87% of pensions professionals say their firm uses AI. 75% say it currently touches only 1–5% of their services. The gap between those two numbers is the story. AI is in the room, but it is not yet in the governance of the room, and that distinction is where the risk, and the opportunity, actually sit.
How is AI being used in pension scheme governance today?
The honest answer is: unevenly, and often without the board knowing it.
The most established applications sit with administrators and investment managers, away from the trustee table. Administrators use AI to cleanse and reconcile member data (a task whose importance is difficult to overstate and whose tedium is difficult to exaggerate). Investment professionals use predictive analytics to model funding scenarios, assess employer covenant strength, and monitor portfolio risk in real time. Intelligent chatbots handle routine member queries. None of this requires the trustee board to have formed a view on AI, because none of it sits within the board's direct decision-making process.
What is newer, and more directly consequential for trustees, is the application of AI to the governance process itself. This is where the technology moves from back-office efficiency to boardroom capability, and where the design choices embedded in the platform determine whether the result is better governance or merely the appearance of it.
AI-powered search across the governance record
This is the capability that changes the shape of what a trustee board can do. A platform with AI search reads across every document, minute, discussion, and risk register entry the scheme has ever produced, and answers questions in plain English, cited to the source.
A trustee asking "What was the rationale for the 2022 hedge ratio change?" no longer depends on a colleague's memory, a scheme secretary's filing, or a manual search through years of papers. They get the answer in seconds: the minute, the paper that informed it, the discussion that preceded it, and the actions that followed. The institutional memory of the scheme is no longer bounded by the tenure of its longest-serving member.
This matters in ordinary meetings. It matters more in extraordinary ones (the regulatory enquiry, the Ombudsman complaint, the post-wind-up claim). In each case, the question is the same: can the board produce the evidence that its decisions were sound? AI search across a connected governance record produces the answer. Manual search across a fragmented archive often does not, at least not at the speed the situation demands.
AI-assisted minute drafting
Platforms with AI transcription can capture the trustee meeting via audio, identify decisions and actions, and produce a structured first draft of the minutes within the hour. The scheme secretary reviews and finalises rather than drafting from scratch.
The time saving is significant (days of work per meeting cycle for schemes with lengthy quarterly meetings). The more important effect is structural: the minutes are produced while the meeting is fresh, with the discussion captured as it happened rather than reconstructed from notes taken under the pressure of a full agenda. The fidelity of the record improves, not because the secretary was doing a poor job, but because the task itself was designed for failure: comprehensive note-taking during a meeting that demands the note-taker's full attention as a participant.
Automated action tracking
When AI captures decisions and actions from a meeting, those actions can be assigned, tracked, and followed up automatically, linked to the document or agenda item they originated from. The action tracker updates itself from the meeting record. The follow-up that used to depend on the scheme secretary's diligence (and, occasionally, their memory) becomes a feature of the system rather than a feature of the individual.
Document summarisation
Trustee boards receive substantial meeting packs (hundreds of pages per cycle in a well-run scheme). AI summarisation produces a brief of each paper before the meeting, highlighting key points, changes since the last review, and questions for the board. Trustees arrive better prepared. Meetings run more efficiently. The cognitive bandwidth that used to be consumed by absorption is redirected to judgement.
The caveat matters: the summary is useful precisely to the extent that it makes the trustee more likely to engage with the underlying paper, not less. A summary that replaces reading is not preparation. It is the smooth-ratification problem in a new wrapper.
What can AI do for a trustee board that was not possible before?
The most significant shift is the end of dependence on individual memory as the mechanism for institutional knowledge.
Trustee boards rotate by design. A trustee typically serves for six to nine years; chairs may serve longer, but eventually step down. With each departure, institutional knowledge leaves the boardroom. The new trustee who asks "Why did we take this approach?" often gets a shrug, a vague recollection, or (worst of all) a confident but inaccurate answer from someone who was not in the meeting where the decision was made but believes they know what happened. The institutional memory of a trustee board is, in effect, a consensus narrative maintained by whoever happens to still be in the room. It is not evidence. It is not searchable. It is not reliable.
AI search across a connected governance record replaces that fragile mechanism with something better: the actual record, fully indexed, fully sourced, and accessible to anyone authorised to ask. The answer to "When did we last review this?" or "What was the rationale?" is no longer a function of who remembers. It is a function of what was captured.
This matters most in two situations.
Regulatory enquiries. When TPR asks how a scheme evidences its compliance with the General Code, or how the ORA was informed, or when the SIP was last reviewed, the answer needs to be specific, sourced, and fast. AI search produces it. A shared drive does not.
Post-wind-up claims. When a scheme has wound up and a claim surfaces years later, the trustees (or their successors) need to reconstruct the decision-making history. If the governance record is in a searchable, AI-indexed archive, the evidence is there. If it is in a zip folder on a decommissioned server, it is effectively gone. The capability that makes an archive useful is not storage. It is intelligence.
The institutional memory of a trustee board is a consensus narrative maintained by whoever happens to still be in the room. It is not evidence. AI search replaces it with the actual record.
What are the risks of AI in pension governance?
Trustees do not need to become AI specialists. They do need to ask the right questions: of the technology, of the advisers who use it, and of the platforms that embed it.
Data security and confidentiality
Pension governance data is sensitive: personal member information, financial strategy, employer covenant assessments, legal advice. Any AI system processing this data must operate within a secure, compliant environment. The questions trustees should be asking are specific: Where is the data processed? Is it sent to external AI models? Is it used to train models that serve other clients? Does the data leave the platform's own certified infrastructure?
Platforms built for governance should process AI queries within their own secure environment, with no data leaving the certified infrastructure. This is not a premium feature. It is a precondition.
Accuracy and hallucination
Large language models can generate plausible-sounding answers that are factually wrong (a phenomenon known as hallucination). In a governance context, an inaccurate AI answer about a past decision could lead to poor decision-making or an indefensible regulatory response.
The safeguard is grounding: AI that answers questions only from the scheme's own documents, cited to the source, rather than generating answers from general knowledge. A well-implemented governance AI says "Based on the Q2 2024 Investment Sub-Committee minutes…" and provides the citation. If it cannot find the answer in the record, it says so. It does not invent one. The distinction between these two behaviours (answering from the record versus answering from general knowledge) is the distinction between a governance tool and a liability.
The smooth-ratification risk
This is the subtler danger, and the one least discussed. AI that makes it easier for the board to confirm a recommendation, without making it easier for the board to challenge one, does not improve governance. It degrades it (while producing a more comprehensive audit trail, which makes the degradation harder to detect).
A well-designed governance AI surfaces inconsistencies, unanswered questions, and gaps in the record. It makes the second, uncomfortable question (the one most trustees lack the time or context to formulate) almost free to ask. A poorly designed governance AI produces confident affirmations of whatever the board is inclined to approve. The first reduces the cost of challenge. The second reduces the cost of appearing to have challenged. These are not the same outcome, and the design choices that produce them are largely invisible to the buyer.
Governance of the AI itself
A 2025 survey found that a majority of global businesses had no formal AI governance strategy. For pension schemes, PASA's Data for AI guidance recommends that trustees understand which advisers and service providers use AI, for what purposes, and under what safeguards.
Trustees should ensure that the scheme's AI governance policy is documented and reviewed periodically; that service provider contracts address AI use, data processing, and liability; and that the board has sufficient understanding of how AI is used within their governance platform to make informed decisions about its deployment. This does not require technical expertise. It requires the same diligence trustees apply to any other material risk.
What does the regulatory landscape look like?
The regulatory position is evolving but clear in direction.
TPR's General Code does not currently include specific AI provisions, and the Trustee Toolkit does not cover AI governance. This will change. The DWP's 2026 consultation on trustees and governance signals an expectation that trustee competence will extend to understanding the technology that supports their decision-making: not at the level of engineering detail, but at the level of informed oversight.
PASA's Data for AI guidance, published in 2025, provides the most practical framework available to schemes today. It highlights high-quality data as the bedrock of AI, sets out how schemes should approach adoption, and provides examples of current use cases including chatbots, predictive analytics, fraud detection, and intelligent document processing.
The direction of travel is unambiguous: AI in pension governance is not a question of if, but how. The schemes that engage with it proactively (understanding the capabilities, setting the guardrails, choosing platforms that implement it within a governed framework) will be better positioned than those that wait for regulation to force the issue and then scramble to comply.
How should trustees evaluate AI in a governance platform?
Not all AI is built the same. A marketing claim of "AI-powered" can describe anything from enhanced keyword search to a genuine intelligence layer across the governance record. The questions that separate the two are specific and answerable.
Is the AI grounded in your scheme's own data? The AI should answer questions from your documents, your minutes, your governance record: not from general knowledge or training data. Every answer should be cited to a specific source within the scheme's record. If the AI produces an answer it cannot cite, it has produced something closer to opinion than evidence.
Where is the data processed? AI queries involving scheme governance data should be processed within the platform's certified infrastructure, not sent to external APIs. The question is not whether the vendor uses a third-party AI model (most do), but whether scheme data leaves the controlled environment in order to be processed.
Can it search across everything? Some platforms offer AI on documents only. The real value is AI that reads across documents, minutes, discussions, risk registers, action trackers, and compliance records simultaneously, because governance decisions span all of these.
Does it replace human judgement or support it? AI in governance should surface information, identify patterns, and reduce the cost of asking the right question. It should not make decisions, recommend actions, or cross the line into advice. The trustee signs off on the minute, not the AI. The board approves the strategy, not an algorithm.
What is the audit trail? When an AI-assisted minute is produced, can the board see exactly what was captured, what was edited, and who approved the final version? AI-generated content needs the same evidential rigour as any other governance output (arguably more, because the provenance is less intuitively obvious).
Where Knowa fits
Knowa has embedded AI into every layer of its governance operating system, designed specifically for the evidential and regulatory requirements of UK institutional boards.
Knowa Q reads across the entire scheme record (trust deeds, SIPs, minutes, risk registers, conflicts, correspondence) and answers questions in plain English, grounded exclusively in the scheme's own data. Every answer is cited to its source. No hallucination from general knowledge. No data leaves the platform's ISO 27001-certified infrastructure.
Knowa Verse captures trustee meetings via AI transcription and produces a structured first draft of the minutes (with decisions, actions, and owners identified) within the hour. The scheme secretary reviews and finalises. The record is produced while the meeting is fresh, not reconstructed days later from partial notes.
Connected compliance modules (risk register, action tracker, conflicts, CPD, ESOG) are populated from meeting decisions and discussions, creating a live compliance evidence base that updates as the board works. The evidence is not a reporting exercise. It is a byproduct of the governance process itself.
For schemes approaching wind-up, Knowa Archive preserves the full AI-searchable governance record in a dedicated archive with no ongoing subscription, ensuring that the institutional memory remains accessible for years after the scheme closes.
Every AI feature operates within Knowa's UK-hosted, ISO 27001-certified environment. Scheme data is never sent to external models, never used to train other systems, and never accessible outside the scheme's own access controls.
AI is already in your pension scheme's governance ecosystem. The question is whether to engage with it deliberately or allow it to arrive by default.