An estimate covers the interface and AI prompt, then adds a small line called integration. That line may contain the work that determines whether a customer's booking, payment or document update actually succeeds.
For founders reviewing engineering proposals, ask what will be tested across system boundaries. An estimate that omits this work can make a project appear cheaper by leaving acceptance unresolved.
What is integration testing in an AI product estimate?
Integration testing checks how connected components behave together. It covers request formats, authorisation, remote state, callbacks and recovery, rather than only whether an individual function returns the expected value.
The scope depends on the workflow. A document assistant needs different fixtures from an appointment receptionist. Ask the engineer to trace the customer outcome and identify each dependency that can alter it.
Turn a vague integration allowance into inspectable work
Use a proposed appointment workflow to discuss the estimate. The following review matrix is illustrative, not a claim that every product needs the same testing budget.
| Boundary | Scenario to budget | Evidence of completion |
|---|---|---|
| Identity and account | Caller or user belongs to wrong account | Request rejected without data exposure |
| Remote write | Response is lost after save | State reconciled without duplicate action |
| Callback delivery | Event arrives twice | One intended business outcome |
| Data mapping | Customer corrects a field | Final verified value appears remotely |
| Operating environment | Credential or route is unavailable | Owned failure and recovery path |
For each row, ask who supplies access, creates fixtures, checks remote records and approves the result. These responsibilities consume time even when the underlying API call is short.
Separate unknowns from omitted work
A responsible early estimate may include uncertainty because the developer has not inspected your integration access or existing code. Ask for assumptions and a discovery step instead of demanding false precision.
An omitted category is different: the proposal might assume that a successful request means the complete business process works. Request an explicit acceptance scenario and inspect whether its setup, execution and recovery fit within the quoted work.
Use the paid engineering trial framework when one narrow integration experiment can reduce a major estimate uncertainty. Keep that trial bounded so it does not become an unpaid product implementation.
Compare proposals against the same completion boundary
Give each candidate the same customer journey, dependencies and review expectations. One proposal may include a deployable pilot, while another includes only a local prototype. Comparing the totals without that distinction produces an unreliable decision.
The cheapest useful milestone may deliberately leave some steps manual. That is acceptable when the founder understands the operating cost and the estimate includes the handoff needed to perform those steps safely.
Does every integration need a complete automated test suite?
Choose coverage based on impact and repeatability. Some external behaviour still needs an authorised manual verification procedure.
Can a fixed estimate include unknown APIs?
It can state assumptions, exclusions or a discovery milestone, but those boundaries should be visible before approval.
What should we ask the engineer to revise?
Request named integration scenarios and acceptance evidence in the external technical brief.