AI in Financial Services and the risk of automating complexity

I attended the FinTech Week London Summit this week, focusing on sessions most closely connected to our work in operational transformation: AI strategy, governance, agentic workflows and the measurement of value.

The discussions approached AI from different directions, but a consistent message emerged. The strongest applications begin with the outcome, not the technology. They start with a genuine customer or business problem, examine how the work is performed today and then determine where AI contributes to the overall solution.

For financial services organisations, this distinction matters. AI can remove friction, improve service and release capacity. It creates an opportunity to address parts of the operation that have resisted conventional automation and transformation.

It also creates a risk that is useful to consider through the lens of operational debt: the cumulative effect of historical decisions that constrain an organisation's ability to change, serve customers and scale. Operational debt rarely results from one obviously bad decision. It builds through individually rational choices that collectively create fragmentation, workarounds and additional control.

AI could help organisations reduce that debt. But if it is added through a series of disconnected use cases, without redesigning the underlying operation, it may automate existing complexity and create the next generation of it.

Start with the outcome rather than the technology

Several speakers warned against becoming fascinated with the technology itself. An executive instruction to use AI can encourage teams to search for applications rather than begin with a problem that customers or the business genuinely need to solve.

A better starting point is the durable need of the customer. Where is there friction in the journey? How much of that friction is genuinely required, perhaps because it protects the customer or manages risk, and how much is simply a consequence of how the organisation has evolved?

From there, the organisation can define the required outcome, understand the decisions and work needed to achieve it, and decide where AI should form part of the design. The aim is not to find something for AI to do. It is to improve an outcome the organisation already cares about.

That does not mean waiting until the technology is perfect. The pace of development makes that unrealistic. It means experimenting against real problems, learning from operational experience and retaining the ability to adapt as the technology develops.

Understand how the work is actually completed

One of the most relevant observations from the day was the need to watch people perform the work before deciding that a workflow is ready to be handed to an agent.

Process documentation rarely captures the whole operation. A colleague may appear to be following a procedure while also applying learned intuition, interpreting incomplete information, recognising unusual customer circumstances or using informal workarounds that keep the service functioning.

This is particularly important in financial services. A process may contain moments where vulnerability, possible coercion or another high-consequence issue changes what should happen next. The formal process alone may not reveal how experienced colleagues recognise and respond to those moments.

Understanding the work therefore requires observation at the coalface. Organisations need to distinguish repeatable steps from contextual judgement, and documented controls from the practical knowledge people use every day. It is difficult to automate or redesign a workflow responsibly without first understanding how it really works.

Moving from copilots to agents changes the operating model

The distinction between a copilot and an agent was another useful theme. A copilot helps a person to do the work while the person retains agency. An agent can act independently and proactively within its own environment.

That shift is not simply a more advanced form of automation. It changes the role of people, the allocation of decisions and the design of control. The human needs to become the manager of the agent's work, with meaningful involvement where judgement or accountability is required, rather than acting as a nominal approver.

Agents also need work to be defined differently from people. Humans can often interpret ambiguity and balance several objectives at once. Agents perform more reliably with a narrow purpose, explicit instructions and clearly defined boundaries. A workflow may therefore use one agent to collate information, another to analyse it and another to check the output.

That specialisation may be necessary for control, but it also introduces a risk. If each agent is designed separately, the organisation can recreate the same fragmentation, handoffs and unclear ownership that exist across many human operating models. The individual tasks may be efficient while the end-to-end customer journey remains complex.

Every agent needs a clear mandate

The phrase Know Your Mandate, used during the governance discussion, captures an important design principle. Before an agent is deployed, the organisation needs to codify:

• What the agent is authorised to do and which decisions it can make
• The rules and contextual guardrails within which it must operate
• Where its authority ends and when work must pass to a person
• Who remains accountable for the outcome
• How its actions and decisions can be audited and explained

These questions need to be addressed as part of the workflow design. Governance added afterwards is more likely to create extra checks, approvals and manual interventions that slow the operation without necessarily making it safer.

There is also an organisational issue. The people asked to approve AI models may carry much of the accountability and downside risk while receiving little of the benefit. If decision rights, accountability and incentives are misaligned, governance will become a barrier regardless of the quality of the technology.

Measure business value rather than AI activity

The discussion about return on investment exposed another common weakness. Measures such as the number of employees using AI, the number of use cases launched or the volume of hours saved demonstrate activity. They do not prove that the organisation is creating value.

Capacity creation is not automatically business value. Its value depends on what the organisation does with the capacity released. It may reduce cost, increase throughput, improve customer service, strengthen risk management or support growth, but that conversion needs to be explicit.

Speed also needs to be considered alongside quality. Completing work faster is valuable only if the outcome is at least as good. Even then, measurement cannot stop at the task boundary.

For example, AI may reduce average handling time while increasing repeat contacts, complaints or downstream rework. A local measure improves, but the end-to-end journey becomes less effective. The apparent efficiency has simply moved cost and effort somewhere else.

The economic value measure and the organisation's risk appetite should therefore be established before implementation. Measures should follow the outcome across the journey, including successful resolution, quality, recontact, failure demand, complaints, losses and the productive use of any capacity created.

Build the capability to keep evolving

The organisations making the greatest progress are not necessarily those making the biggest AI announcement today. They are often those that have experimented, learned and built capability iteratively over several years.

This matters because the potential applications are evolving quickly. In only a few years, the conversation in banking has moved from using chatbots to produce simple narrative towards agents that may support personalised guidance, exercise judgement within defined limits and enable developments such as perpetual KYC.

Organisations cannot design a fixed solution for the next five years and assume the work is complete. They need an operating model capable of learning and changing as the technology, customer expectations and associated risks develop.

There is an important difference between cumulative capability and cumulative complexity. Both are built over time. Deliberate iteration strengthens the organisation's ability to adapt. Disconnected additions make every subsequent change harder.

This is where AI could become another source of operational debt. Not because the technology inherently creates complexity, but because it may be introduced through the same pattern of local decisions, point solutions, additional controls and short-term workarounds that created much of today's complexity.

AI adoption is an operating model decision

The central lesson from the sessions I attended was that successful AI adoption is not primarily about deploying more technology. It is about designing better operations.

That requires organisations to begin with the customer and business outcome, understand the current operation end to end, involve the people who know the work, define clear mandates and decision rights, and measure whether the whole journey improves.

AI has significant potential to improve financial services. The organisations that realise that potential will not be those that can report the most users or the longest list of use cases. They will be those that can show that customers are receiving better outcomes, work has genuinely been removed rather than shifted, and the operation has become easier to change rather than more complex.

Used in that way, AI can help pay down operational debt. Adopted as another series of disconnected point solutions, it risks becoming its next source.

Want to explore what this means for financial services?

AI can only deliver sustainable value when it is built on simple, well-designed operations. In my previous article, “Why AI for Financial Services Requires Simple Operations,” I explore why simplifying the underlying operation is becoming an essential foundation for successful AI adoption.

Read the blog: Why AI for Financial Services Requires Simple Operations