Your vendors added AI. Did anyone tell you?
The CRM you approved three years ago now drafts emails with a large language model. Your support desk summarises tickets with AI, and your analytics tool has a chat assistant that reads your dashboards. None of these arrived through a purchase order or a security review. They arrived in release notes.
That's the shift PwC describes in its piece on Responsible AI and third-party risk management: vendors are building AI into their products and services, "often without full visibility or the understanding of their customers." Most vendor programmes were built to vet new vendors. The bigger exposure now is the vendors you already trust.
The quiet shift: AI arrives through updates
When you onboard a new AI tool, it goes through intake, security review and legal. When an existing vendor adds AI to a product you already use, nothing triggers that process. The contract doesn't change, the invoice doesn't change, and the feature is often switched on by default. Your data may now be flowing to a model provider you've never heard of, under terms written before the feature existed.
This isn't vendors behaving badly. Shipping AI features is what their customers ask for. The problem is that nobody on the buying side is asked to look again.
Why your existing checks miss it
SOC 2 reports weren't written for it A SOC 2 report tells you a vendor's controls were operating over a period. It rarely says which models they use, what data those models see, or whether your data trains them.
Questionnaires predate the feature If your security questionnaire was written before 2023, it probably has no questions about AI at all. Even newer ones tend to ask "Do you use AI?" and stop there.
Contracts are silent, or worse Most older agreements say nothing about AI. Some grant the vendor broad rights to use "aggregated" or "usage" data to improve its services, which can quietly cover training models on your data.
Reviews happen once A vendor reviewed at onboarding is rarely reviewed again until something goes wrong. AI capabilities change with every release.
Six things to do this quarter
1. Find the AI you already have List the vendors that touch customer, employee or financial data and check their last year of release notes for AI features. Record, for each vendor, whether it uses AI and what the AI does. You can't govern what you haven't written down.
2. Ask whether your data trains their models This is the single most important question. Look for "service improvement", "aggregated data" and "product analytics" rights in the contract, check the admin console for a training opt-out, and get the answer in writing.
3. Put AI disclosure in the contract Require notice before the vendor adds or materially changes AI in the service, with a right to object. Pair it with a no-training clause and AI incident notification. There's sample wording below.
4. Ask AI-specific questions Send vendors a short questionnaire about their AI: models, data, human oversight, testing and incidents. Ten focused questions get better answers than a hundred generic ones. Ours is below.
5. Tier vendors by what the AI actually does "Uses AI" isn't a risk level. An assistant that drafts text for a person to approve is very different from a model that makes decisions about customers, or an agent that acts on its own. Weigh the AI's role, the sensitivity of the data it can reach, and whether it learns from your data.
6. Re-check every year, and at every renewal Confirm once a year that each AI vendor's answers still hold. Renewal is your best moment: it's when you have leverage to add the clauses above.
Sample clause: AI use disclosure
A starting point for your legal team, not legal advice. Adapt it to your agreement and jurisdiction:
"Supplier shall notify Customer in writing at least thirty (30) days before introducing, or materially changing, any artificial intelligence or machine learning functionality used to provide the Services or to process Customer Data. The notice shall describe the functionality, the Customer Data it processes and any third-party model providers involved. If Customer objects, the parties shall discuss the change in good faith, and if the objection is not resolved within thirty (30) days Customer may terminate the affected Services without penalty."
Pair it with a clause that stops the vendor using your data, or anything derived from it, to train or improve AI models without your written consent, and one requiring notice of AI incidents within a fixed time.
Ten questions to send your vendors
AI in the service Does your service use AI or machine learning, including AI used by your staff to deliver it?
What it does What does the AI do in the service you provide to us?
Models and hosting Which models or AI providers do you use, and where are they hosted?
Training Is our data used to train or improve AI models?
Data Which of our data does the AI process, and how long is it kept?
Autonomy Does the AI make decisions or take actions without a person approving them?
Oversight How do people review, correct or override the AI's output?
Testing How do you test the AI for accuracy, bias and security before and after changes?
Incidents How and how quickly will you tell us about AI incidents or material changes to the AI?
Assurance Do you hold ISO/IEC 42001 or another AI governance certification or audit?
How Granur helps
We built these steps into Granur so they happen as part of managing vendors, not as a separate project:
AI use on every vendor Each vendor's page records what its AI does, whether it trains on your data, whether the contract requires disclosure, and how much it decides on its own. Together these give the vendor an AI risk tier, and Trust status shows that tier for every vendor in one view.
The questionnaire, answered by the vendor Send the ten questions above through a secure upload link. The vendor answers without an account, and their answers fill in anything your team hasn't answered yet.
AI terms in the contract review Granur's AI contract review now reports whether the contract requires AI disclosure and whether it allows training on your data, and flags the gaps.
Clauses ready to use The clause library suggests five AI clauses (disclosure, no training on your data, incident notice, human oversight and information rights) that you can add in one click and edit.
Yearly re-attestation Every AI vendor gets a yearly reminder to confirm its answers still hold, and the renewal wizard asks whether new AI features have arrived before you renew.
The vendors most likely to surprise you are the ones you stopped looking at. Start with the five that hold your most sensitive data, and ask them the training question this week.