Call reduction became a question of service resolution.

The bank wanted to reduce customer-support cost by enabling more customers to resolve their needs digitally.

Existing reporting could show which products generated calls, but not why customers still needed a person. And fewer calls would not necessarily mean better self-service: customers could also abandon the attempt, switch channels, or simply stop asking for help.

Before discussing interface changes, we needed a better question: where was customer progress breaking down, and what kind of response was missing?

The reframing needed something concrete enough to use in analysis. I developed five forms of customer progress: Identify, Understand, Decide, Act, and Evaluate.

They are not a prescribed journey. A single customer problem can involve several of them at once. The model helped ask two questions at each point: what does the customer need to move forward, and what must the service be able to provide?

Existing categories could tell us which product or topic a call concerned. They could not show where the customer had become stuck, what response was needed next, or whether digital self-service was actually capable of providing that response.

I added a second diagnostic view: instead of asking only “What is this call about?”, ask “Where has customer progress stopped?”

Working with research and customer service, we translated the customer-progress model into language agents could apply immediately after a call. Over two weeks, nearly 400 interactions were categorized by the kind of progress customers had needed.

The pattern was useful, but the more important finding was why digital resolution had failed. Some customers needed help reaching or understanding a capability that already existed. Others needed information, state, or processes the digital service simply did not possess.

The categories showed what kind of progress customers needed. Looking across the cases also revealed why digital resolution had failed.

Once access problems and capability gaps were separated, one universal self-service solution no longer made sense.

I used the diagnostic to structure a comparison of four possible intervention routes: what each could address, what it depended on, and what new problems it might create.

Contextual search was still only an intervention direction. I defined an MVP around four moments, each connecting a customer question to a responsibility of the service.

Interaction level. The research team evaluated the MVP with ten customers. One immediate issue was the interaction contract: some participants were unsure whether search expected keywords or a natural description of their problem.

That led us to revise how the feature introduced itself and how the search field prompted customers to express their situation.

Service level. A second finding went beyond interaction mechanics. Some customers reached a relevant resolution path and still wanted confirmation from a person.