KudoDesk
An AI support desk that answers before a human has to.
A support desk that answers from a team's own knowledge, and knows the moment it should stop and hand over to a person.
- 68%
- Tickets resolved without a human
- Under 9 seconds
- First response
- 340
- Teams on the platform
How it started
It began as a conversation, not a brief.
The founders had a hunch and a spreadsheet of the questions their customers kept asking. The obvious build — point a model at the inbox — was the one we argued against. Answering confidently and wrongly is worse than answering slowly, so the product we shaped was less about generating replies and more about knowing when not to.
What we did
- Product strategy
- Conversation design
- Model evaluation
- Full product engineering
Product decisions
What to build, and just as importantly what to leave out.
Incoming tickets are read, sorted and answered from the team's own knowledge, with a person stepping in only when it matters.
Confidence is the product
Every draft answer carries how sure it is and what it is based on. Below the line it never reaches a customer; it goes to a person with the context already gathered.
The handover is designed, not an afterthought
When a person takes over they get the history, the sources and the reason it stopped, in that order. No agent has to re-read a thread to work out where things are.
It learns from corrections, not from noise
When a human rewrites an answer, that correction becomes the reference. Volume alone never changes behaviour, which keeps quality from drifting quietly.
How it is built
The part nobody sees, which is the part that has to hold.
The same engineers who designed this still run it. Here is how it is put together and why we made those calls.
Retrieval before generation
Answers are grounded in the team's own documents and past resolved tickets, retrieved and ranked first. Nothing goes out that cannot point at where it came from.
Work happens in the background
Ingestion, re-indexing and evaluation run as queued jobs with retries, so a slow document or a provider hiccup never holds up a live conversation.
Tenants kept genuinely apart
Each team's content, index and history are isolated at the data layer. One customer's knowledge can never surface in another's answer.
- Answer grounding
- Every reply cites the source it came from
- Isolation
- Per-team data and index separation
- Quality
- Evaluated against a fixed set on every change
- Support
- Monitored and maintained by us
Where it is now
What it does for the people using it.
- About two thirds of tickets are resolved without a person touching them.
- First response lands in under nine seconds, at any hour.
- Agents spend their day on the cases that actually need judgement.
Building something like this?
Bring us the idea, or bring us what you already have.
We can build it with you from nothing, or look at what exists today and tell you what to fix, what to build next and what to leave out.