Yann Kudelski and Vlad Magereanu speak to TWM about designing trustworthy artificial intelligence (AI) architectures – including the challenges of messy data, fragmented systems, and the practicalities of moving AI from controlled testing environments into day-to-day workflows.
What typically prevents wealth managers from successfully transitioning AI from pilots to being embedded in operations day-to-day?
Yann Kudelski: There's quite a variety of wealth managers at this stage that are transitioning AI from pilots to embedding it in operations. But what we've seen is, a lot of times it's designed in isolation, for a perfect world with clean data, a narrow scope, and limited exception-handling. The challenge arises then when you embed it into day-to-day operations. You're confronted with the real world – messy data, more fragmented systems, and more complex rules and exceptions. That's usually the biggest challenge – to move from pilot projects to embedding it operationally.
How are wealth managers approaching the design and governance of AI capabilities today, and what do you see as the most effective ways in which firms are addressing those requirements in practice?
YK: Institutions need to think carefully about when to use AI and when to rely on more deterministic rules-set; they have to find the right balance – when to use which tool to tackle what type of problem. That’s the biggest design choice, although it’s usually a combination of both – leveraging AI tools, but also using a traditional deterministic business rule engine, and combining them in the right balance. Another choice is between single-model or multi-model approaches – using the right model for the right problem, either statically or dynamically.
Vlad Magereanu: Models of different natures solve different problems. If we move away from the generative AI space and the current large-language model (LLM) trend, and go into machine-learning models, then a different set of problems can be solved through mixing and matching an LLM with a more tailored machine-learning model. For example, fraud detection uses very specialised models, which in conjunction with an LLM could give better feedback to the outside world and to the end-user.
What data-related challenges most frequently limit progress, and how are firms addressing these limitations in practice?
VM: The biggest problem firms have is the fragmented data, from the silos they’ve built over time. The big incumbents have a far more fragmented data landscape than a new neo-bank or other challenger, which might have started with a modern infrastructure. This fragmentation is the problem to solve to have more efficient AI use-cases and pilot initiatives, because AI doesn’t have a direct access layer to this data.
How do you overcome this? Well, companies are trying to build so-called data lakes, where they try to bring the data in a uniform format, so it can be consumed in a more consistent way with predictable results. But data lakes might not be suitable for generative AI solutions.
Interested in reading more about designing AI architectures? Mosaic I is available to read in full here.
Where do you see the biggest gaps between client-facing innovation and underlying operational or compliance systems?
YK: Institutions have invested heavily on the client-facing side – portals, apps, and specific optimisation within fragmented landscape silos. The biggest gap is in connecting the two and then benefiting from a step-change in efficiency and process automation – having the glue that brings the legacy systems together with the client experience that can then also enable agentic AI-driven workflows, combined with deterministic business rules.
As AI takes on more responsibility, how should firms maintain trust, accountability, and control?
VM: That’s a difficult one, because to maintain accountability you need transparency, but currently the observability in the space of LLMs and what they do, it’s not in the best state. There’s research going on, but it’s not there yet. But by building observability into the process, which is AI enhanced, that’s already one step ahead. In our regulated world, financial institutions should still have the human in the loop for now – at least until the guardrails that are built within the models, and in the use-cases themselves, are strong and reliable enough that you can take the human out of the loop and still have a deterministic result, a trusted result, out of this process.
And with regulators increasingly focused on outcomes, transparency, and repeatability, how does this shape the way firms design and deploy AI-enabled processes?
VM: I think this ties to what we discussed previously, because regulators need auditability, explainability, and a deterministic result. That’s why at additiv, what we’re doing and what we’re building, is a combination of deterministic rules and processes enhanced with AI, to bridge the structured and the unstructured data.
Because, after all, the financial industry is more rule-based than unstructured data-based. But by bridging this gap, for example by understanding regulations from a regulatory paper through AI and applying them through deterministic rules on top of platform – that’s one way to achieve this kind of AI-enabled process. Maintaining the balance between AI and what their responsibility is, and the deterministic rules which are recognised by the regulators – that’s all about reporting: you need to report the results of your processes.
What makes personalisation hard to achieve at scale? Are firms getting the trade-offs right as they attempt to do so?
YK: First of all, personalisation is not just a content problem, but also a data and orchestration problem. Wealth managers have a significant amount of client data – interaction data, engagement data, preferences, holdings, and transactions. You need to make sense out of all that data together to really personalise the experience for your client. It is a trade-off between superficial personalisation, like relying on one preference and having a superficially personalised experience, but also not to over-engineer the personalisation – it has to still be operationally efficient that you can run it and reap the benefits. And there you have to find the right trade-off between the approaches that you run.
How do you see wealth managers rethinking decisions around buying, building, or partnering for technology, as AI becomes more central to their strategies?
YK: Historically, and in fact still today, especially at the larger end a lot of wealth managers have a tendency to build on their own. But we’re now seeing a tendency to partner more, especially for the foundational platform approaches – connecting everything, having the foundation there, but still building bespoke experiences on top because you have to differentiate yourself.
As firms push AI deeper into decision-making, where do you expect the next architectural bottlenecks to emerge, and how is additiv preparing for them?
VM: The technical architectural bottlenecks are still represented by the limitations of LLMs, the current design and architecture of the foundational models, by limited context, and long-term memory – when they have processes which spread over multiple steps and multiple days, for example, then you need memory to back it up because LLMs cannot hold much data. Then you have the trade-off between reasoning, how much time AI has to think about the problem, and latency, how fast you want to give the answer to the user. Then, as we discussed earlier, there’s the regulatory explainability of what happens within these processes – that’s still something which is not fully solved.
Then, there’s the multi-model approaches and the trust boundaries – what kind of problem you delegate to which kind of model. That’s an architectural decision, which must be taken on a use-case by use-case basis. What we’re doing here is offering a long-term memory through the additiv platform, holding the state of the system and of transactions, and AI-enabled processes on top of it – which solves the long-term memory problem of such a process. When we are talking about latency versus reasoning depth, then we are choosing our use-cases and applying AI where appropriate. As for regulatory explainability, we build the audit trail which is embedded in our platform, but we also build more observability into what the AI is doing on its own.
By 2030, what will characterise the wealth managers who have used this execution phase well?
YK: It’ll be the wealth managers who look at it as a paradigm shift and reconsider their operating models. Continuing with the same operating model might not be the best choice – those who think about the impact of AI, how it will change the operating model, how it might be made better or more efficient, are the ones who will probably be ahead of the curve in 2030.
Interested in reading more about the news, insights, and trends shaping wealth management today? Mosaic I is available to read in full here.
Want to feature in Mosaic II?
Work on Mosaic II: Summer 2026 is already underway. If you would like to feature in the next edition, you can discover the range of contribution options available here.
About The Wealth Mosaic
The Wealth Mosaic is a UK-headquartered online solution provider directory and knowledge resource, focused specifically on the wealth management industry.
For wealth managers, the buy side of our marketplace, The Wealth Mosaic is designed to enable discovery of key solutions, solution providers and knowledge resources by specific business needs.
For solution providers and vendors, the sell side of our marketplace, The Wealth Mosaic exists to support the positioning, exposure and business development needs of these firms in a more complex and demanding market.
Interested in discovering more? Read our reports!
- WealthTech 2026 – read here
- US RIA Toolkit 2026 – read here
- Future View Toolkit 2025 – read here
- UK Toolkit 2025 – read here
- AI Toolkit 2025 – read here
- Client Experience Toolkit 2024 – read here
- US WealthTech Landscape Report 2024 – read here
Join our community and follow us on LinkedIn here.
