From dead ends to clear paths: How I redesigned a chatbot fallback for BMO Assist
Problem
Half of 1M+ chats ended in "I'm sorry, I didn't understand that."
My role
Content strategy lead, working with PM, design, engineering and QA.
The decision
Fix it with content now, not wait for model training: turn fallback into "fall-forward."
Result
Response success rate went from 45% to 86.5% in week one.
Before: the dead end
The hard call: content first, not more training
The roadmap answer was to improve intent recognition through training. That was right long-term, but customers were failing today. I argued that success should mean the customer finishes the task, not the AI guessing correctly, and got PM to change the KPIs to match.
"What if fallback wasn't the end of the conversation, but helped them move forward?"
The system: options, not apologies
Short acknowledgement, then 2–4 clear options, then a safe way to reach a person. Over 300 options, written to three patterns: verb + noun ("Lock card"), noun ("Account balance"), question ("Why am I receiving alerts?").


Shipped



Results, first week
For many customers this was their first try at self-serve banking. Replacing the dead end made the channel feel worth coming back to.