ClearFeed's customers run support inside Slack, and the same questions kept coming back. I designed how a support head plugs their docs into GPT, checks the answers, and switches automated replies on.
Lots of B2B companies, especially in the US, support their customers in shared Slack channels. It's fast and people love it, until there are hundreds of channels and nobody knows which question is open, who owns it, or which ones are about to breach an SLA.
ClearFeed sits on top of Slack and turns those channels into a proper support queue, with a dashboard for response insights.
Support teams were answering the same questions again and again, by digging through docs and old replies each time.
It slowed every agent down. And after ChatGPT, every support leader was asking some version of this:
“How can these AI systems be more efficient? How can our agents not have to respond to so many user queries at all? And how can we make it easier for agents to resolve problems?”
We ran interviews, rewatched old recordings, read what support leaders were writing, and I sat with ClearFeed's own support team to watch how they worked. Interviews were expensive, so we saved them for support heads.
Support teams have been aware of this problem, repeat questions, for years.
On top of their docs, they had invested in FAQs and knowledge bases for customers.
85% of users prefer asking in Slack instead of searching a support portal.
Keeping those knowledge bases up to date was a constant struggle.
Ironically, Slack channels themselves are one of the best sources of knowledge.
The Head of Customer Support, the person who sets this up. Think David: 15 years in support, 7 leading a team, measured on SLAs, response times and how stretched his team feels.
His job with our feature came down to three things.
Connect public and private docs to GPT within hours, by pasting URLs or using ready-made connectors.
Ask real questions and check how accurate and relevant the answers are, before any customer sees them.
Switch GPT answers on for a triage channel, so the easy questions get deflected.
I owned the setup in the web app. Suraj Mahraj, my senior, owned how it shows up in Slack. We met often, since each half depended on the other, and he unblocked me more than once.
Coffee, paper, and a lot of versions of how the three steps could sit together in one flow, before anything went into Figma.
There are dozens of possible sources, so I asked customers directly. ClearFeed can broadcast a message to every customer channel, so I used it as a poll. Public documentation and Confluence came out on top, and became the first release.
Then I sat with the data science team to learn exactly what each source needs to connect, so the forms asked for nothing extra.
Hi, I'm looking for your input on which data source you use most. I'm designing GPT Powered Automated Answers, and the most popular sources will be released first. Options below; reply if you use something else.
Iterations
Iteration 1
Iteration 2 Shipped
Iteration 1
Iteration 2 Shipped
Iteration 1
Iteration 2 Shipped
Minimal on purpose: fewer choices meant fewer setup mistakes and a faster first run.
The first screen animates what the feature does, so nobody has to read a wall of text to get it.
Clear labels and tooltips where people usually got stuck.
Indexing takes time, so live status keeps expectations right and builds trust.
Colour-coded indexing, failed and indexed states across every source.
Support heads wanted to double-check that their docs were really indexed before trusting the bot. I borrowed the pattern from simple ask-and-answer web apps, mapped the information it needed, and showed both the generated answer and the search results behind it.
Pricing was set per generated answer, so answering everything could get expensive. I added a second mode where an answer is only generated when an agent reacts with an emoji. A single toggle turns it on or off, and teams choose how many answers to generate.
A customer asks a question and the virtual agent replies from the indexed docs. If the answer misses, the agent taps "No, file a ticket", replies by hand, and AI suggests turning that reply into a new FAQ. Suraj designed this side.
Measured over two weeks, across two customer buckets. It rolled out to ClearFeed customers bucket by bucket, and teams like CastAI used it to mix their canned responses with official documentation.
Customer testimonial · CastAI
“It's interesting to have a combination of a previous set of canned responses by the support engineering team for customer queries, and the official Kubernetes documentation as input to the large language model to form a hybrid approach and be able to help our customers with automatic response.”
Shravan AshokStaff Technical Program Manager, CastAIWe collected feedback from each bucket, and three requests stood out.
Teams wanted control over what got indexed, so I added include and exclude rules for public docs.
Some wanted several drafts to pick from, so activation settings got a number of responses.
A few asked for replies with no human check. We held off: a wrong answer sent to a customer costs more than a slow one, and the model needed more data first.
My internship ended here. The feature kept growing, and it now works across all public docs and many private knowledge sources.
Suraj Mahraj, who designed the Slack side and unblocked me more than once, and Ankit Jain, Lalit Indoria and Joydeep Sen Sarma for the reviews along the way.
The story behind the work, or my résumé on one page.
