AgeBridge Guided ProjectReviewed by AgeBridge
Customer Support Automation
A support agent that drafts answers grounded in the help centre, always cites the article it used, and refuses to guess — anything it is not confident about goes straight to a human with the context attached.
DR
Daniel Roth
Support automation builder
Business problem
A subscription product received about 900 tickets a month, and roughly 40% were the same eleven questions. Answers varied by agent, and first response time on a weekend stretched past a day.
Solution
A support agent that drafts answers grounded in the help centre, always cites the article it used, and refuses to guess — anything it is not confident about goes straight to a human with the context attached.
What was built
- •Help centre ingestion with a cleanup pass that removed 60 dead articles
- •Retrieval tested on its own before any answer was generated
- •Answer drafting that must cite at least one source article
- •A confidence threshold that routes to a human instead of guessing
- •Automatic ticket tagging and routing by topic
Architecture / workflow
- 1Help centre articles chunked and indexed nightly
- 2Incoming ticket classified by topic and urgency
- 3Relevant chunks retrieved and scored for relevance
- 4Answer drafted with citations, or escalated below the threshold
- 5Human reply is fed back as a better answer for next time
Integrations
ZendeskIntercomNotion
Tools / stack
ClaudeVector searchTypeScript
Screenshots
Screenshot slot
Screenshot slot
Demo
Two answers side by side: one answered, one escalated
Results
Reported by the builder: 38% of tickets resolved without a human touching them.
Evidence
- Demo video · Answered vs escalated (1:12)60 seconds showing the thing actually running. The single strongest proof.
- Architecture notes · Retrieval evaluation and thresholdsHow data flows and why you designed it that way.
- Screenshots · Cited answer in the helpdeskReal screens from the build, not mockups.
Proof
Proof score
55/100Add live link — A working link a buyer can click and try.
An AgeBridge reviewer checked the evidence behind this project.
Decisions and tradeoffs
- •Retrieval quality was measured before answer quality. Fixing the docs improved answers more than any prompt change.
- •Escalation is a feature, not a failure. The agent says 'I'm not sure' out loud rather than inventing a policy.
- •Every answer cites its article, so a wrong answer points straight at the document that needs fixing.
What went wrong
Early answers cited outdated pricing articles that were still published. The fix was editorial, not technical: archive the stale pages.
What I'd do next
Track which cited article most often precedes a follow-up question — that is the page that needs rewriting.
Need something like this?
Talk to the builder behind the project, or build your own version.