Two quotes for a similar brief may assume different acceptance, data, review and operating responsibilities. Compare those assumptions before comparing amounts; no market price or currency is established by this example.
Driver 1: the acceptance test
"It answers questions about our documents" and "it extracts these eleven fields with a measured accuracy we report to a client" are separated by evaluation infrastructure, a labelled test set, and a correction loop. Evaluation scope can materially change the estimate.
Driver 2: the regulatory surface
Access control, audit trails, residency, retention and evidence export can affect architecture. Their cost depends on requirements and the system; retrofitting after review may add significant work.
Driver 3: integration count
Each external system adds authentication, rate limits, failure modes, reconciliation and someone else's release schedule. Confirm vendor support, access, dependencies and recovery assumptions before estimating integration work.
Driver 4: who operates it
A platform that runs unattended needs monitoring, alerting, a rollback path and a person accountable when a run goes wrong. Confirm whether operations are included rather than assuming future support or costs.
Assess model usage alongside engineering and operations
Model usage, engineering, human review and operations all contribute to cost. Compare candidate models against measured quality, latency and usage; no cost share is assumed.
How to get comparable quotes
- State the acceptance test in measurable terms for each feature.
- Name the compliance frameworks you must satisfy within eighteen months, even if not today.
- List every system that must be integrated, with owner and access status.
- Say who operates the platform after launch, and for how long.
- Ask for a scoped proposal and separately agreed terms before work begins.
A quote without an acceptance test is not a quote. It is an estimate of how long someone is willing to be optimistic.