations of the same capability run days, and a client-specific integration scoped at about 40 hours is a real reference point. Run economics are separate and small; see the private-AI guide.
Why do AI quotes differ so wildly?
Because the invisible line items (data preparation, evaluation against real cases, guardrails) are either priced or skipped. Skipped is cheaper until launch.
Does the model choice change the cost much?
Less than founders expect. The engineering around the model dominates; swapping models inside a well-built feature is days, which is also your insurance against vendor drift.
Work with us
