The client
Flowdesk is a workflow-management SaaS serving 8,000+ teams. Fast product iterations meant documentation always trailed reality — and their four-person support team paid the price: 1,100 tickets a month, first responses averaging four hours, and a CSAT score drifting downward every quarter.
The challenge
Three problems compounded each other:
- Volume: 60% of tickets were "how do I…" questions already answered somewhere in the docs.
- Latency: Customers in US time zones waited overnight for answers any hour of the day deserved.
- Drag on roadmap: The founding engineers spent every Friday afternoon in the support queue instead of shipping.
Flowdesk had evaluated off-the-shelf chatbots and found them rigid: keyword-matching that broke the moment a question was phrased differently, and no awareness of Flowdesk's actual product logic.
Our approach
We treated the agent as a product, not a widget. Over six weeks:
- Ticket archaeology. We exported 18 months of resolved tickets and clustered them. 41% of volume mapped to just 28 question patterns — that cluster became the agent's first scope.
- Grounded knowledge base. We rebuilt the docs' structure so every article was machine-retrievable, wrote 23 new articles for the gaps, and wired the retrieval layer so the agent answers only from verified sources — with links, so customers can verify too.
- Guardrails before autonomy. Billing, security and data-deletion topics route to humans immediately. The agent escalates with a full transcript summary, so customers never repeat themselves.
- Human-quality bar. The agent drafts in Flowdesk's tone, cites the doc it used, and ends unresolved conversations with a clean handoff rather than a confident guess.
The solution in production
The agent lives in Flowdesk's help center widget and email pipeline. Every answer is logged with the source doc and a thumbs up/down; anything rated down twice triggers a docs review. The analytics dashboard we shipped alongside it shows resolution rate, escalation reasons and doc gaps in real time — so the system improves itself weekly.
Results
Ninety days after launch:
- 63% of all tickets resolve without human involvement, at a 4.6/5 helpfulness rating.
- First response time went from four hours to instant — including nights and weekends, which is when Flowdesk's US customers write in.
- The CS team reopened 30% of their week and moved to proactive onboarding outreach. Trial-to-paid conversion rose 9 points in the following quarter, which the team attributes partly to faster answers during evaluation.
The AI support agent Boost Web built now resolves 6 out of 10 tickets before a human ever sees them. Our team finally works on problems that matter. — Sarah Kim, COO, Flowdesk
What we'd tell the next team
Start from the ticket data, not the model. The agent's accuracy came from knowing exactly which 28 questions mattered — the LLM was the easy part.