Bringing customers with you

Customers can’t trust what they can’t see. Today, AI-touched journeys in retail banking are mostly indistinguishable from human or rules-based ones. That ambiguity risks being corrosive as the capability of AI evolves.

In our view, customer interactions that involve an AI decision, from a service recommendation to an agentic action, should carry a persistent, plain-English marker. This should not be a legalistic disclosure but something as simple and recognisable as a virtual sticker.

The food industry does this best. Food labelling that uses traffic-light systems on packaged food has been proven to work because it is glanceable, standardised and helpful. Energy ratings on appliances do the same job.

For banking this could look like a small, consistent visual indicator alongside any AI-driven outcome (‘This decision was made by our AI assistant. Tap here to see how’). A necessary pre-step, of course, is to be clear on which decisions do in fact involve AI and where AI is being used in combination with human judgement. Such visual indicators need to have limits to be meaningful to customers. For example, nearly all retail credit decisions will involve machine learning, which could lead to ‘over-stickering’ and confusion.

The test should be that any retail customer, in any channel, can tell at a glance which parts of their banking experience were AI-driven, within a given scope of activity defined by the bank. This moves the disclosure from a compliance overhead into a brand asset. We believe customers will reward visibility, not punish it.

Most customers don’t know what banking tasks AI can helpfully do. The emerging nature of the technology means that they underestimate the simple stuff (budgeting, payment categorisation, summarising statements) and overestimate the complex stuff (regulated advice, tax planning). Because of this, explicit, customer-facing signposting of an agent’s scope creates trust (‘Our AI assistant can do A, B, C. It can suggest D, but you’ll need to confirm. It can’t do E’). In practical terms, banks should explore building into a mobile app or desktop home screen an “ask your agent” guide, refreshed as agent capability evolves. Clear escalation paths when the customer hits the edge of the agent’s competence should be engineered now. The first thing a customer should see when they enter an agent journey is an interstitial screen that says “here’s what I do”. Such clear signposting is the strongest evidence a firm can give that it is supporting, not advising.

Customers will tolerate AI in routine interactions if they know a human is reachable when it matters. The single fastest way to lose trust in an AI system is to make that human path harder than it needs to be.

A guaranteed, one-tap route from any agentic interaction to a named human channel, with the agent handing over the full context, avoiding the need for the customer to start again, builds trust and is something customers experience in other contexts. Premium experiences, for example with airlines, don’t mean that you avoid an IVR, but they often do mean that when you need a human, you reach one fast, and they already know who you are and what you need.

In practical design terms, we are therefore likely to see a proliferation of “speak to someone” buttons, present on nearly all agent-mediated screens. The agent’s context, what the customer was trying to do, what was offered and where they got stuck, passes to the human automatically. The customer service assistant is then empowered to override the agent’s decision, on the spot, with a clear audit trail.

Importantly, this is not about ‘more humans’ but it is that the availability of a human, and the quality of the handover, are made visible parts of the AI product.

Our expectation is that customers will accept AI errors at roughly the rate they accept human errors, provided the recourse path is faster, clearer and demonstrably fair. We need to protect against AI-driven outcomes in retail banking having worse recourse than human ones because no one has thought to design the appeals flow.

Banks need to provide customers with an explicit right to appeal any AI-driven decision, ensuring that they publish a Service Level Agreement, commit to a human reviewer and a written explanation of the outcome.

This may make alarming reading for bank executives focused on delivering improvements in cost-to-income ratios and concerned about creating additional complaints and generalised customer service demand.

Yet without the right to appeal, agent mistakes will not only drive complaints, they risk destroying the levels of trust in banking that provide banks with a unique position in their customer’s lives.

A ‘challenge this decision’ link on AI-driven outcomes, with a guaranteed response time and a commitment to a human in the loop, would deliver accountability and unlock broader agent uses cases. Visible aggregate data on appeal volumes and reversal rates, published quarterly, would not only act as training data for agents but also support a commitment to transparency that customers will support.

Customers must be able to inspect what the agent has done on their behalf and, under certain conditions, have the option to undo it. In the future, agentic systems will quietly accumulate actions that the customers can neither see nor reverse cleanly unless customers are better able to inspect what they have been doing.

In practice, this could be a regular, plain-English digest of agent activity on customer’s accounts or interactions with the bank. In the future this could cover limits adjusted, subscriptions cancelled or spending categories changed. An ‘undo’ option on every line where unwinding is low stakes and feasible would put customers in control.

The best analogy here is Apple’s Screen Time weekly summary or the Google Your Activity digest. Both work because they prioritise disclosure, summarise behaviour the user wasn’t tracking and let customers act in one tap.

A weekly in-app digest, opted-into by default, summarising every action the customer’s agent has taken, with clear undo paths, can help banks build the cleanest evidence base for Consumer Duty outcomes testing and bring home the value of agents to the banking relationships (‘your agent has saved you £43 this month by automatically sweeping balances into your current account’). This would also serve as a clear example of putting not just Humans in the Loop, but Customers in the Lead.

Most consent debates in AI assume the customer is being asked to agree to an agent’s behaviour after the fact. The better model is to let customers pre-commit to what they want and use the agent to honour those pre-commitments.

A ‘ground rules’ interface where the customer sets standing instructions for what the agent can and cannot do would be powerful. As customer-facing agents evolve, this could cover spending caps, subscription reviews and even times of the day the agent is forbidden from acting. The analogy here is smart-thermostat providers enabling pre-sets or streaming services providing parental control systems. They work because they put the customer in the position of control.

Such a ‘ground rules hub’, discoverable at a journey or account summary page, with sensible defaults that reflect both customer protection and Consumer Duty foreseeable-harm logic, are inevitable if banks are to explore new and creative ways to put agents in the service of customers. This is also the cleanest way to ensure vulnerable-customer policy drives actual product behaviour, with vulnerability characteristics becoming tighter default pre-commitments not just status flags.

The previous six moves all have the potential to generate one thing: human escalations.  A customer who taps ‘speak to someone’ after a stickered AI decision, a customer who appeals an agent’s account configuration move, or a customer whose pre-commitment guardrails were tripped will all ladder up to a colleague. If that colleague cannot see what the agent did, cannot explain it, and confidently override it, these new strategies just won’t work.

The frontline readiness gap is large and well documented. The Financial Services Skills Commission’s Annual Skills Report 2026 found that AI skills are among the largest supply-demand gaps in the sector.¹ For a UK bank with several thousand customer-facing colleagues across telephony, branch, complaints, and relationship management, the implication is that the people most likely to receive an AI related escalation are the last prepared to resolve one.

A bank-wide frontline AI literacy programme built around three capabilities – comprehension, override authority, and vulnerability recognition, so that colleagues are trained to spot vulnerability indicators that an agent have missed – can position banks to create a differentiated position on trust.

This frontline enablement is the operationally grounded reality of the human oversight pillar of KPMG’s Trusted AI framework. Human oversight as a principle is meaningless if the humans on the other end of the escalation cannot see what the agent did or override it confidently. Upskilling is the mechanism by which bank accountability becomes a tangible, customer-side experience.