Grounding evidence
Whether the sources are owned, current, permissioned and testable—not which model is fashionable.
Generative AI planning
Estimate the path from a GenAI demo to a governed production system
A prompt that works in a demo is a small part of the project. This estimator sizes the data, evaluation, permission, integration and operating work around it.
What the estimate measures
Production conditions, not model fashion
The public result is driven by evidence that can be inspected before a project is signed.
Release gate
Whether task success, unsupported claims, latency and cost per completed task can block a bad release.
Action boundary
Answering, recommending and changing a customer record are different engineering and approval problems.
Product operations
Tenant isolation, support visibility, incident ownership and regression testing are counted as build work.
Worked examples
How the inputs move the starting point
These examples explain the rules. They are not invented client results.
| Situation | Likely first path | Why |
|---|---|---|
| Internal document assistant, curated corpus, named owner, no evaluation set | AI Pilot · $28K–60K | The retrieval path is clear; the highest risk is assembling and signing the release gate. |
| Existing SaaS feature, multi-tenant sources, held-out set, support workflow | Production Build · $75K–180K | Evidence exists and the remaining work is product integration, controls and operations. |
| Autonomous actions across six systems in a regulated workflow | Enterprise AI · $180K–500K+ | Permissions, validation, rollback and integration sign-off dominate the scope. |
Methodology v1.0.0
A versioned rule, open to inspection
The score has five dimensions: problem definition 20, data readiness 25, integration readiness 20, acceptance readiness 20 and operational ownership 15.
- 01The score rewards written evidence: a named workflow, representative data, specified interfaces, a held-out release gate and an operational owner.
- 02Model hosting economics are deliberately excluded. The result estimates the project required to make the application safe to release and possible to operate.
- 03A low score selects feasibility or a pilot even when the desired system is large, because multiplying an unproven architecture only multiplies uncertainty.
Limitations
What this result cannot establish
- The estimate is a planning band derived from the answers supplied. It is not a quote, warranty, or substitute for reviewing real data and interfaces.
- Third-party licences, cloud or hardware consumption, travel, and unusual certification work are outside the build band unless a proposal explicitly includes them.
- No model, accuracy, savings, or return claim is inferred from a questionnaire. Those require a feasibility test or measured baseline.
Questions
Before you use the range
No. Model choice belongs behind the evaluation. The estimator sizes the application, data, permission, integration and release work that remains whichever model clears the gate.
No. It is a build-budget band. Provider consumption is workload-specific and is measured separately once a representative task and traffic shape exist.
Yes. Actions, approvals and integration count raise the scope because every tool needs least-privilege access, failure handling, audit evidence and a tested stop condition.