loading…
AI adds two workstreams that ordinary application estimates often omit: controlling uncertain behavior and proving that the behavior is good enough.
AI System Complexity ≈
Software Complexity + Model Uncertainty + Evaluation BurdenThe first model call may take an afternoon. The project is the rest: data access, prompt and tool boundaries, evaluation, human fallback, tracing, cost limits, failure handling, and the process for improving bad cases.
If payment succeeds, mark the booking paid. The same input should produce the same result.
Decide whether an interview answer is strong. There may be several reasonable judgments.
| Version | Included work | Typical scope |
|---|---|---|
| Prototype | Chat UI, one prompt, fixed interview flow, final feedback | A few days |
| MVP | Auth, session state, multiple modes, RAG, rubric, structured report, history | Several weeks |
| Production | Eval set, tracing, prompt versions, regression tests, cost controls, fallback, analytics | One to several months |
Retrieval quality depends on the source material and the ingestion pipeline. Budget time for document cleanup, access rules, metadata, chunking, filters, freshness, citations, and retrieval tests. If nobody owns updates, the system will quietly become stale.
Before the pilot, agree on representative cases and expected behavior. Include clean examples, ambiguous requests, missing data, unsafe requests, and cases that should escalate. After launch, every important failure should become a regression case.
AI quality improves through an evaluation loop, not a one-time prompt.
What belongs in the estimate “The model can do it” is the beginning of the AI project, not the end.