How to use this checklist
We quote fixed prices. One workflow, executable acceptance criteria signed on day one, a two-week build in your environment, and $10,000 due only after every criterion passes. You cannot quote that way on vibes. You quote it by asking the questions that expose where the real work hides, before you commit.
These are the 20 questions we ask in a scoping session, refined across more than 200 production deployments since 2014. They group into the six drivers we actually price on: systems to integrate, document classes, regulated data, teams and approval stages, edge cases, and scale. Each question comes with one line on why it moves the price, so you can see the quote forming as you answer.
A vendor who quotes without asking most of these questions is guessing. The guess does not disappear. It comes back later as a change order, a slipped date, or a pilot that never ships.
Systems to integrate (questions 1 to 4)
Integration, not the AI model, is where most of the engineering time goes. The count of systems matters less than how each one lets us in.
| Question | Why it moves the price |
|---|---|
| 1. Which systems does the workflow read from, and which does it write to? | Every system is a surface to connect, test, and keep stable. Reads are cheaper than writes, which carry rollback and data-integrity work. |
| 2. Does each system expose a supported API, or are we working with exports, file drops, or screen-level access? | A documented API is a known cost. A nightly CSV export or a UI-only system can double the integration effort for that connection. |
| 3. Who controls authentication for each system, and can a service account be granted within the first two days? | Access delays burn sprint days that cost the same as build days. A named owner who can grant access fast keeps the quote small. |
| 4. Are any of these systems heavily customized or running on-premises? | Custom objects, custom fields, and network boundaries turn standard connectors into bespoke work, and bespoke work is what you are paying for. |
Documents and data (questions 5 to 8)
AI workflows eat documents and data. The variety, not the volume, is what sets the build effort.
| Question | Why it moves the price |
|---|---|
| 5. How many distinct document classes does the workflow touch? | Each class, whether contract, invoice, slide deck, or form, needs its own extraction logic and its own slice of the test set. Classes are a bigger cost driver than page count. |
| 6. Are the documents digital-native, scanned, or mixed? | Scans bring OCR, layout variance, and quality triage. A workflow that is 10 percent scans still needs the full scan-handling path built. |
| 7. Does a reviewable sample set exist, say 50 real examples per document class? | If we can freeze a gold set on day one, acceptance is cheap to verify. If the gold set must be assembled and labeled first, that work lands in the quote. |
| 8. Where does the source data live today, and how clean is it? | Duplicates, stale records, and inconsistent fields must be handled before any model sees them. Cleanup is honest scope; pretending it is not is how projects slip. |
Regulated data and compliance (questions 9 to 12)
We build to SOC 2, HIPAA, and GDPR standards, and every implementation is aligned with the EU AI Act. The controls are routine for us, but they are still scope, and the answers here decide how much of that scope your workflow carries.
| Question | Why it moves the price |
|---|---|
| 9. Does the workflow touch personal data, patient data, or financial records? | Regulated data classes pull in minimization, retention, and audit-logging requirements that must be built and demonstrated, not just promised. |
| 10. Could the system's output later be read by a regulator, an auditor, or a court? | Evidence-grade output needs versioned rules, cited sources, and replayable verdict records. That is a deeper build than a system whose answers only inform a human. |
| 11. Are there data residency or tenancy constraints, such as data that cannot leave your region or your tenant? | Residency limits constrain model and infrastructure choices, and the constrained option is sometimes the costlier one to operate. |
| 12. Does an AI governance, legal, or works council review need to approve before go-live? | Approval cycles that cannot finish inside two weeks must be sequenced in parallel from day one, or the sprint ends with a system that passes and cannot launch. |
Teams and approval stages (questions 13 to 16)
Every team that touches the workflow is a source of requirements, and every approval stage is a screen someone has to use. People-shaped scope is the scope buyers most often forget to mention.
| Question | Why it moves the price |
|---|---|
| 13. Who owns this workflow today, and will that person sit in the scoping session? | Criteria signed by someone without authority get renegotiated mid-build, and renegotiation is the most expensive activity in software. |
| 14. How many teams handle the work between trigger and final output? | Each handoff hides rules nobody wrote down. Surfacing them in scoping is cheap; discovering them in week two is not. |
| 15. How many approval stages does the output pass through, and which ones need a human-in-the-loop screen? | Review queues and approval UIs are real engineering, frequently more of it than the model work itself. |
| 16. Who can sign acceptance criteria on day one? | Our model only works if pass-fail criteria are signed before the build. A missing signer adds calendar time, and calendar time is money on both sides. |
Edge cases and scale (questions 17 to 20)
The average case is usually easy. The price lives in the exceptions and in the volume the system must survive.
| Question | Why it moves the price |
|---|---|
| 17. What are the five ugliest real examples from the last quarter? | Edge cases set the true difficulty of the build. We price the exceptions, because the exceptions are why the workflow still needs humans today. |
| 18. What share of volume is allowed to fall to a human queue at launch? | A system that automates 80 percent with clean escalation is a two-week build. Chasing 99.5 percent is a different architecture and a multi-sprint program. |
| 19. What volume does the workflow run at, per day and at peak? | Throughput drives architecture, infrastructure sizing, and per-run model cost, which is an operating expense you should see in the quote, not after it. |
| 20. If this first workflow ships and passes, what is the next one? | The roadmap decides whether we build single-purpose or lay foundations the next sprint reuses. Larger programs run as repeated two-week sprints on the same pay-on-pass terms. |
What a complete answer buys you
With these 20 answers written down, a competent builder can hand you two documents in days: a fixed quote and a draft acceptance criteria sheet with pass-fail checks you could run yourself. If a vendor cannot produce both from your answers, the scoping is not done, or the vendor is pricing risk into a number they will defend later.
Scope one workflow. Get the criteria in writing. Pay when they pass. Everything in this checklist exists to make that deal quotable.