The Impact Of GenAI In The Pension And Wealth Space
Guest: Matt Gosden, Strategist & Enterprise Architect Advisor on AI, Data, and Innovation in Financial Services
Matt Gosden is a strategist and enterprise architect working on pensions, wealth and long term savings, advising organisations on applying AI and data in regulated parts of financial services. He has spent close to thirty years in the industry, starting in technology. This conversation is about what innovation means in a sector built on decades old systems, where generative AI is already useful, and what he would standardise if he could start again.
What does meaningful innovation look like in long term savings?
Three things, and the first is the one the sector keeps failing. Customer experience, because engagement with savings and pensions has been low for a long time: customers are often not very engaged with what insurers or banks offer, and even with challenger banks in the market, traditional banking can be sluggish and pensions more so. The useful version of innovation is helping customers understand and use the products they already have.
The second is technology, which in this sector means the platform conversation and whether to replatform in order to deal with legacy.
The third is commercial, and it is the one that moves the numbers. In insurance particularly, innovation in pricing strategy, customer retention and cross selling, usually enabled by technology, has a direct effect on the bottom line.
His explanation for the pace is the familiar pair, regulation and legacy, and he is matter of fact that financial services does not innovate as quickly as technology does.
Why is the infrastructure so fragmented, and is legacy the real problem?
Because every company built its own, and no dominant platform ever arrived. Core systems in insurance and pensions demand high resilience, security and privacy, which produces cautious choices, and those choices accumulated over decades into fragmentation.
He is clear that these firms were not technology laggards. In the late 1990s they were running valuations on mainframes, which looks dated now and was the right answer for that scale then.
What is missing is a standard. There are platforms in the wealth space, and Salesforce is present inside large insurers, but no universally accepted dominant player has emerged in the way that cloud platforms dominate infrastructure, and integration remains costly and complex. That absence is why everyone is dealing with fragmented legacy.
His conclusion is more encouraging than the diagnosis. Not all legacy is bad. You can build service layers around it and modern APIs for integration, which gets contemporary functionality without discarding what works.
Should you fix the data before doing AI?
No, and this is the most practically useful part of the conversation. The standard move is a replatforming programme and a data tidying programme, on the grounds that the data is messy compared with a technology company's.
His position is that messy data still supports scrappy work, and that a use case driven approach beats a data programme. He would not build a large data infrastructure up front. Start with specific use cases, find out what they actually need, test against the data you already have, and productionise iteratively once something works.
The cost of the alternative is the part people miss. Focusing on data preemptively can consume resources without producing anything, when quick moves with AI would have told you what you needed to build.
Where can generative AI actually help?
Internally first, because the human stays in the loop. Most current initiatives are internal facing and aimed at making knowledge workers more efficient, including AI assisted search across corporate data, which improves what is effectively a document management problem. These work because the AI is an assistant with somebody checking it.
The bigger opportunity is customer facing, and his argument for it starts from what customers are already doing. People ask general purpose chatbots their financial questions and get well composed and mostly accurate answers. His estimate of the accuracy is the important number: roughly 60 to 70 percent on complex questions, and 80 to 90 percent on simpler ones.
That is the opportunity and the risk in one statistic. Significant financial decisions are being made on information that is sometimes wrong, which gives the industry an imperative to make sure that advice delivered through AI is sound, and an incentive, since customer engagement is something providers want anyway.
What AI does well here is specific: simplifying complex jargon, and enabling asynchronous communication with a provider rather than a phone queue. His standard is comparative rather than absolute. These systems may never be 100 percent accurate, and a well structured customer facing application can still outperform typical human interactions and improve on the status quo.
The pensions example makes it concrete. Many people access their pension pot without understanding the tax implications or the benefits they may lose. AI can explain the implications of that action without giving regulated advice, which leaves the customer better informed and aligns with the provider's responsibility for consumer outcomes.
How do you plan a roadmap when AI moves this fast and the industry does not?
You accept rebuilding. In two and a half years his stack was rebuilt four times as the underlying capabilities improved, and competing directly with the leading AI labs is not realistic given their resources, so the cost is continuous reinvestment.
Against that sits an industry where buying decisions take a long time: he puts the average enterprise software decision at around eighteen months. What follows is herd behaviour, and he treats it as an opportunity rather than a complaint. Once one firm adopts something successfully, others follow, so the strategy is to get something good live and let that pull the industry along.
The reason to ship rather than perfect is what you learn. Putting solutions in front of consumers produces insights, and early trials teach unexpected lessons, which is the case against perfectionism. He credits regulatory support for making that possible, naming the FCA's sandbox in the United Kingdom as an invitation to test innovations safely.
What would he redesign from a blank sheet?
He would standardise the product, because it is not actually complicated. Underneath, these are pots of money with options for withdrawal under varying tax and regulatory conditions. A standardised approach would simplify a great deal.
The second thing he would build is the technology base as a commons. He wants standardised, open source inspired systems producing commoditised solutions, on the grounds that semi custom technology is expensive and delivers little differentiation, and that streamlining the base is what would unlock innovation across the industry.
His view of the end state is agentic. AI handling transactions much as a financial adviser does today, managing the intricacies in the background and consulting the user only on the critical decisions, which would democratise advice that is otherwise expensive. He is honest about the obstacles: varying regulatory standards make it complex, and distribution channels are their own problem.
The build advice attached to that is immediately actionable. Making services agent friendly is not difficult, and mostly means richer inputs and outputs. His example is a retirement calculator: today it returns a number, where a richer output would support actual understanding, which serves the agent and the human equally.
Takeaways
- Engagement is the real product problem in pensions, and it has been low for a long time regardless of interface quality.
- Fragmentation came from every firm building its own core and no dominant platform ever emerging. Integration cost is the tax everyone pays.
- Legacy is not automatically the enemy. Service layers and modern APIs around it buy most of the benefit without a rebuild.
- Do not run a data programme before an AI use case. Test against the messy data you have and build the infrastructure the use case turns out to need.
- Internal AI works because a human checks it. That is why assisted search and document work came first.
- Customers are already asking chatbots about their pensions, at roughly 60 to 70 percent accuracy on complex questions. That is the argument for the industry doing it properly, not for waiting.
- The benchmark is not perfection. A well built customer facing system can beat a typical human interaction.
- Explaining the tax consequences of accessing a pension pot is guidance, not advice, and it is where AI is immediately valuable.
- Four stack rebuilds in two and a half years against eighteen month enterprise buying cycles is the real tension in this sector.
- Herd behaviour cuts both ways. One live success pulls the industry forward, so being first is strategically useful.
- Semi custom technology costs a lot and differentiates little. Standardised, commoditised plumbing is what frees up innovation.
Transcript
Alex: Hello and welcome to the Curiosity Code podcast. I'm your host, Alex. And today we're joined by Matt Gosden, a strategist and enterprise architect working at the forefront of pensions, wealth, and long-term savings. He advises organizations on how to apply AI and data in practical, responsible ways across highly regulated areas of financial services, from long-term savings to protection products and customer engagement.
Matt Gosden: Yeah, lovely to be on the show. Thanks so much, Alex.
Alex: Let's dive into the world of long-term savings and wealth. What does meaningful innovation actually look like in that sector? Especially when you're working with decades-old infrastructure and deeply regulated environments.
Matt Gosden: Yeah. So I've been in the industry for a long time, maybe nearly 30 years now. I started off, my very first job was in tech actually, but I've spent most of my time in financial services. Financial services don't innovate as quickly as some other industries, especially tech, which innovates rapidly. This is partly due to regulation and legacy. We can talk more about that in a minute. But if you speak to most business leaders, meaningful innovation usually comes from thinking about how to improve the customer experience, particularly in savings and pensions, where engagement has been low for a long time.
Customers often aren't very engaged with what insurers or banks offer. Even with the rise of challenger banks, traditional banking can still be a bit sluggish, and certainly, pensions can be too. Meaningful innovation focuses on helping customers understand and use their products better. There's also a tech aspect, with discussions about platforms or re-platforming to deal with legacy issues. But for me, a lot of it is about the customer experience. The third area is the commercial levers. In industries like insurance, innovation around pricing strategies, customer retention, and cross-selling, often using tech, can also be quite meaningful and impact the bottom line.
Alex: Mm-hmm. So the infrastructure is quite old, right? My take on why it's old is because it works and there's not much interest in changing it since it's doing what it was designed to do. In the world of AI, we carry heavy technical debt from legacy systems that are still operational. They work, but as you said, innovation is driven by user experience and business opportunities. How do you navigate updating these core systems without disrupting the core trust, compliance, and continuity of a business?
Matt Gosden: Yeah, I can talk about the insurance and pension sides, but less about banking. These core systems require high resilience, security, and privacy, leading to cautious approaches with potential fragmentation over decades. Many financial services firms were early adopters of technology. In the late 90s, we still used mainframes for things like valuations, working with what, in hindsight, might seem dated but was necessary for scale.
The fragmentation arises because each company set up its own system, leading to complexity. We haven't seen one dominant platform taking over. There are platforms in the wealth space, but no universally accepted dominant player has emerged. This lack of a dominant standard means you're dealing with fragmented legacy systems. But not all legacy is bad; you can build service layers around it and modern APIs for integration, allowing for more contemporary functionality without discarding everything, despite slow innovation in this space.
Alex: So you mentioned a hypothetical provider that could overtake the industry. I'm surprised there's none currently. There must still be several key players in the market, right? Based on your experience, what are the biggest architectural challenges they face? Is it data unification as AI relies heavily on the state of data in the organization, user experience, or something else?
Matt Gosden: Yeah. There are some solutions, but no singular dominant platform like AWS or Azure. Salesforce is present in big insurers, but integration remains costly and complex. Using systems effectively is crucial and more expensive if not fully utilized. For AI, there's much talk about messy data needing modernization. Although data isn't neat compared to tech firms, you can still do scrappy work. Firms often focus on re-platforming and tidying data programs, but use case-driven approaches can be more effective. Starting with scrappy projects can prove concepts and build out data infrastructure once something works.
In AI initiatives, I wouldn't build a massive data infrastructure upfront. Start with specific use cases to see what you need, test with existing data, and iteratively productionize once something proves successful. Preemptively focusing on data can waste resources; it's often better to make quick moves with AI.
Alex: Everyone wants to move quickly with AI; it's a catch-up game. Talking about AI, you've written about the responsible rollout of AI in financial services. Where do you think GenAI can make the biggest impact in the pensions and wealth space?
Matt Gosden: Yeah, the pace in AI is incredible. Currently, many initiatives focus on internal-facing use cases to help knowledge workers become more efficient. This includes AI-assisted search across corporate data to improve document management. Internal-facing efforts work well as the human remains in the loop, acting as an assistant. However, we're exploring bigger opportunities in customer-facing applications.
Many customers now use tools like ChatGPT and Bard to ask financial questions online, receiving well-composed, mostly accurate responses. But these models' knowledge can run only 60-70% accuracy for complex questions and 80-90% for simpler ones. Consequently, some crucial decisions might derive from flawed information, posing opportunities and challenges for the industry.
The industry has an imperative to instill trust by ensuring advice via AI is sound. Encouraging customer engagement aligns with provider incentives. AI can simplify complex jargon and enable asynchronous communication with providers. While AI may never achieve 100% accuracy, well-structured customer-facing applications can outperform typical human interactions and improve the status quo in overall customer relations.
Alex: So your focus right now is on customer-facing solutions, right? As I picture it, I, as a user of the system, can ask to look at different solutions available in the market regarding my pension or long-term investment, analyze what's happening, and come back to advise something. It's more like an advisor, right?
Matt Gosden: Yeah, mostly. It's about enabling consumers to use services effectively. Let's take pensions. Many people tap into their pension pots without understanding tax implications or potential benefit losses. AI offers guidance, helping consumers grasp the implications of such financial actions without giving direct advice. This empowers customers to make informed decisions, aligning with providers' responsibility to improve consumer outcomes.
Alex: I work with AI daily, developing AI solutions and using them personally. With the fast-paced changes, it's like playing catch-up. Your industry is slow-moving and highly regulated, juxtaposed with the rapid AI landscape. Consumers also have high expectations from using AI solutions daily. How do you create a roadmap amid these dynamics?
Matt Gosden: Yeah, keeping up is challenging. In two and a half years, we've rebuilt our tech stack four times as capabilities improve. Competing with leading AI labs is tough since they have immense resources. Innovations mean constant reinvestment, which can be hard to manage.
The industry moves slowly because financial decision-making, particularly software buying, takes extensive deliberation, averaging 18 months for enterprise solutions. Herd mentality then kicks in; once someone adopts an innovation successfully, others follow. Our strategy involves getting something good live, sparking industry-wide progress.
Getting solutions into consumers' hands yields valuable insights. Early trials teach unexpected lessons, making the case for iterative, customer-facing solutions rather than perfectionism. Regulatory support, like the FCA's sandbox in the UK, promotes such testing, inviting firms to explore innovations safely.
Alex: Yeah, I can relate. In financial services, the process is slow but sticky, and the industry has immense capital. What unique innovation opportunities are overlooked in the pensions and wealth space?
Matt Gosden: I've discussed customer engagement: it's crucial. AI can enhance customer interaction with pensions and insurance, reducing jargon and frustration while encouraging engagement on their terms. Many already use chatbots for financial inquiries, indicating demand.
Additionally, strengthening technological backbones could catalyze further innovation. There's a need for standardized, open-source-inspired systems, driving commoditized solutions. Currently, semi-custom tech is expensive, with little differentiation. Streamlining tech bases could catalyze industry-wide innovations and foster ease of use.
Alex: If you had a blank sheet to redesign how this industry works, where would you start, and what would you leave behind?
Matt Gosden: The underlying products aren't complex—essentially being pots of money with options for withdrawal under varying tax and regulatory conditions. A standardized approach would simplify things. I envision a future where AI handles transactions similarly to financial advisors today, managing intricacies in the background and consulting users solely for critical decisions.
Such an agentic system could democratize professional advice, otherwise costly, generating outcomes driven by technical expertise. Although this is complex due to varying regulatory standards, it's an intriguing innovation. Distribution channels pose challenges, but this could enable better consumer experiences and financial outcomes.
Alex: This aligns with previous episodes discussing agents driving infrastructure development. I foresee AI agents assisting with daily finances across many agent-driven APIs. Modernizing infrastructure towards API-centric designs makes it agent-friendly, rather than relying on layers built for manual data handling.
Matt Gosden: I agree. Creating agent-friendly APIs isn't difficult, merely enabling enriched inputs and outputs for more engaging interactions. For instance, retirement calculators could benefit from richer outputs beyond simple numerical results, allowing deeper understanding. Developing agent-friendly services makes interactions meaningful for both agents and users. In reality, companies must integrate these agents while enhancing internal services. Thank you.
Alex: Well, I'd like to wrap this up and thank you for your time. I enjoyed the conversation a lot. Thanks a lot, Matt.
Matt Gosden: No, thanks very much for inviting me along. It's been great.
Alex: And to our listeners, as usual, thanks for listening to the end. Don't forget to hit the like button on YouTube, subscribe to the channel, and if you're listening on any podcast platform, please leave a positive review. See you in the next episode. Bye-bye.