1 min readchatbots · rag · integrations
What got our support bot to a 58% resolution rate
The Forethought–Zendesk integration I engineered at Upwork reached a 58% chatbot resolution rate. Much of the work was in what we connected it to.
By Asif Iqbal
I was at Upwork from 2017 to 2024, on a marketplace serving millions of users. I owned the support chatbot platforms there: Ada from 2017 to 2021, then Forethought until 2024. From 2021 I was Engineering Lead for Zendesk, Help Center, and Support Chatbot engineering.
The Forethought–Zendesk chatbot integration I engineered reached a 58% chatbot resolution rate. Zendesk published it in its Upwork customer story.
Context was a big part of it. A support bot that only knows the help center can only give help-center answers, and a person with a problem about their own account doesn't want a generic article.
We wired the bot into Upwork's GraphQL layer and into Zendesk Guide, so it answered with live account context rather than generic help-center text.
The integration itself was built for reuse: an embed wrapper the bot sat behind, with event-tracking hooks. When you can see where conversations succeed and where they stall, you stop guessing about what to fix next.
When the bot couldn't resolve something, the conversation still had to land with the right people. On Ada, before the switch, I added real-time routing data so live-chat handoffs reached the right agent group instead of a general queue.
Chatbot resolution only counts as a win if the cases the bot doesn't resolve are handed off cleanly. (Opinion.)