the setup

Customers calling a bank's IVR system often need information they could get faster on the website. Checking a rewards balance, updating an address, finding a policy answer. The opportunity was pretty clear: redirect them to digital self-service when it would genuinely serve them better.

The key word there is genuinely. We weren't trying to dump people off the phone to save the bank money. We were trying to get people to the right channel for the right task — and respect their choice either way. The hard part was doing that without turning "digital redirection" into a cost-cutting euphemism.

the assumption

The original design made a big bet: that the NLU engine could reliably distinguish between “how do rewards work” (a general knowledge question) and “check my rewards balance” (a personalized account action). These sound different to a human. To a natural language understanding model, they're pretty dang close.

The plan was to use that classification to route customers down two completely different paths. General knowledge gets an article. Personalized action gets an account link. Clean, elegant, logical.

One problem: it didn't work. The NLU couldn’t make the distinction reliably.

Why this was risky: forcing customers to articulate their needs with the precision required for accurate intent classification meant brittle intents and fragile branching logic that would only grow more complex over time. Every new use case would mean more branches, more edge cases, more failure modes.

the breakthrough

So we stopped trying to be perfect and started being helpful.

Instead of complex branching based on intent detection, we designed a single adaptive flow that bundles both general knowledge AND personalized information. If someone calls about rewards, they don't need to specify whether they want to learn about the program or check their balance. We give them both.

This sounds obvious in retrospect. The best design solutions usually do. But getting there required the team to let go of a technically elegant architecture in favor of one that actually worked for humans. That's harder than it sounds, especially when the engineering is already underway.

Design principle: Meet users where they are. Don't assume they’ll perfectly articulate their needs. Don't make AI perfectly classify intent. Design for graceful uncertainty.

the flow

Here's what it looks like in practice. A customer calls and says something about rewards:

The system gives a quick general answer — “rewards are calculated based on your card type” — while simultaneously checking their account eligibility. Then it offers a choice: “I can send you an article about how rewards work, plus a link to your personal rewards balance. Sound okay?”

The customer says yes, and they get an SMS with two separate links: one to a knowledge article, one to their personal rewards hub.

Design decision: two links, not one. Cramming everything into a single destination would have forced the same classification problem we just escaped. Education and account action serve different needs and deserve different paths.

Simple. No branching logic nightmare. No NLU gambling. One flow that adapts.