"How much would it cost to add an AI feature to our service?" There's a reason this question is hard to answer with a number right away. Even the same "summary feature" can mean very different amounts of work depending on what data it reads, who checks the result on which screen, and how accurate it needs to be.
We covered model usage fees separately in our earlier post on estimating LLM API costs. This post covers the step before that: development cost, meaning the factors that set the amount on a quote. Once you know them, it's easier both to compare quotes and to narrow the scope so you can start small.
Why quotes differ from vendor to vendor
AI feature quotes rarely come out as "one AI feature = X." The code that calls an AI model isn't long in itself, but there's far more work around it.
- Before: what data to collect and in what form to hand it to the model
- After: where to show the result and how people check and correct it
- Around: how to catch wrong answers, and who re-checks things when the model changes
When vendors scope this before-and-after work differently, the price differs too. So it's worth checking first whether the scope is the same, before looking at the amount.
The four factors that set the price
| Factor | What to look at | When the price goes up |
|---|---|---|
| Data to connect | Where it reads from, and what | Multiple systems, inconsistent formats |
| Screens · workflow | Who uses the result, and where | Many review · edit · approval steps |
| Accuracy bar | How right it must be to "pass" | An eval set has to be built from scratch |
| Operations | Who watches it after launch | Monitoring and model swaps included |
Data to connect. Reading one well-organized database is a different amount of work from reading spreadsheets, email, and external APIs together. Shaping the data into a form the model can read well also falls under this.
Screens · workflow. Some features are done once the result appears behind a single button; others only save after a person edits and approves it. The more human review steps there are, the more screen and permission design is needed.
Accuracy bar. Anthropic's official documentation advises defining success criteria that are "specific and measurable" and evaluating with test cases that resemble real work. OpenAI's official guide likewise describes evals as tests that check model outputs against criteria you've defined. The time it takes to build this eval set goes into the quote.
Operations. AI usage fees are charged per token (a chunk of text) for what you use. Anthropic prices input and output tokens separately per million tokens, and gives a 50% discount on batch requests processed in bulk. Separately from these usage fees, you need to decide whether the quote includes watching answer quality after launch and re-evaluating when you switch to a new model.
Want to check this against your own data?AI Readiness Assessment · we review your material and map candidates and orderContact usStart small with base + options
If you scope all four factors large from the start, both cost and timeline grow. When we add AI features to custom development, we talk about scope in two parts: base and options.
- Base: one data source, one screen, one human review step, a small eval set
- Options: more data connections, approval steps, admin statistics, re-evaluation on model swaps
Using the base version first shows you "where it's worth spending more." Options chosen after real use waste less than options decided before anyone has tried it.
Five questions to ask when comparing quotes
- What data, and from how many sources, is it scoped to read?
- Does it include a screen where people check and correct results?
- Who defines what "working well" means and writes the eval cases?
- Are AI usage fees included in the quote, or separate?
- After launch, where does re-evaluation go when the model changes?
Ask every vendor these same five questions, and most of the reasons the prices differ will come to light.
Summary
The cost of developing an AI feature is set less by the model name than by four things: data, screens, the accuracy bar, and operations. Split the scope into base and options, and put the same five questions to every quote. Comparing gets much easier.
References
- Anthropic, Define success criteria and build evaluations: https://platform.claude.com/docs/en/test-and-evaluate/develop-tests
- OpenAI, Working with evals: https://developers.openai.com/api/docs/guides/evals
- Anthropic, Pricing: https://platform.claude.com/docs/en/about-claude/pricing
In a one-hour AX consultation we map AI candidates and similar cases for you.
We build web apps, mobile apps and business systems, then add and run the AI they need. We build and operate our own meeting notes, contract and company analysis AI.
See our projects