How a B2B SaaS cut L1 support time by 40% with an AI agent
From a flooded inbox to 70% of tickets resolved without a human · email + chat · 24/7 · human in the loop
TL;DR
- Saturated support team: most tickets were variants of the same questions (account status, billing, access, how-to).
- Clara in production: classifies every ticket, autonomously answers recurring questions with validated responses, drafts replies for complex cases.
- Human in the loop from day 1: below the confidence threshold, the ticket goes to a person with the draft ready.
- Outcome: −40% time on L1 support and 70% of tickets resolved without human intervention, with 24/7 coverage.
The context
A B2B SaaS publisher with hundreds of business customers. Every new customer adds users — and every user adds tickets: order or account status, billing, access, usage questions. The support team answered during office hours; customers wrote at all hours. Support cost grew linearly with the customer base, and quality dropped at every peak.
The problem
L1 support was a funnel: a few people answering, one by one, questions that repeated every week. Three direct consequences:
- The team spent most of the day on variants of the same 20 questions.
- First-response times degraded at every peak (launches, month-end closes, incidents).
- Senior staff ended up covering L1 instead of doing onboarding, retention and key accounts.
Hiring to keep up only bought time: ticket volume grew faster than the team, and every hire arrived already late.
Why classic chatbots failed
Before the agent, the usual paths had been tried:
- Decision-tree chatbot: customers routed around it ("talk to a human") because it didn't resolve — it deflected. Satisfaction dropped and the ticket arrived anyway, with an already-irritated customer.
- Help center / FAQ: well written and rarely read. People write before they search; documentation doesn't answer, it waits to be found.
The metric isn't "tickets deflected". It's "tickets RESOLVED without a human and without friction". A bot that deflects manufactures angry customers, with a delay.
The approach
Four design decisions, each motivated by how a real support team works:
1.An agent on real knowledge, not a script
Clara answers with RAG over the product documentation and the history of resolved tickets. Every answer comes from the company's validated knowledge — not a rigid script, not the model's imagination.
2.Classify first, answer second
Every incoming ticket is classified by topic, priority and risk. Only recurring, low-risk categories get autonomous answers; the rest goes to a human with the context already prepared.
3.Human in the loop with a confidence threshold
Below the confidence threshold, Clara doesn't send: she prepares a draft and hands it to the team. That turned the support team into the agent's ally from week one — they reviewed, corrected, and the agent improved with every correction.
4.Live where support already lives
No new tool: Clara integrated into the existing flow (email + chat). The team kept working in their usual inbox, with one new column: "resolved by Clara".
Implementation
Phase 1 · Ticket audit
Classification of the history: recurring topics, volume per category, what share was genuinely automatable, and where the knowledge gaps were in the documentation.
Phase 2 · Shadow mode
Clara answered in draft, sending nothing. The team compared each draft to what they would have written. Prompts, knowledge base and thresholds tuned until quality was indistinguishable.
Phase 3 · Progressive activation
Highest-volume, lowest-risk categories first, autonomously. Then expansion category by category, measuring resolution and satisfaction at every step.
Phase 4 · Continuous operation
Quality monitoring on sent answers, weekly review of escalated cases, and knowledge-base updates for every gap detected.
Measured results
In production, on the real ticket flow:
- L1 support time: −40%
- 70% of tickets resolved without human intervention
- 24/7 coverage, peaks included, with no new shifts and no new hires
- Senior staff went back to onboarding, retention and key accounts — the work that actually moves revenue
Lessons
Resolving ≠ deflecting
The value isn't that the customer never reaches a human — it's that they don't need one. Every autonomous answer must truly close the topic; otherwise you're manufacturing reopens.
Shadow mode buys adoption
The support team decides whether an agent lives or dies. Watching it work in draft, correcting it, and seeing it improve turned skepticism into ownership.
The agent exposes your documentation gaps
Half the work was cleaning up the knowledge: contradictory answers, outdated docs, undocumented cases. That cleanup alone is worth the project.
Defining what it must NOT touch matters as much
Angry customers, churn risk, legal topics, sensitive billing: direct escalation to a human, always. One badly handled autonomous case costs more than a hundred well-resolved ones.
Tech stack
Does your support only scale by hiring?
If your team answers the same questions every week and peaks bury you, let's talk. Free 15-min diagnostic: I'll tell you which part of your support is automatable and what to honestly expect.