The Pensions Regulator's new AI guidance tells trustees it creates no new rules for a sector worth 1.8 trillion pounds. It then sets ten mandatory-sounding questions, gated by a word the guidance itself never defines.
Estimated reading time: 7 minutes
In short. The Pensions Regulator published guidance on 29 October 2026 telling pension trustees how to govern their use of AI [1]. The guidance says plainly that it "is not intended to establish new and separate regulatory requirements" [1]. It then says every trustee "should be able to answer" ten specific questions "for each significant use of AI" [1]. The word doing all the work, "significant", is never defined, only illustrated with examples the guidance itself calls "non-exhaustive and illustrative only" [1]. The guidance sets no review date for itself, while telling trustees to keep their own approach under review. Two documented mistakes, four stars.
TPR regulates a sector it describes, in its own words, as covering "around 22 million" people saving into a workplace pension with "£1.8 trillion of assets under management" [2]. On 29 October 2026 it published its first dedicated guidance on how trustees should govern AI use across that sector, alongside a press release in which its Executive Director of Market Oversight, Ben Gunnee, said trustees "don't need to be AI experts, but they are accountable" [1] [3].
The guidance opens, under the heading "About this guidance", with a clear boundary on what it is: "It is not intended to establish new and separate regulatory requirements, but to help you apply existing regulatory expectations to AI use in a practical and proportionate way" [1]. That framing matters, because it is also how TPR avoids the slower, more accountable route of a statutory code of practice.
The checklist the "no new requirements" line does not fit
A few paragraphs later, the same document states: "Every trustee or scheme manager should be able to answer the following questions for each significant use of AI" [1]. Ten questions follow: what AI is used, why, what member outcome or process it improves, what decisions it may influence, what data it uses and how that data is handled, who is accountable, how it is tested, how a decision can be challenged by users or members, how humans can override it, and what happens if the AI fails or becomes unavailable [1]. That is a specific, itemised documentation and governance exercise, triggered by a specific condition, AI use that is "significant". Nothing like it exists in the statutory duty the guidance says it is merely helping trustees apply.
That statutory duty is trustees' knowledge and understanding, set out in sections 247 to 249 of the Pensions Act 2004. Section 247 requires an individual trustee to "have knowledge and understanding of... the law relating to pensions and trusts" and "the principles relating to... the funding of occupational pension schemes, and... investment of the assets of such schemes", to a degree "appropriate for the purposes of enabling the individual properly to exercise his functions as trustee" [4]. TPR's own guidance page on this duty describes it the same way: trustees must have general knowledge of pensions law, trust law and funding and investment principles, and be "conversant with" their scheme's own policy documents [5]. That page contains no AI-specific checklist, no ten-question test, and no requirement to document an answer for each "significant" use of any particular technology [5]. The AI guidance's own ten questions, and the specific trigger that activates them, are not in that statute or in TPR's existing account of what it requires. They appear for the first time in this document.
A trigger word the guidance cannot pin down
The guidance uses the word "significant" five separate times, and never converts it into a rule a trustee could apply in advance [1]. It asks trustees to "recognise when it may have significant consequences for the scheme or members", to explain their reasoning "where AI has significant influence on a decision", and to "establish clear governance and accountability for significant AI use" [1]. The closest it comes to a definition is three worked examples, graded low, medium and high risk: an internal tool drafting documents and summarising meetings (low), AI producing first-draft benefit specifications from scheme rules and deeds (medium), and an AI customer service system handling highly sensitive member data around the clock (high) [1]. TPR labels these examples "non-exhaustive and illustrative only" and tells trustees to "consider risks within the context of your scheme" [1]. A trustee board weighing up a mid-range use, an AI tool that flags cases for human review rather than drafting documents or running member-facing chat, cannot read the threshold off the document; it has to guess where its own use sits between three illustrations the regulator says are not exhaustive.
No date set for its own review
The guidance closes its framing section with a commitment that runs one way. It tells trustees: "As AI technologies, uses, and regulatory frameworks continue to develop, trustees should keep their approach under review" [1]. TPR's own undertaking is softer: "We will continue to monitor developments and engage with relevant regulators and organisations to help identify emerging risks, share good practice and support a co-ordinated approach where appropriate" [1]. The page carries no version number, no "last reviewed" date distinct from its 29 October 2026 publication date, and no stated date by which TPR will revisit it [1]. The press release accompanying the guidance names no monitoring or compliance-checking mechanism either; Gunnee's quoted line is about trustee accountability, not about how or when TPR will check whether the guidance is working [3].
The claims, tested
| TPR's own words | What the record shows | Verdict |
|---|---|---|
| The guidance "is not intended to establish new and separate regulatory requirements, but to help you apply existing regulatory expectations" [1] | The statutory trustee knowledge and understanding duty it points to (Pensions Act 2004, ss.247-249) requires general knowledge of pensions and trust law, funding and investment principles [4]; TPR's own page on that duty contains no AI-specific checklist [5]. The ten-question test is new | A specific, itemised obligation sits inside a document that says it adds nothing new |
| "Every trustee or scheme manager should be able to answer the following questions for each significant use of AI" [1] | "Significant" is used five times and illustrated only by three examples the guidance itself calls "non-exhaustive and illustrative only" [1] | A trustee cannot tell from the document alone when the obligation applies |
| "Trustees should keep their approach under review" [1] | TPR commits only to "continue to monitor developments"; no review date, version number or enforcement mechanism appears on the guidance page or in its press release [1] [3] | The review duty is put on trustees; TPR's own review commitment to the document is unquantified |
Credit where due
TPR has not tried to hide the limits of its own document. It says outright, in its own words, that the guidance creates no new regulatory requirements, rather than letting trustees assume otherwise [1]. The risk-graded examples, low, medium and high, are a genuine attempt to give trustees something concrete to reason from, and the guidance is explicit that they are illustrative rather than a closed list, which is honest about their own limits even if it leaves the central term undefined [1]. Publishing detailed guidance on a fast-moving technology within a year of it becoming a board-level issue for trustees, rather than waiting for a slower statutory process, is itself a defensible choice; the gap here is in how the document's own disclaimer sits against its own checklist, not in the decision to publish early.
Verdict
Four stars, two documented mistakes. The Pensions Regulator's new AI guidance tells trustees, in its own words, that it creates no new regulatory requirements. It then sets a specific, ten-question governance exercise, gated by a word, "significant", that the guidance illustrates but never defines. And it asks trustees to keep their own approach under continuous review while committing itself to nothing more dated than "we will continue to monitor developments". For a document covering a sector TPR itself values at £1.8 trillion, covering 22 million savers, that is two gaps a trustee board would have to fill in by itself.
The star score counts two documented mistakes. First: the guidance's ten-question checklist for "each significant use of AI" has no counterpart in the statutory trustee knowledge and understanding duty (Pensions Act 2004, ss.247-249) it says it is merely helping trustees apply, and the word "significant" that triggers it is illustrated five times but never defined, only given three examples the guidance itself calls non-exhaustive. Second: the guidance asks trustees to keep their approach under continuous review but sets no review date, version number or enforcement mechanism for itself, on the guidance page or in its companion press release. Two falls in the 1 to 3 band: four stars; the bands are on the ratings page. This piece makes no finding about the conduct of any named TPR official; the quoted remark from TPR's Executive Director of Market Oversight is cited as the organisation's public position, not assessed as personal performance. All sources below were fetched directly on 11 November 2026 and returned HTTP 200.
Sources
- Use of artificial intelligence (AI) in pension schemes, The Pensions Regulator, published 29 October 2026
- Annual Report and Accounts 2025-26 (PDF), The Pensions Regulator, p.8
- TPR urges pension schemes to embrace AI opportunities responsibly, The Pensions Regulator, press release, 29 October 2026
- Pensions Act 2004, section 247, legislation.gov.uk
- Trustee knowledge and understanding, The Pensions Regulator