The right process, found during the call

Pitched before a line of it existed

The operation's knowledge lived in Google Sites. I designed what should replace it, including the visual identity, the UX and UI, and the structure, then took that to the managers before anything was built. The distance between a Google Sites page and what I was proposing was wide enough that it went through without much argument. It was built on HelpJuice, with the features that mattered written on top of what their system gave us. I coded, and my team built it out with me.

A wizard instead of a memory test

The centre of it is a flowchart of the service processes, turned into a wizard. The agent answers the questions the case actually raises and arrives at the correct process while the customer is still on the line. Before that, an agent had to know which of several hundred processes applied, from memory, and the cost of being wrong was a customer handled under the wrong rules.

In 2022 there was no model to ask

This was before LLMs, and the chatbots available then were not good enough to hand a live case to. There was no box an agent could describe a situation into and get the right process back. So the routing had to be built as a decision tree: the branching of the rules themselves, mapped out and turned into questions, one at a time. That was slower to build, and it is the part that aged well. A model can only route a case correctly if someone has already worked out how the rules branch.

Processes were only part of it

OrangeBase also held the instructions for every tool the agents used, along with a news section for process changes, new features, promotions and institutional updates. The point of the news was that an answer given today should not be running on last month's rules.

Two thousand pages, and a controlled first wave

The migration moved more than two thousand pages off the old CMS. Before the whole operation saw it, we opened access to a wave of early adopters, the strongest agents, who used it on real calls and told us what was wrong while it was still cheap to change. Then the training team ran the rollout waves, on material my team wrote and I led. From the start of the migration to the operation running on it took two months.

What changed

There was a learning curve, and it still came out ahead: average CSAT across the operation rose 27%. The largest gain was where the guessing had been worst. Return & Refund went from 56% to 72%. I never saw the financial side of it, but a customer who leaves the call satisfied is a customer who keeps buying in the app, so the direction is not hard to read.