With the growing adoption and capabilities of agentic AI, many wealth managers are looking to the technology as a way to grow their customers without needing to dramatically increase their adviser team. However, who is responsible if an AI adviser gets something wrong?
AI is becoming a core part of the wealth management business. It is being used across various use cases, from streamlining back-end workflows to providing detailed insights to support decision-making. Demand to use AI is not just coming internally; customers are also exploring how the technology could help them improve their finances.
A recent report from Opinium, which surveyed 2,000 UK adults, found that 54% of investors have used AI to support their investment decision-making. How much you can trust the output of AI always comes into question, and around half of investors appear to have faith in its guidance. The report claimed that 45% of the investors trusted the technology, while 49% did not.
One of the biggest challenges wealth management firms are facing is the growing demand for personalisation. Customers are no longer satisfied with the one-size-fits-all approach and want a firm that treats them as an individual. However, this is difficult to achieve when advisers are already managing multiple relationships. Rather than increasing the team, many firms are looking to AI as a solution. This includes using AI to streamline manual workloads, but another route is using AI to assist the adviser. This can either be as a tool for the adviser to leverage to speed up decision-making or as an AI adviser that can either act autonomously or work as a first response to handle smaller tasks until the adviser is needed.
AI has the alarming ability to always looking right, even when it is not. The confidence in its answers can give people a false sense of certainty and while it might be right the majority of the time, it still has a big window for error. If that mistake breaks through and reaches the client, who bears responsibility if an AI-generated recommendation is wrong and a client loses money?
For Fredrik Davéus, CEO and co-founder of Kidbrooke, the answer is confidently the firm. He said, “That answer doesn’t change because there was a model in the loop. Regulated advice has always had an accountable party, and you can’t delegate that accountability to software any more than you could to a spreadsheet.
“The mistake some firms are heading towards is deploying systems they can’t actually stand behind, and then discovering that “the AI decided” is not a defence anyone will accept, not a client, a regulator, and not a court. So the real question isn’t who to blame afterwards.
“It’s whether you’ve built something you can take responsibility for in the first place. That’s a design decision. If the recommendation comes out of a process you can inspect, explain and reconstruct, you can own the outcome. If it comes out of a black box, you’ve accepted a liability you can’t manage.”
Friedhelm Schmitt, co-founder and CEO Fincite, held a similar opinion. He noted that if a customer ever experiences this situation, the industry has failed. He said, “If this question ever reaches the client, the industry has already failed. You cannot capture the efficiency gains of AI and then push part of the technology risk back onto the client. The client should never have to ask: Was it the bank? The adviser? The model? Responsibility has to be crystal clear.
“If an institution integrates AI into its advisory process, it owns the way that AI is used. A bad investment outcome does not automatically mean somebody made a mistake. But accountability cannot become less clear simply because software takes over a larger part of the value chain. If AI is supposed to create efficiency, institutions have to absorb the complexity internally. They cannot dump it on the client.”
Building governance and protections
If wealth management firms are going to leverage AI in customer-facing situations, there needs to be strict governance and oversight to ensure AI advisers are held accountable.
To achieve this, Davéus believes the single most important thing is to separate the reasoning from the interface. This means keeping the financial logic, projections, suitability checks and portfolio construction in a transparent, testable engine that will behave the same way each time and has version control and auditability. Then language models can be used for what they are good for, understanding the question and explaining the answer in simple terms.
He added, “Once you’ve drawn that line, governance becomes tractable: people sign off on the methodology rather than trying to review every individual output, and you can keep an audit trail that lets you reconstruct why any given recommendation was made. The failure mode to avoid is letting a probabilistic model make the financial decision itself. A system that produces a slightly different recommendation each time you ask it is impossible to govern and frankly has no business making decisions about someone’s retirement.”
One way that some firms might want to ensure the AI system is not going to make mistakes is by having a human check its output prior to it being used further. However, this just stifles the productivity boost it was brought in for.
Schmitt said, “If every AI decision still needs human approval, you might as well forget about automation. More human control does not mean better governance. We need smarter governance.”
Instead, a better system is one where humans define the control framework that defines the data AI can use, what decisions it can make, the boundaries and when it must stop or escalate. This then needs to be continuously tested, monitored and improved.
He added, “The hard part is not adding more approval layers. The hard part is defining rules so clearly that a machine can understand and follow them without ambiguity. Once those rules are in place, the AI has to be allowed to operate within them. The future of governance is not tighter control of every individual transaction. It is better control of the system.”
Strong governance is not just necessary for internal peace of mind, but to adhere to regulatory requirements. Regulators around the world are starting to increase rules around the deployment of AI within financial services, for instance, the EU is continuing to strengthen its AI Act, the UK’s regulators have also been exploring AI guidelines, and in the US FINRA continues to explore AI guidance.
However, opinions are split on how well existing regulatory frameworks are equipped to deal with increasingly autonomous AI systems. Davéus believes the existing advice frameworks are more robust than often believed and are capable of adapting to the new environment. For instance, rules like MiFID II are largely technology-neutral and process-based, requiring firms to establish suitability and justify advice, regardless of how it was created. These rules still work when AI is involved.
He said, “Those obligations apply perfectly well to an AI system; if anything, they expose the systems that can’t meet them. The EU AI Act adds a risk-based layer on top, which is sensible. Will the rules evolve? Certainly, and they should as the technology matures. But I don’t think the core principles, accountability, suitability, and transparency are missing or wrong. The gap is less about the rulebook and more about firms adopting tools that can’t satisfy rules that already exist. A requirement to explain your advice isn’t an obstacle to work around. It’s a fairly good test of whether you should be using a particular system at all.”
On the other hand, Schmitt is less confident on the adaptability of existing regulations and believes more work is needed to account for AI.
He said, “They need to evolve fundamentally. We spend too much time asking how AI can comply with existing regulation. The more important question is whether regulation itself is ready for machines.
Today, regulation is still largely built around text, interpretation, local variation and institution-specific judgment. Humans can operate in that ambiguity. Machines cannot do that reliably at scale. As more of the economy becomes software, regulation will become machine-readable as well. Regulation will become code.”
He added that regulatory requirements will need to be positioned in a way that systems can read them, apply them and continuously prove compliance to them. This will move compliance away from checking individual cases retrospectively, to where rules are embedded directly into a system.
“This is one of the biggest changes AI will force on financial services. We do not simply need more regulation for AI. We need regulation that AI can actually understand.”
Are you speaking to an AI?
As the models behind agentic AI continue to evolve, they are going to become increasingly human-like. However, that could lead to situations where some people might not be fully aware whether the person behind a message is a human or an AI. That raises the question as to what level of transparency should clients expect when AI is being used to advise them.
For both Davéus and Schmitt, it is simple: clients should always know when the AI is involved.
Davéus noted that they shouldn’t only know when AI is behind used, but how it is being used, whether that is simply to phrase an answer to shape the recommendation. He said, “Those are very different things and lumping them together does no one any favours. What matters most is that the client can get an explanation they can understand for any advice that affects their money, and that should be true whether or not a machine was involved.”
Despite how common AI becomes, there will be people who are for it and against it. This makes it important for firms to give their clients an option to opt out. For Davéus, the option to speak to a person should be a choice of preference, not trust. Firms should ensure their AI is explainable and defensible so that people not wanting to use AI are not doing so due to fear of errors.
Schmitt also sees a need for opting out to be simply a case of preference. He said, “Clients should know when AI plays a material role in their advice. And they should be able to opt out. But this is not primarily a moral question. It is a market question. If AI performs certain tasks better, faster and cheaper than humans, using AI can become a mark of quality rather than something firms need to apologize for. Clients who still want predominantly human advice should be able to get it. But it should come with a price tag. That is how every other industry works. You can buy industrially manufactured shoes. Or you can go to a shoemaker and have them made by hand. Both can be good. Nobody expects them to cost the same. Advice will move in the same direction.”
Copyright © 2026 FinTech Global
