Technology was the easy part: six lessons from bringing AI to banking clients

Our previous blog two set out what conversational AI promises in investment advisory. This blog three focuses on what it takes to keep that promise once real clients and banking operations come into play.

Deloitte has helped build conversational AI in advisory, and the challenges rarely appeared where teams expected them. What looked difficult turned out to be simple, and what looked simple took weeks to decide. The reason was usually the same. Banks are used to controlling every screen, word and step. In a conversation, however, the client leads. Here are the six lessons we have learned on operationalising conversational AI in advisory. We explain why a question about cash can split a bank three ways, and where the adviser needs to step back in.

1. Define the answer before the model

Start with the basics: "How much cash do I have?" Banks already struggle to agree on one answer. The moment you ask this question, ambiguity surfaces. Does cash mean current account balances? Should an overdraft or a Lombard loan be netted off? What about money market holdings that behave like cash but are not booked as such, and in which currency? Which accounts should be considered (current, joint account, power of attorney)? Each interpretation is defensible but each leads to a different number.

Front Office teams, Risk, and Compliance must reach agreement on what the AI agent may calculate, how it is validated, and who owns the output. Connecting the systems is straightforward; reaching an agreement is not. A better starting point is the definition of the answer, not the capability. The useful question is never whether to add a chatbot, but which client interaction should improve and who remains accountable for the outcome.

2. When a word becomes advice

Discussions about implementation are often not about whether an answer is factually correct. They are about which words the agent is allowed to use. A phrase such as “recommend” can have a different regulatory implication from “most preferred”. Language is part of the bank’s control environment. It requires clear rules for when the agent may inform, when an adviser must validate, and when an output is blocked. It also exposes a deeper tension. Conversational AI is useful because it interprets context and composes a response. But banks need output that is controllable, explainable, and consistent with their approved research perimeter. If a client asks about a fast-moving market theme and one company dominates the public news, should the agent mention it if the bank’s own research does not cover it? If it does, a public headline can become part of what the client reads as the bank’s research house view. If it does not, the answer may feel incomplete. The response principle must be settled early. In successful implementations, language controls are built directly into the workflow. This is consistent with FINMA Guidance 08/2024, which expects institutions to manage explainability and documentation as part of how AI systems operate.

3. Govern the components behind the answer

In a conversational interface, the client sees a single response. For the bank, that response comes from a combination of: approved content, product data, client information, regulatory rules, and the language model used to formulate the answer. Treating these as one undifferentiated "AI answer" is where risk begins and opens a larger question every bank now faces: how much of the system should run on an external LLM, and with which data. External models accelerate development and are intuitive to use. They also raise questions a Swiss bank cannot ignore: data protection, outsourcing, client confidentiality, information leakage, auditability, and data sovereignty.

In Deloitte's experience, this decision is rarely all-or-nothing and treating it as such stalls the project. The more workable framing is narrower: which tasks may use an external model, with which data, under which controls, and for which purposes. Drafting a neutral explanation of what "volatility" means is a different risk decision from letting the same model see a client's positions and suggest a course of action. Separate those cases explicitly, and the external model question turns from a stand-off between innovation and compliance into a set of manageable choices.

4. Data remains your foundation

When a client sees "CHF Hedged acc no HF UCITS" in a response, the reasonable conclusion is that no one thought about technical jargon. A conversational interface does not conceal complexity; it puts it on display. Reuse the internal taxonomy unchanged, and the assistant hands the bank's jargon straight back to the client. Data must be prepared first, with clean interfaces, proper documentation and a useful glossary. Much of what is described as conversational AI work is, in practice, data preparation. This principle decides whether the solution can move beyond a controlled pilot and operate at scale. Each capability needs a defined data foundation: what it draws on, who owns it, how current it is, and what it may be used for. Once established, it is re-used by the next capability. The difference between a pilot that works and a system that scales is not the sophistication of the model; it is the quality and governance of the data foundation.

5. Let real use define the roadmap

Across our work with banks, we often see the same pattern. Institutions design the flows they believe clients will value: refining an investor profile, running portfolio quality checks, constructing modular investment proposals. These are sensible ideas, and they look convincing. But once a conversational interface is live, client behaviour is usually more direct and less elaborate. Clients ask what is on their mind: “Show me my portfolio; what should I do after this news; why has this fee changed; can I block my card?” The real client journey rarely looks like the one drawn up inside the bank, and it starts changing the moment the client begins typing.

This is why the question of gradual versus "big bang" integration matters. Our experience points clearly to a staggered approach. Start with the questions the bank can answer confidently, launch within a controlled scope, and let real usage show where demand, risk and value sit. Real client behaviour is the most reliable guide to what matters. In this context, being reactive is not a lack of ambition. It is discipline: build from proven behaviour, expand where trust is earned, and scale only where the operating model can support it.
 

6. The handover between AI and client adviser

A representative survey by finpension and the IFZ of the Lucerne University of Applied Sciences found that about one in four people in Switzerland have already put a financial question into an AI tool. The behaviour is already visible: clients will ask direct, consequential questions, including the one that banks most need to handle carefully: "What should I do?"

The challenge begins when a supportive answer starts to look like advice. Frameworks such as FIDLEG give that distinction real weight, but in a live conversation the line is rarely static. It shifts with the client's intent, the data being used, the wording, and the possible consequence of acting on it. Escalate too early, and the experience loses value. Escalate too late, and the bank may already have crossed a line it cannot easily re-draw. That is why the handover cannot be treated as an exception or a failure of automation. In the implementations that work best, it is designed as a core feature of the product: the moment when the agent steps back and a human adviser takes over. The agent absorbs routine, information-heavy exchanges; advisers focus on the moments that require judgement, trust and relationship depth.

But that shift does not come from adding a tool to today's process. It comes from redesigning the operating model around clear responsibilities, escalation triggers, and the points where human advice creates the greatest value.
 

The question to reflect upon

One question runs through all six of these lessons. What can the bank safely let the system do, what must remain under human control, and how should that boundary be drawn deliberately?

Public models already answer financial questions quickly and confidently, often with far fewer boundaries than a bank would rightly require. For a regulated institution, the issue is not whether an answer can be generated. It is whether the bank is prepared to stand behind it.

This is the real distinction between an impressive demonstration and an institutional capability. Bringing use cases to life requires decisions about accountability, data, language, sequencing, oversight, and the boundary between information and advice. Swiss banks that make those decisions early will not just automate advisory work. They will design the next standard for trusted, AI-enabled client advice.