Guides · Vendor risk
Your AI vendor can rewrite the contract underneath your product with a web-page edit and an email you will not read. Most edits are cosmetic. A few change what happens to your users' data, what you owe your own customers, and whether your compliance story still holds. This guide is about telling those apart — and catching the real ones in time.
By Ingenious AI · Published 1 November 2026
AI vendors are pricing, and re-pricing, questions that did not exist five years ago: Is your input data used for training? How long are prompts retained for abuse monitoring? Which subprocessors touch the data, and in which countries? What happens to outputs you generated if the model is withdrawn? Regulators keep moving too — the EU AI Act's transparency duties alone forced a wave of documentation updates across the industry. The result is that AI terms and privacy policies change several times a year per vendor, often with the changes effective immediately or on a few weeks' notice. A vendor review done at procurement is a photograph of a moving target.
The highest-stakes sentence in any AI agreement: whether inputs and outputs may be used to train or improve the vendor's models, and whether the answer differs by tier (consumer vs API vs enterprise). A change here can silently void the promise you made your own customers. Watch for the wording shifting from "we do not train on your data" to "we may use de-identified or aggregated data" — that is a different promise wearing similar clothes.
How long prompts, files and outputs are stored, and for what stated purpose (abuse detection is the usual one). Retention changes interact directly with your own privacy policy and your data-processing agreements: if your policy tells your users "deleted within 30 days" and your vendor quietly moves to 90, your policy is now false. Retention is also where "zero retention" enterprise arrangements get quietly re-scoped to named services or regions.
New subprocessors mean new companies — sometimes in new jurisdictions — processing your users' data. For teams with GDPR or sector duties, a subprocessor added in a country your DPA doesn't cover is a compliance event, not an administrative detail. Vendors usually maintain a dedicated subprocessor page; it changes more often than the main terms and is the page nobody watches.
Deprecation schedules, default-model swaps and acceptable-use redefinitions often surface first in policy or documentation pages rather than in a changelog your engineers follow. An acceptable-use change can make a feature you already ship a violation. Pair terms monitoring with model deprecation watch; the two risks travel together.
Output-IP indemnities (who defends you if generated content infringes), liability caps, and the vendor's right to change pricing or rate limits on notice. These rarely change — which is exactly why a change here is always material.
Materiality is about consequence, not length. A two-word edit — "may retain" replacing "will delete" — can outrank a three-page reorganisation that changes nothing. When we triage a diff, we ask four questions: Does it change what happens to data? Does it change what we promised our customers? Does it change what the product is allowed to do? Does it change what it costs or who bears the risk? A "yes" to any one makes it material, and material changes deserve a named owner and a decision: accept, renegotiate, switch the feature flag, or switch the vendor.
Keep the diffs themselves. The pair of documents — before and after, dated — is your evidence that a promise changed and when. In a customer dispute or a regulator's questionnaire, "the vendor changed it in March, here is the text" is a complete answer; "we think they changed something" is not.
This is the workflow TermsWatch automates for our own stack: it tracks AI vendor terms and privacy policies continuously and flags the changes that affect your data and your contracts, so the weekly human job is reading a short list of material diffs instead of re-reading forty pages of legalese hoping to spot a moved comma. However you run it, run it as a system — the vendors are certainly changing their terms as one.