14 questions to ask an AI vendor before you sign.

"Which models are actually under the hood — yours, or someone else's — and will you name them in the contract?"

Why it matters: most AI products are an interface over a third-party model, and every downstream obligation depends on knowing whose. A vendor who won't name the model provider in the contract is asking you to inherit obligations from a party you can't identify. Expect a deflection: "that's proprietary architecture." The counter-move: you don't need the architecture — you need the legal identity of each model provider, as a named subprocessor, and notice when it changes.

"Will you state in writing who builds, who modifies, and who deploys — the facts our counsel needs for the role analysis?"

Why it matters: the EU splits duties between providers and deployers; Texas and Colorado write obligations for developers and deployers by name. The same product can put you in different roles in different jurisdictions. Expect a deflection: "we're a software-as-a-service provider; our obligations are defined in the agreement." A vendor will rarely bind itself to a statutory role label — so don't ask for the legal conclusion. Ask for the facts: who develops and substantially modifies the system, who places it on which markets, who controls the deployment. Your counsel makes the role call from that record.

Reg. (EU) 2024/1689, Art. 3(3), 3(4) · Texas HB 149 · Colorado SB 26-189 (developers & deployers)

"Do you train on our data — and where is that stated in a document that can't change without notice?"

Why it matters: the answer buried in a settings toggle or an updateable policy page is not an answer; it's a variable. Get the commitment into the agreement itself — not a hyperlinked policy that updates silently: what is used for training, what is retained, for how long, and what notice you receive if any of it changes. Timing warning: this clause is negotiable before signature and nearly immovable after.

"What training-data disclosures have you — or your model provider — actually published?"

Why it matters: EU-market general-purpose model providers must publish a training-content summary on the Commission's mandatory template — a duty in force since August 2025, with the Commission's power to fine up to €15M or 3% of worldwide turnover activating 2 August 2026. California requires training-data documentation from generative-AI providers, in force since January 2026. These are public documents. A vendor who can't point you to theirs — or their model provider's — is telling you something about the documentation you'll get when you need it.

Reg. (EU) 2024/1689, Art. 53(1)(d), Art. 101 · California AB 2013

"If your product talks to our customers, whose job is the AI disclosure — yours in the interface, or ours in deployment?"

Why it matters: the EU requires that people interacting with an AI system be informed; Korea requires advance notice of AI use. Disclosure can live in the vendor's interface, your deployment layer, or fall between the two — and "falls between the two" is the default unless the contract says otherwise. Whichever side carries it, the other should warrant not to break it.

Reg. (EU) 2024/1689, Art. 50(1) — applies 2 Aug 2026 · Korea AI Basic Act, Art. 31

"Do your outputs carry machine-readable provenance marks — which standard, and will your engineers answer that in writing?"

Why it matters: the EU's machine-readable marking duty and California's watermarking provisions both become operative 2 August 2026. "We support content provenance" is a roadmap sentence, and the account executive on the demo call can't inspect metadata live anyway — so don't ask them to. Put the requirement in the technical questionnaire: which standard (C2PA, IPTC, watermarking), on which output types, and a sample output with its provenance metadata intact, delivered in writing by their engineering team. Either the marks are there or they aren't.

Reg. (EU) 2024/1689, Art. 50(2) — Dec 2026 deferral for pre-existing EU-market generative systems only · California SB 942

"Which certifications do you hold — scope, certificate number, and which public registry confirms it?"

Why it matters: an ISO/IEC 42001 certificate covers a defined management-system scope, which may or may not include the product you're buying. A SOC 2 report covers the controls it was scoped to cover — it is not an AI-governance credential by default. Scope statement and registry entry are checkable facts; a badge on a website is neither. Expect a deflection: a trust-center link. The counter-move: ask for the certificate's scope statement itself.

"What documentation will you hand us — the package the law already obliges developers to provide?"

Why it matters: Colorado's rewritten act requires developers to give deployers understandable technical documentation before the system is used — intended uses, training-data categories, known limitations, and human-review instructions — plus what a deployer needs for its own disclosure duties. EU general-purpose model providers owe downstream integrators an information package. If a statute already obliges your vendor to produce this bundle for someone, the question is whether you get it too — as a contract deliverable with contents and a delivery date, not a sales deck.

Colorado SB 26-189 — eff. 1 Jan 2027, dates may move · Reg. (EU) 2024/1689, Art. 53(1)(b)

"What can change without telling us — models, weights, subprocessors — and what counts as a 'material update' you must notify?"

Why it matters: Colorado's rewritten act requires developers to notify deployers of material updates and substantial modifications within a reasonable time. Beyond any statute: a silent model swap changes your risk profile, your documentation's accuracy, and possibly the role analysis from Q.02. Define "material" in the contract; the vendor's default definition is "whatever we didn't tell you about."

Colorado SB 26-189 (material-update notice)

"When the compliance requirements change, who pays for the platform changes?"

Why it matters: the rules under this product are still moving — Colorado's framework takes effect in 2027, Article 50 guidelines are still being finalized, and the state map grows every session. When a regulator's next rule requires a new disclosure surface or a new export, is that a compliance update included in your subscription, or a "premium enterprise feature" priced at your next renewal? A commitment that regulatory-compliance updates are included in the base subscription is cheap to ask for now and expensive to discover missing later.

"What counts as an incident, when do we hear about it, and what do you owe us then?"

Why it matters: your insurer, your customers, and your own record all depend on what your vendor's notification commitment was — and the vendor's standard paper defines the trigger narrowly. Expect the fight over one word: their draft will say a "confirmed, material breach"; the scope of what counts as an incident, and how fast notice arrives, is precisely the clause your counsel redlines. This guide's only insistence: the definition, the timeline, and the remediation owed must exist in writing before signature — not be discovered during the incident.

"Where does your paper put liability for all of the above — caps, carve-outs, IP and indemnity?"

Why it matters: this is the question the previous eleven exist to inform — and it is a question for your counsel, not an answer this guide provides. The liability allocation in a vendor's first draft is a starting position; read where the caps sit, what the indemnities actually cover, and which exposures are carved out, then let counsel redline from there. Timing warning: raised at diligence, this is leverage; raised at renewal, it's an admission.

"Which of these risks does your own insurance actually cover — and does your E&O policy now carry an AI exclusion?"

Why it matters: an indemnity is worth the balance sheet — or the insurance — behind it. Carriers have attached AI exclusions to commercial policies at renewal since January 2026, including professional lines. A vendor whose own errors-and-omissions coverage excludes AI risk is offering you an indemnity their insurer may not stand behind. Their answer belongs in your counsel's file and your broker's.

ISO endorsements CG 40 47 / CG 40 48 / CG 35 08 · public carrier filings — see the record at certian.com

"What do we get at exit — our data, our fine-tunes, our logs, and the records we'd need to show a regulator two years from now?"

Why it matters: disclosure and documentation duties outlive the subscription. If a regulator asks in 2028 about a decision the system touched in 2026, the records either left with the vendor or they didn't. Exit terms are the cheapest clause to negotiate at signature and the most expensive to negotiate at termination.