AI use at managers has run ahead of governance. This piece sets out why AI governance now belongs in operational due diligence, what a reviewer is really assessing, and where regulators stand.
In brief
- The diligence question is no longer whether a manager uses AI. It is whether the manager governs it like any other operational process that can cause loss.
- The SEC’s fiscal 2027 examination priorities, published October 23, 2026, cover how firms monitor AI and whether controls match what investors were told.
- A reviewer is assessing five areas. Four apply to every manager; the fifth applies only where the manager builds its own models.
- No rule settles this for allocators. The FSB, IOSCO and US bank guidance are reference points, not requirements.
The diligence question about AI has changed. It is no longer “does this manager use AI.” It is “does this manager govern its use of AI the way it governs any other operational process that can cause loss.” Those are different questions, and the second is the one that tells you about risk.
Examiners are already asking the manager’s version of it. On October 23, the SEC’s Division of Examinations published its priorities for fiscal 2027. Examiners will review firms’ policies and procedures for monitoring AI, which the SEC now calls “super intelligence,” including its use in back-office operations. For automated investment tools, they will ask whether representations are fair and accurate, and whether operations and controls are consistent with the disclosures made to investors.
Not long ago, a manager’s use of AI was a curiosity. A reviewer might note it and move on. That window has closed. At many managers, AI is now part of how research is summarized, documents are drafted and investor questions are answered; the SEC’s own priorities name research, drafting, coding and client communications. With that adoption comes a set of risks that look familiar to anyone who has run operational due diligence: undocumented processes, data leaving the building, outputs no one can trace, and dependence on a provider the firm does not control.
I have seen operational due diligence from three seats: doing it hands-on at a fund of funds, advising allocators on how to build and run their ODD policy, and now helping design the AI workflows that put it to work. From each of them, this looks like a familiar problem arriving in a new place.
Adoption ran ahead of governance, quietly
At many firms, the tools arrived through individuals rather than through procurement, and increasingly as features inside software the firm already licenses. Nobody decided that the firm would adopt AI; it simply did.
In a manager, that gap shows up in recognizable ways. Confidential data typed into a consumer tool. An investor response drafted by a model and sent without review. A valuation input produced by a system that cannot be reconstructed after the fact.
None of these requires a sophisticated AI strategy to go wrong. Each requires only that nobody owns the question.
A manager using AI is not a red flag
Used well, AI makes a small team more capable and a process more consistent. The point of diligence is to distinguish a manager that has thought about control from one that has not, because the second kind owns a risk it may not know it is carrying.
Delegating work to a tool, or to a vendor’s tool, does not delegate accountability for the result. That holds for the manager. It holds equally for the allocator who selected the manager. In Canada, CAPSA’s risk management guideline for pension plan administrators makes the same point: an administrator that delegates a task retains its fiduciary duties.
The bar should scale with consequence. A tool that summarizes internal meeting notes does not need the controls of one that touches valuation, investor reporting, trading or regulatory records. Start where AI touches something an investor, auditor or regulator relies on, and expect the evidence to deepen from there.
What a reviewer is really assessing
In practice, four areas separate a governed operation from an exposed one, plus a fifth that applies only where the manager builds its own models.
- Governance and control. Is there a policy, an owner, and a clear line between where AI assists and where humans decide? Does what the firm says about its AI match what it does?
- Data and confidentiality. What client or position data reaches AI tools, on what terms, and what records do those tools create?
- Output integrity and oversight. Can the firm catch an AI error before it reaches an investor, reconstruct how a material output was produced, and keep working if a provider fails?
- Agents, access and AI security. Where AI acts rather than drafts, what can it reach, and can a wrong action be detected, stopped and reversed? Are AI-specific risks, such as instructions hidden in a document a tool reads, inside the firm’s security testing?
- Proprietary model development. If the manager builds or fine-tunes its own models, the bar rises to data provenance, validation and auditable evidence.
The aim is the discipline a cybersecurity or business-continuity review already brings, applied to a risk that barely registered in a diligence questionnaire two years ago.
So far, enforcement has turned on what firms say about AI to investors
The SEC’s first “AI washing” cases, against two advisers in March 2024, concerned what the firms said about their AI, not what their systems did. Examiners are watching the same gap: as the fiscal 2027 priorities put it, whether representations are fair and accurate, and whether operations and controls are consistent with disclosures made to investors.
For a reviewer, the test is simple. Compare what the DDQ and pitch book say about AI with what you hear in the meeting, and with what you see when you ask to be shown.
The gap can also sit around a tool rather than in it. In September 2026 the SEC settled charges with an adviser whose client-matching algorithm did not favor its own platform, but whose salespeople, working around the algorithm, steered clients in ways that created an undisclosed conflict.
There is no rule, but there is now a reference point
On October 15, 2026, the Financial Stability Board published twelve sound practices for the responsible adoption of AI by financial institutions. They cover governance and accountability, materiality and risk assessment, data, explainability, performance, human oversight, cyber risk and third-party dependencies, including concentration in the AI supply chain. The FSB is explicit that they are non-prescriptive and not a binding international standard, and Canada’s OSFI endorsed them the same day. For allocators, their practical value is a shared vocabulary.
For securities markets, IOSCO published its Supervisory Toolkit for AI Use in Capital Markets in May 2026. It is non-binding, written for securities regulators, and covers everything from traditional machine learning to generative and agentic AI. Its four areas of focus are governance and risk management, third-party and outsourcing risk, disclosure, and recordkeeping, and it includes example questions for supervisors to put to the firms they oversee.
In banking, the Federal Reserve, the OCC and the FDIC issued SR 26-2 on April 17, 2026, replacing SR 11-7 as interagency model risk management guidance. It placed generative and agentic AI outside its scope, describing those models as “novel and rapidly evolving,” while keeping traditional and non-generative AI models inside the framework, and left their governance to each institution’s own risk management practices.
None of this is a rule for an endowment, a foundation or most pension plans, and we would not suggest otherwise. But the signal travels. Model risk management is one of the most mature governance disciplines in finance, and when US bank supervisors looked directly at generative and agentic AI this year, they declined to bring it inside that framework.
Outside bank supervision the gap is wider, and principles rather than rules are now filling it. No certification settles the question.
The subject is new; the discipline is not
As AI becomes load-bearing in more of a manager’s operation, allocators should expect the same evidence they already ask for on cybersecurity and business continuity: not assurances, but artifacts, policies, logs and review records.
The managers that will diligence well are the ones treating governance as a design requirement now, rather than something to assemble under pressure later. The allocators that will review them well are the ones asking for that evidence before it becomes standard. That is the instinct good operational due diligence has always rewarded. The subject is new; the discipline is not.
Next in this series: five questions to ask in investment manager due diligence, one from each area, with what a strong and a weak answer sound like.
If you would like to see manager due diligence run to your own checklist, across the whole manager book, it is worth thirty minutes on one workflow. Nothing to prepare, nothing to sign.
Book thirty minutes on one workflow
Daryl Purdy is Head of Business Solutions, Americas, at Romina Day and holds the CAIA designation. Romina Day is the intelligence layer for investment operations: AI agent teams that run a firm’s most complex, repeatable work, configured to its methodology, with every figure traced to source and a named person signing off. The questions here apply to any manager, whatever tools they use, and we expect to answer them about our own work.
