All services Every engagement has a fixed, published price. Book a strategy call
Most of your AI estate belongs to other people. The five-step method for assessing supplier AI risk: find the silent AI, tier by what could break, verify claims instead of collecting them, fix the contract, and re-check on a cadence.
Most organisations discover an uncomfortable fact when they finally build an AI inventory: the majority of their AI estate belongs to other people. It arrived inside the HR platform, the CRM, the service desk and the finance tool, switched on by vendors through routine updates, with no procurement event and no approval. Your customers, your regulator and your insurer do not care whose model it is. If it makes a decision about your data or your people, it is your risk wearing someone else’s badge.
This is the method we use to assess supplier AI risk for clients, published in full. It is built on one principle that separates it from the questionnaire industry: claims are verified rather than collected.
The short answer. Assessing supplier AI risk means finding every place a supplier’s AI touches your operations, data or customers, tiering suppliers by what they could break rather than what they cost, verifying their claims against evidence instead of questionnaire answers, fixing the contract so AI change requires notice, and re-checking on a cadence. Five steps, and the first is an inventory rather than a questionnaire.
Ordinary supplier cyber risk moves slowly: a vendor’s security posture this quarter looks like last quarter’s. Supplier AI risk does not behave. It arrives silently: AI features appear inside products you already licence, enabled by default in a release you did not read. It changes without changing: the vendor swaps or retrains a model and the feature behaves differently, with no new version number on your invoice. Your data may be feeding it: unless the contract says otherwise, some vendors treat customer data as training material. Accountability does not transfer: under the EU AI Act, an organisation using an AI system under its own authority carries deployer obligations even though the vendor built it, and your customers will hold you responsible either way. And you cannot see inside: when a vendor’s model scores, screens or decides something about your customers or staff, its reasoning is invisible to you, but the outcome carries your name.
Before any questionnaire, build the supplier AI inventory. Ask every material supplier two questions in writing: what AI or machine learning does your product use today, and which AI features have you added in the past twelve months. Then check their release notes yourself, because the answers are frequently incomplete. Every finding goes into your AI register alongside the systems you built, which is where AI governance and supplier assurance meet. You cannot questionnaire what you have not found.
Spend is the wrong sorting key. A cheap tool touching payroll data outranks an expensive one that formats slides. Tier by four questions: how sensitive is the data their AI touches, does it decide or score anything about people, how dependent are your operations on it, and would its failure be visible to your customers. The top tier earns full verification, the middle tier a focused review, the bottom tier contractual basics and a watching brief. Proportionality is what makes the whole method affordable.
A questionnaire answered by a supplier’s sales team is marketing with checkboxes. Verification is different: certificates are checked at source against the issuing registers, the “we never train on customer data&rdquo. Claim is tested against the data processing agreement rather than the brochure, sub-processors are named and looked up, and the “human in the loop&rdquo. Claim is probed for whether the human can intervene or merely observe. Most of what suppliers claim is checkable in public registers and contract schedules. The gap between claimed and checkable is precisely where the risk lives.
Four clauses to ask your lawyers for: advance notice of material AI changes and new AI features, with the right to switch them off. No use of your data for model training without written consent. Incident notification that explicitly includes AI failures and model misbehaviour as well as breaches, and evidence rights, so the claims you verified once can be re-verified. Renewals are the moment of leverage. A supplier who resists all four is telling you your tier one assessment was correct.
Supplier AI assurance is point in time, and the estate changes with every vendor release. Top tier suppliers get a full annual re-verification. Every tier gets a monthly release note sweep, the same rhythm an AI governance office runs for internal systems, and any material change, a new model, a new AI feature, an incident, triggers reassessment ahead of schedule.
| Ask the supplier | Why it matters | The weak answer sounds like |
|---|---|---|
| Is our data used to train your models? | Your confidential data may be improving a product sold to your competitors. | “Only to improve the service.” |
| What AI features have you added in the last year? | Finds the silent AI that arrived without a procurement event. | A pause, then a list shorter than their release notes. |
| Can we switch AI features off? | Without an off switch, their roadmap is your risk appetite. | “It is core to the product now.” |
| Who are your AI sub-processors? | Your data may be flowing to model providers you have never assessed. | “We use leading providers.” |
| What happens when your model is wrong about one of our customers? | Tests whether a correction route exists at all. | “Accuracy is very high.” |
| Which of your certifications cover the AI itself? | ISO 27001 covers information security. Only ISO/IEC 42001 covers the AI management system. | A security certificate offered as an AI answer. |
The method is published because a capable procurement and security team can run it. If you want it done for you, done fast, or done independently, the engagements are priced: a Supplier AI &. Cyber Risk Assessment tiers your critical suppliers and verifies their claims from £8,500 + VAT. A single Supplier Assurance Review checks one supplier properly at £4,500 + VAT in five to ten working days. A Supplier Assurance Framework builds the tiering, question sets and evidence standards as your own capability for £14,500 + VAT in four to six weeks, and the Supplier Assurance Office runs the whole function monthly from £3,500 + VAT. If your team has the headroom and this article in hand, build it yourself. The method does not mind who runs it.
Supplier AI risk is the exposure created when a vendor’s artificial intelligence touches your data, decisions or operations: models embedded in software you licence, AI features switched on through updates, and vendor sub-processors handling your information. The defining feature is that the risk changes without your involvement, because the supplier controls the model and its updates.
No. Questionnaires collect claims. Assessment verifies them. Certificates should be checked against issuing registers, data training claims against the data processing agreement, and sub-processor lists against reality. Questionnaires are useful as a starting map, never as evidence.
The ones that could break the most: suppliers whose AI touches sensitive data, decides or scores anything about people, underpins daily operations, or fails visibly in front of your customers. Rank by potential harm rather than by contract value.
In many cases yes. An organisation that uses an AI system under its own authority is a deployer under the Act, with obligations of its own, even though the vendor built the system. Buying AI does not outsource the accountability for using it.
Fully, once a year for your top tier. By release note sweep, monthly across the estate, and immediately on material change such as a new model, a new AI feature or an incident. Point in time assurance decays, because the supplier’s roadmap keeps moving after your assessment stops.
Emmanuel Olatunji
Chief Enterprise Architecture Partner
Emmanuel founded Intelance in 2017 and has led architecture, AI governance, cyber and deal engagements across the UK, Europe, the Middle East and Africa. Every Intelance engagement runs under his direction, and every guide he publishes comes from that work.
See his record on LinkedIn About IntelanceThirty minutes with a licensed assessor. You leave knowing which route applies to you, what it costs and what to do first.
Talk to a senior adviser
