← Back

Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice

← All guides

What to Ask Your AI Vendor Before January 1, 2027: Twelve Questions With the Answers You Need in Writing

The short answer: Every question on this list exists because one of your duties needs the answer. Colorado's developer data names the core set — intended uses, categories of training data, known limitations, instructions for appropriate use and human review, notice of material updates — and the deployer's practical needs add data export, sub-processor transparency, per-file run logs, and the tool's exact trade name for the notices. Ask them before January 1, 2027 (and at every renewal after), ask them in writing, and score the answers against your duties — because the duty stays with you, the deployer, whatever the vendor says.

Check your duties and dates — free Build your first notice — free

What is this list for? It is the procurement instrument for the deployer position set out in Vendor vs Deployer and the documentation set detailed in Colorado Developer Duties. Each question states what you need, why (which duty consumes it), and what an adequate answer looks like — so a sales conversation can be graded against a compliance standard.

The documentation five (Colorado's developer set)

1. Intended uses — 'State the decisions and populations this tool is intended for, and the uses it is not intended for.' Adequate: a written statement with boundaries. Consumed by: your purpose statements and register. 2. Training-data categories — 'List the categories and sources of the training data, with coverage notes per population.' Adequate: categories with honest thin-coverage notes. Consumed by: California's risk-assessment data element, Connecticut's notice fields, your own evaluation.

3. Known limitations — 'What are the known limitations, and what should deployers do differently where each applies?' Adequate: operational limitations with deployer consequences. Consumed by: reviewer training and safeguards. 4. Appropriate-use and human-review instructions — 'Give us your numbered instructions for appropriate use and for how human review should be configured.' Adequate: testable instructions, versioned. 5. Material-update notice — 'Will you notify us of material updates in advance, with content and a pre-production window? Put it in the contract.' Adequate: a clause, not a promise.

The operations four (your record and notices)

6. Trade name — 'What is the tool's exact trade name for consumer-facing notices?' Connecticut's written notice requires it; every state's notice names the tool. Adequate: one exact string, confirmed. 7. Data export — 'Can we export all of our decision data — inputs, outputs, reasons data — at any time, in readable form, including at contract end?' Adequate: a tested export, not a datasheet claim. Consumed by: your 3-year record, which must outlive the contract.

8. Run logs — 'Can we obtain per-file run logs showing which files the tool processed and when?' Adequate: a log or API that reconciles against your opt-out register (the negative-proof requirement). 9. Sub-processors — 'Which embedded services or sub-processors touch our people's data, and will you notify changes?' Adequate: a maintained list with change notice. Consumed by: your inventory's data-flow pass, which must stay true over time.

Worked example: Crosswinds Property Group

Example — fictional business, real duty structure. Crosswinds, the fictional housing operator from the deployer guide, ran the twelve questions at its screening vendor's renewal. The scored outcome:

The honesty three (claims you will be tempted to rely on)

10. Human-oversight claims — 'Your materials describe human oversight. What exactly does the tool itself do, and what does it assume our reviewers do?' The answer separates the vendor's mechanism from your duty; no vendor claim excuses your notice duty. 11. Evaluation evidence — 'What accuracy and discrimination evaluation exists for uses like ours, and can we see the method and results?' Adequate: method-visible evidence you can weigh — remembering California's hiring exception requires your evaluation, which vendor evidence informs but never replaces.

12. Bias-testing support — 'What outcome data can we extract to run our own bias testing and monitoring on our population?' Connecticut's no-defence rule makes your own testing file the evidence that matters; a vendor who cannot feed it has capped your compliance. Adequate: exportable outcome data at the granularity your testing design needs, confirmed by a trial extraction.

Scoring the answers: the file, not the meeting

File answers per question with dates: document received (version), commitment made (contract clause), or gap noted with its consequence. Gaps are decisions: a missing limitations document goes in the register as a risk with a mitigation (your own sampling, enhanced review), and a missing export becomes a contract-end retrieval plan before it becomes a hostage negotiation. The register row for the tool should link the vendor file the way it links the checker result.

Re-ask on triggers: renewals (your leverage moment), material updates, new decision types, new states. The January 2027 deadline makes this autumn's round the founding one — the answers you file now become the baseline every later answer is compared against. Ask from the checker's documentation list in hand, and let the vendor's responsiveness to precise questions tell you what deployment will feel like.

If the answers do not come

A vendor who will not answer has answered. The deployer's options, in order of preference: constrain the use until the gap stops mattering (a genuinely minor role, evidenced), replace the tool for covered decisions, or stop the use — each recorded in the register with the vendor file attached. What is not on the list: deploying anyway on assurances. January's duties will be discharged — or not — by your notices, your record, and your reviewers; the vendor file exists so those are built on facts.

One encouragement to close: most vendors, asked precisely and early, can produce most of this — the documentation set is a stated duty on their side in Colorado's data, and your questions give their compliance team the deployer's shopping list they rarely receive in usable form. Ask well, file the answers, and the vendor relationship becomes part of the evidence rather than part of the risk.

  1. Ask the documentation five in writing — intended uses, training categories, known limitations, review instructions, update notice as a contract clause
  2. Ask the operations four — exact trade name, tested data export, per-file run logs, sub-processor list with change notice
  3. Ask the honesty three — decompose oversight claims, inspect evaluation methods, confirm outcome-data extraction for your own testing
  4. Score and file every answer (document, clause, or gap-with-mitigation) in the register beside the tool
  5. Re-ask at renewals, material updates, new decision types, and new states — and treat refusal as an answer with consequences

Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.

Build your first notice — free

Related guides

See all ADMT guides for the full set.

Frequently asked questions

Why must the answers be in writing?

Because your duties consume them as artefacts: the risk assessment cites the data categories, the notices carry the trade name, reviewer training uses the limitations and instructions, and the record relies on the export. A meeting answer cannot be filed; a document can.

Our vendor is small and informal. Is this list realistic?

The core five are the developer duties in Colorado's checker data — intended uses, training-data categories, known limitations, appropriate-use and human-review instructions, and material-update notice. A vendor selling into consequential decisions should expect exactly these questions.

What if the vendor's evaluation looks strong — do we still need our own?

Yes, where an exception turns on it: California's hiring and work-assignment exception requires the business to have evaluated the tool for accuracy and discrimination. Vendor evidence informs your evaluation; it does not replace it. Connecticut's rule makes the same point — your own testing file is the evidence; the vendor's certificate is an input.

When should we ask — contract, renewal, or now?

Now, and again at every renewal and material update. Renewals are the leverage moments; January 1, 2027 is the deadline that makes this round the founding one.

What if a vendor refuses the data export?

Treat it as a gating gap: your 3-year record duty outlives the contract. Secure the export before deployment, constrain the use, or choose a different tool — and record the finding and path in the register.

Build your first notice — free