Measure before automating
We clustered twelve months of tickets and found 61% of volume sat in nine intents — all answerable from documentation the company already maintained.
SaaS · 2026
Deflecting 65% of support tickets without a single angry tweet
A retrieval-grounded support agent with confidence gating and human review that resolved two thirds of inbound tickets while raising customer satisfaction.
10 weeks
Engagement length
5 specialists
Delivery team
65%
Tickets fully deflected
+9pts
Customer satisfaction
The situation
What we walked into.
Support volume was growing 14% a month against a hiring freeze. First-response time had slipped to nine hours and the team was re-answering the same forty questions. An earlier off-the-shelf chatbot had been switched off after it confidently invented a refund policy that did not exist.
Client profile
Client
B2B SaaS platform
Industry
SaaS
Duration
10 weeks
Team
5 specialists
Approach
The decisions that mattered.
Not a chronology — the four calls that determined how the project turned out.
Grounded, not generative
Hybrid retrieval over docs, changelogs and past resolved tickets, with every answer required to cite its sources. No source, no answer.
Confidence gating
Below a tuned threshold the agent does not guess. It summarises the case, attaches its research and routes to a human — which is why satisfaction went up rather than down.
Evaluation as a product
A 900-case golden dataset scored by the support leads runs on every prompt, model and content change, blocking regressions before release.
What we shipped
Delivered scope.
- LangGraph agent with hybrid retrieval and reranking
- Citation-required answers with source links in every reply
- Confidence gating and human handoff with research attached
- Zendesk and Slack integration inside existing workflows
- Golden-dataset evaluation harness wired into CI
- Cost, accuracy and deflection dashboard for support leadership
Technology
LangGraphOpenAIAnthropicpgvectorFastAPILangfuseZendesk APIGrafana
Results
What changed for the business.
65%
Tickets resolved without a human
-82%
First response time
+9pts
CSAT versus pre-automation baseline
$0.11
Average cost per resolved ticket
- Hallucination rate measured at 0.4% on the evaluation set, with every flagged case traceable to a documentation gap.
- Support headcount stayed flat through a 40% growth in customer base.
- Documentation improved as a side effect: unanswerable questions became a weekly content backlog.
The confidence threshold is the whole product. It knows when to shut up and get a person, and that is why our customers trust it.
B2B SaaS platform
Keep reading
Other engagements.
Fintech
2025
Rebuilding a lending platform that was losing deals to its own load times
We replatformed a monolithic loan-origination system into a modular web platform with automated underwriting checks, cutting application time from 22 minutes to under 6.
-73%Application completion time
+41%Approved applications
12minCommit to production
Web DevelopmentDevOps & ArchitectureUI/UX Design
Read case study
Retail & commerce
2025
One app for 340 stores, shipped on both platforms in fourteen weeks
A Flutter loyalty and commerce app with offline-tolerant checkout, unified inventory and a design system that finally made the brand consistent across channels.
4.8★Combined store rating
+29%Repeat purchase rate
14 wksConcept to both stores
Mobile AppsUI/UX DesignWeb Development
Read case study
Your project
Have a comparable problem?
Tell us where you are stuck. You will get a senior response within one business day, including whether we think we are the right team for it.