From dead ends to clear paths: How I redesigned a chatbot fallback for BMO Assist
The problem: A chatbot that didn’t actually assist
When I joined BMO, the chatbot had a problem.
​
It wasn’t that the bot didn’t have answers: it was that two-thirds of the time, it didn’t know what to say. Instead of helping users, it gave a generic fallback:
​
“I’m sorry, I didn’t understand that.”
​
That’s chatbot-speak for “You’re on your own.”
​
And when customers are managing their finances, “on your own” isn’t where they want to be.
​
I knew the fallback needed more than a polite shrug. It needed to guide users forward — like a helpful banker, not a brick wall.
Challenge: Stop the dead ends
The product owner wanted to improve accuracy by training the chatbot on more intents. The logic? If the bot understood more queries, fewer users would hit the fallback.
​
But that was a long-term fix. Training takes time, and even the best AI misses the mark sometimes. We needed a solution that worked now.
​
So I asked:
​
“What if the fallback wasn’t the end of the conversation? What if it was a second chance?”
​
That’s how the fall-forward strategy was born.
A poor user experience. The chatbot couldn't match a response even if it were close.
Solution: Offer a way forward
Instead of a dead end, I designed a fallback that suggested “Did you mean…” options when confidence was low. Even if the chatbot wasn’t 100% certain, it could still try to be helpful.
​
Here’s how I led the change:
-
Built a scalable content pattern: I developed 300+ clear, action-oriented CTAs to ensure suggestions felt helpful, not robotic.
-
Aligned cross-functional teams: I collaborated with developers and QAs to test implementation, ensuring the new fallback worked across all scenarios.
-
Created stakeholder buy-in: I shifted the conversation from “Train more” to “Help more.” The results proved it was the right call.
.png)
Structuring clear, user-friendly CTAs
"Not sure if I got it right. Do you need information on any of these?"
Locking a card
Cancelling a payment
None of the above
To make chatbot responses intuitive and actionable, I built a content pattern that guided users with clarity. Each CTA followed a simple, effective structure:
​
-
[Verb + Noun] → Locking a card
-
[Noun Only] → Account balance
-
[Question Format] → Why am I receiving alerts?
The result? Users always knew what to do next — no guesswork.
Real examples
The results? Users didn’t just stop at fallback. They moved forward.
Within the first week of launching the Fall-Forward Model, the chatbot’s response success rate jumped from 45% to 86.5%—nearly doubling its effectiveness overnight.
13,150
Fall-forward responses invoked in the 1st week
90%
Increase in correct responses
43%
Responses contributed to total thumbs up
My takeaway
Sometimes the quickest way to fix a chatbot isn’t with more AI training. It’s with better content strategy. By giving users a way forward instead of a dead end, I improved not only the chatbot’s performance but also the overall customer experience.