A pension scheme member asks ChatGPT: "Should I transfer my Defined Benefit pension. ChatGPT delivers a confident, detailed answer - and it's dangerously wrong.
AI is quickly becoming part of the pensions conversation.
Some schemes are already experimenting with AI-powered member support, administrators are exploring automation, and members themselves are increasingly turning to tools like ChatGPT for pension-related questions.
But the industry is asking the wrong question.
The real question is not:
“Can AI answer pension questions?”
It clearly can.
The more important question is:
“Can schemes trust the answers enough to bet a member's retirement on them?”
Because in pensions, accuracy, governance and trust matter far more than novelty.
AI is already entering pensions - whether the industry is ready or not
The conversation around AI in pensions has changed significantly over the last 12-18 months and especially since the final Dashboards connection deadline of October 2026.
Pensions UK recently stated that AI is already:
- improving operational efficiency,
- reducing costs,
- supporting member engagement,
- and changing how schemes communicate with savers.
At the same time, it warned that risks around cyber security, governance, data privacy and regulatory compliance must be carefully managed.(Source:https://www.pensionsuk.org.uk/news/article/artificial-intelligence-presents-opportunities-for-pension-schemes-but-risks-must-be-carefully-managed)
The Pensions Regulator (TPR) has also started speaking much more openly about AI.
In a 2026 speech discussing the future of UK pensions, TPR CEO Nausicaa Delfas said:
“AI systems must be governed carefully, with strong oversight, transparency and accountability.”
She also stressed that AI should support human judgement, not replace it.(Source:https://www.burges-salmon.com/articles/102msjo/tpr-highlights-ai-benefits-but-issues-a-note-of-caution-too)
That distinction matters more than most trustees realise.
Because most pension schemes are not trying to build “clever chatbots”.
They are trying to protect members while operating inside heavily regulated environments where incorrect information can have serious financial and emotional consequences.
Generic AI creates a trust problem
Public AI tools are impressive.
They can summarise documents, explain pension terminology and generate convincing responses within seconds.
The problem is that “convincing” is not the same as “correct” - and in pensions, the gap between them is measured in lost livelihoods.
The FCA has already warned about AI hallucinations within financial services, where AI systems generate plausible but inaccurate information presented with high confidence.(Source:https://www.ft.com/content/a16bbdf4-a8c9-4fb2-b61e-e83bb5bfa11b)
That risk becomes particularly serious in pensions.
Questions around:
- retirement age,
- transfer values,
- survivor benefits,
- contribution rules,
- McCloud implications,
- tax allowances,
- or ill-health retirement
are rarely straightforward.
A generic AI chatbot has no understanding of scheme rules, governance boundaries, or the legally critical line between guidance and regulated advice.
It predicts answers based on patterns in language.
That may work reasonably well for general knowledge questions. It becomes a libility when members start relying on those answers to make retirement decisions.
Members are already using AI anyway
Here’s the uncomfortable truth: members are already asking AI pension questions - regardless of whether schemes officially support it.
Research published by PensionBee in 2026 suggested savers are increasingly using AI tools as a “thinking partner” for retirement planning and pension guidance.(Source:https://www.pensionbee.com/uk/press/ai-til-i-die-the-rise-of-chatgpt-in-retirement-planning)
The same research suggested informational pension guidance websites are already seeing traffic reductions as AI-generated search responses become more common.
In other words: members are no longer always going directly to pension providers or guidance websites first.
They are increasingly asking AI.
That creates an important challenge for schemes. If members are already using AI externally, can schemes afford not to provide safer, governed alternatives internally?
The Rise of AI-Written Member Complaints
Another trend is emerging that warrants close attention.
Pensions lawyers and trustees have reported a significant increase in the number of complaints submitted through the Internal Dispute Resolution Procedure (IDRP). What stands out is not only the rising volume but also the alarming sophistication of the complaints. Members who once struggled to articulate their situations are now accurately referencing pension law.
The likely explanation is that members are using generic AI tools to draft their complaints. But, the real risk is not more complaints, it's that these AI tools rely on outdated information, misinterpret scheme rules, or confuse UK and US legislation. A member with a bad AI draft can waste weeks of trustee time.
This can create additional work for administrators and trustees, who must address plausible-sounding but incorrect claims. This is precisely why governed AI matters, not to silence members, but to give them accurate, scheme-specific answers before they turn to the open internet.
Why pensions are different from most industries
Many industries can tolerate a certain level of AI error.
Pensions cannot.
Pension schemes deal with:
- long-term financial decisions,
- vulnerable members,
- bereavement cases,
- regulatory obligations,
- sensitive personal data,
- and complex scheme-specific rules.
A slightly incorrect product recommendation on a shopping website may be frustrating.
An incorrect explanation of a pension entitlement can push someone into poverty in old age.
That is why governance matters so much more than raw AI capability.
What governed AI actually means
“Governed AI” is becoming one of the most important concepts in pensions technology.
Not because the AI itself is necessarily smarter, but because the surrounding controls are the only thing standing between a good answer and a costly mistake.
PASA’s guidance on AI in pensions administration repeatedly stresses the importance of:
- data quality,
- governance frameworks,
- transparency,
- oversight,
- and controlled implementation.
It specifically highlights that high-quality data is the “bedrock” of AI within pensions administration.(Source:https://www.pasa-uk.com/press-release-pasa-publishes-new-data-for-ai-guidance-to-help-the-industry-embrace-innovation-responsibly/)
One pensions communications practitioner recently put it this way: "AI has changed everything and nothing. I still turn up to do exactly the same job. I just have unbelievable help with it now that I didn't have before".
That is a very different model from simply exposing members to unrestricted public AI systems.
Governed AI in pensions should probably include:
- approved information sources,
- scheme-specific knowledge,
- auditability,
- human escalation paths,
- permission boundaries,
- accessibility controls,
- and clear separation between guidance and advice.
Here's a hidden risk governed AI can solve - outdated content. Pension schemes have years of old documents online. Generic AI can surface a discredited 2012 booklet as fact. What used to be a low-priority content governance task is now a member protection emergency.
The goal is not simply to generate answers quickly.
The goal is to generate answers that are:
- traceable,
- explainable,
- consistent,
- and trustworthy enough to act on.
Self-service is evolving into conversational support
For years, schemes have invested heavily in self-service portals.
That trend will continue.
But conversational interfaces are becoming the next stage of digital engagement.
The issue is that traditional portals still assume members understand:
- pension terminology,
- scheme structures,
- contribution rules,
- and retirement processes.
In reality, many members want answers to straightforward questions in plain English.
That is one reason AI support tools are attracting so much attention across the industry.
Used properly, they may help:
- reduce repetitive queries,
- improve accessibility,
- simplify communication,
- and support overwhelmed administration teams.
But a fast, confident wrong answer is worse than no answer at all.
AI should support administrators, not replace them.
AI should support administrators - not replace them
One of the more sensible industry discussions around AI has been the recognition that the technology should augment experienced administration teams rather than replace them.
A recent eight-month trial reported by Pensions Age explored whether AI could support the interpretation of complex scheme rules and benefit specifications under specialist oversight.
The conclusion was not that AI should replace administrators.
It was that AI could help schemes manage growing complexity and rising demand more sustainably when implemented with proper controls and oversight.(Source:https://www.pensionsage.com/pa/AI-has-clear-role-to-play-in-supporting-administrators.php)
That feels like a much more realistic direction for the industry.
Most members do not want fully automated pensions.
But equally, they probably do not want to wait days for answers to relatively simple questions either.
The challenge is finding the balance between:
- efficiency,
- accessibility,
- governance,
- and human support.
Trust will matter more than intelligence
AI capability is improving extremely quickly.
But pensions is ultimately a trust industry.
The schemes that succeed with AI are probably not going to be the ones using the most advanced models or the most aggressive automation.
They are more likely to be the schemes that:
- implement AI slowly and carefully,
- maintain strong governance,
- prioritise accessibility,
- and understand where human judgement still matters.
Because in pensions, the problem is not whether AI can generate answers.
It is whether members, trustees and administrators can trust them not to cause harm.
Question for pension trustees, administrators and scheme managers:
Have you seen members bring AI-generated answers to you as if they were fact? What did you do?
Share your experience in the comments. The industry needs to learn this together, before a regulator has to learn it for us.
