An AI assistant that deflects support tickets
Support was drowning in repeat questions. A retrieval-based assistant grounded in the company's own docs answers the easy ones instantly and hands the hard ones to people with context.
Most tickets were variations of the same twenty questions, first replies took hours, and agents spent their day searching documentation instead of helping people with real problems.
- 01Hours to a first reply on simple questions
- 02Agents searching docs instead of solving problems
- 03No safe way to put an AI in front of customers
How we got there
Ground every answer
Retrieval over the company's own docs and past tickets, with citations, so answers come from facts rather than guesses.
Guardrails first
Defined what the assistant may and may not answer, with a clean hand-off to a human whenever confidence is low.
Assist the agents too
Gave agents suggested replies and the relevant sources, so even escalated tickets are resolved faster.
Close the loop
Thumbs up and down, review queues and weekly evaluations turn every conversation into a better assistant.
The shape of the change
A simplified before and after. The point is not the boxes, it is that each part can now change, scale and fail on its own.
- Customers get an accurate answer in seconds, any hour
- Agents spend their time on the problems that need a person
- Every answer links back to its source
- Quality is measured weekly, not assumed
- 01Retrieval quality matters more than model choice.
- 02A graceful hand-off builds more trust than a confident wrong answer.
- 03Evaluate on real tickets, not on demo questions.
“It handles the repetitive questions beautifully, and my team finally has time for the hard ones.”