The quiet arrival
AI is already in your pension scheme. Not because anyone decided to deploy it, but because it arrived through the devices and habits of the people who do the work.
A trustee using ChatGPT to summarise a 200-page meeting pack on the train. An adviser drafting correspondence with Copilot. An administrator's fraud detection model flagging exceptions. A member-facing chatbot answering retirement questions.
Some of it is procured and supervised by the board. Much of it is not.
This is the observation Amédée Levillain, CEO of Knowa, made in a recent post that has resonated across the pension governance community. See what he's saying about it on LinkedIn.
What The Pensions Regulator expects
TPR's AI plan, published recently, is direct: trustees remain accountable for decisions and outcomes even where activities are delegated. The board is expected to understand where AI is being used, establish governance, assure itself the same standards apply across administrators and advisers, and test the risks on a continuing basis.
That duty does not change because the AI is convenient, free, or familiar.
The regulatory position is consistent with the broader direction of travel in UK pensions governance. The General Code already places significant expectations on governing bodies around systems of governance, internal controls and knowledge and understanding. AI does not create a new category of obligation. It intensifies the existing ones, because the decisions being influenced by AI tools are the same decisions the board is legally responsible for.
The governance gap
The challenge for most trustee boards is not that they object to governing AI. It is that the AI arrived before the governance did. Consumer tools are ambient. They sit in the browser, in the email client, in the document editor. They are used instinctively by professionals who are trying to do their jobs well under time pressure. The result is that scheme-specific data (confidential meeting packs, investment recommendations, actuarial papers) ends up being processed by tools whose data-handling policies were written for general consumer use, not for fiduciary records subject to data protection, professional privilege and regulatory scrutiny.
A purpose-built platform for trustee work addresses this by design: scheme-level permissions, full audit trails, citations back to source, and a controlled environment that confidential papers do not leave. That is the precondition for AI being trusted with fiduciary material at all.
Nine questions for trustee boards
In response to this landscape, Amédée has launched a series on LinkedIn: nine questions a trustee board can ask whenever AI is pitched, proposed, or already in use in connection with the scheme.
They are an initial governance screen. Not a substitute for procurement or legal review. Short enough to ask out loud in a board meeting. Each designed to produce an answer that is either useful or revealing. Either way, the board learns something it needs to know.
The series is ongoing. Follow Amédée on LinkedIn for the full set of nine questions as they are published.
The duty doesn't change because the AI is convenient, free, or familiar.