Every person in your firm has a reporting line, a job description, permissions that match the role, and a review cycle. The AI your vendors have switched on inside your firm has none of that. The Vendor AI Oversight Review gives it that apparatus in three weeks, inside the governance you already run, and ends in an answer the board and the regulator can read and you can sign.
It is not confined to the back office. Your portfolio and order tools, the research platform that builds the case for an investment, the adviser tools drafting suitability notes, the CRM resolving client queries with no human involved, the administrator’s reconciliations and NAV checks, the transfer agent’s investor onboarding, KYC screening, document management, the HR and finance platforms. Each arrived as a feature in a release note. Each makes your firm the deployer in law the moment it is used on your clients, your investors or your staff, and you have no inventory of where it sits, no named owner for any of it, and a board that has started asking.
The vendor AI you approved followed rules. It did the same thing every time, and one review of the logic covered everything it would ever do. What your vendors ship now makes its own judgements, varies its output, drifts as models and data update, and increasingly acts on its own: moving records, sending communications, making the decision rather than drafting it. The upgrade from one to the other happens in a release, on the vendor’s schedule, with no visible act on your side. Sampling no longer protects you, because one model across an estate makes every decision the same way. Sign-off no longer protects you unless the person signing had the authority, competence and time to overrule, and nobody is measuring whether they did.
Some vendor AI is software you run. The CRM, the portfolio system, the HR platform: their agents act on your data, your clients and your staff under your licence, which makes you the deployer and makes those agents, in every sense a regulator cares about, members of your staff. The questions are what each one does, who in your firm owns it, what it is permitted to do, and how much harm it could cause if it went wrong.
Some vendor AI is inside a service you buy. If your administrator now uses agents to produce your NAV, the administrator is the deployer of its own tools and what you carry is the outsourcing obligation you have always carried. For a service, two things decide whether you are in control: transparency, meaning the vendor tells you when the way the service is produced has changed and can show you what the AI did; and a human in the loop where the risk requires one, with the authority to stop the process, evidenced rather than asserted. The agents a vendor hands you to work with them, the portal, the reconciliation feed, the query bot, belong in the first category. The review keeps the two apart, because a vendor’s attestation that covers one is routinely taken as covering both.
No rulebook is coming. The Central Bank of Ireland expects firms using third-party AI to have governance and controls in place now and is closing its review of administrator and depositary outsourcing this year. The FCA has said it will not write AI-specific rules and will hold firms to the frameworks they already answer to. Under the EU AI Act the obligations on a deployer were never deferred; deferred is not cancelled. And every administrator, transfer agent and platform contract due for renewal in the next eighteen months is the only moment you will have real power over the AI terms in it.
The clock starts when your files arrive, not when the engagement is signed. What you cannot produce is itself the first finding.
Your part. A scoping call, then a one-page list of exactly what I need and where it usually lives: outsourcing and DORA registers, contracts and service levels, existing due diligence, recent board and risk papers, and who holds which responsibility. Your team gathers it; the three weeks begin when it lands. Where a vendor contact will answer directly, a short request goes to them through you now.
Every in-scope vendor: software you run or service you buy, where the AI sits, whether it follows rules or makes judgements, what it touches, and a risk rating of low, medium or high for each agent based on what it can do and what it can reach. A named owner in your firm against every line.
Each use tested against one spine: EU AI Act deployer duties, the Central Bank’s outsourcing and fund management company guidance, DORA, FCA senior manager accountability, and your own three lines of defence. Interviews with the accountable executives and the vendor relationship owners. Vendor interviews only where you want them, with you in the room.
Findings written for a board, not a technologist, with the order to close them and the paper you table as your own.
Week one depends on your files; weeks two and three run on a standing method. The spine, the question set and the clause schedule come with me, and the work is establishing which lines apply to which of your vendors and what your evidence shows. That is what makes three weeks possible and the scope fixed.
A rated inventory of every vendor AI in scope, with a named owner, a declared purpose and a defined scope of authority against each line. A gap assessment against the spine. A documentation report that says, for each vendor AI, what is missing, what to ask the vendor for, and what you should have created yourselves as deployer. A vendor AI due diligence question set you reuse at every onboarding and renewal. The contract terms to ask for at each renewal, ranked by what matters for that vendor. The order to close the gaps over the next quarter, with an owner against each. And a board paper the accountable executive presents as their own.
It is not legal advice. It does not test the vendor’s models for bias or validate them; it establishes whether you can evidence that someone did. It does not implement the remediation, though I can. It is not a data protection impact assessment, though it tells you where one is needed. And it does not cover AI your firm builds itself. That includes the agents and custom GPTs your own staff have set up in ChatGPT, Claude or Copilot, and anything your data science team has built or tuned on a vendor’s foundation model. The model may be a vendor’s, but the agent is yours, and it belongs in the companion In-house AI Oversight Review, which applies the same method and the same three weeks. The two are designed to run together.
What is a Vendor AI Oversight Review?
A three-week, fixed-scope review that inventories every AI your vendors have switched on inside your firm, whether it is software you run or a service you buy, risk-rates each one, names an owner in your firm for it, tests it against the obligations you already carry, and hands you a board paper you can sign.
Why can’t we just rely on our vendor’s own AI assurance?
A vendor’s attestation covering the software you run does not automatically cover the services you buy from the same vendor, and the two are routinely conflated. Using AI in your firm makes you the deployer in law, and that accountability does not transfer to the vendor just because the model does.
Does the EU AI Act apply to AI we did not build ourselves?
Yes. Deployer obligations apply to any firm using AI on its clients, investors or staff, regardless of who built the model, and those obligations were not deferred by the 2026 Digital Omnibus.
Is this the same as an in-house AI oversight review?
No. This covers AI you buy through vendors and outsourced services. AI your own staff have built or tuned, including custom agents or GPTs, belongs in the companion In-house AI Oversight Review, which runs on the same method alongside this one.
Eighteen years on your side of the vendor relationship, then the other side of it.
At HSBC Alternative Investments I was accountable for the oversight of the fund administrators and transfer agents behind a business whose assets grew from $200 million to $31.6 billion. I built the systems that traded with them and reconciled to them, measured them against their service levels every month, and over eighteen years did it with zero FCA breaches. I then held C-suite seats at Backstop Solutions and, as an affiliate, at Leo Regtech, building and selling the platforms firms like yours buy, which is where you learn what a vendor volunteers, what it withholds, and what its assurance actually evidences. I wrote the seven principles for vendor selection and oversight in January 2024, before the EU AI Act was passed. I hold the IAPP Artificial Intelligence Governance Professional certification and use it to fit AI obligations into the structures your firm already trusts rather than build new ones beside them. No payment from any vendor on a purchase I advise on. Referral and advisory relationships with vendors are declared up front, and the full list is available on request.
Use the contact form, or message me on LinkedIn.