Composite Client Profile
A Series B HRTech platform serving mid-market employers with a benefits administration and employee onboarding product. Roughly 320 employees, approximately $28M ARR, ACV between $45K and $110K depending on employee count on the buyer’s roster. The support surface is heavy: HR admins escalate anything touching payroll deductions, ACA reporting, or open enrollment windows, and every ticket has a compliance clock attached. Their customer success team was carrying an unsustainable ticket load, and NPS was slipping in the segment that drives their net revenue retention.
The Problem
The company had grown from 90 to 320 employees in 22 months. Support ticket volume had scaled with headcount on the customer side (an HR admin at a 400-employee prospect generates more tickets than one at a 40-employee prospect), but the internal support org had not kept pace. Average time-to-first-response had drifted from 42 minutes at Series A to 4 hours and 18 minutes by the time we were engaged. Priority classification was manual, done by a rotating triage analyst who read every inbound ticket, tagged it, and assigned it to a queue.
Two consequences were compounding. First, urgent tickets (payroll cutoff missed, ACA filing rejected, benefits eligibility broken during open enrollment) were sitting in the same queue as password resets and cosmetic UI questions, so SLA breaches were happening on the wrong tickets. Second, the CS team had started refusing to onboard new logos above 500 employees because the incremental support load was breaking the org. Sales was livid. The CRO had escalated to the CEO.
Prior efforts had failed for two reasons the leadership team could articulate but not fix. They had bought a well-known support platform with a native AI triage add-on, but the model was trained on generic SaaS support data and misclassified HR-compliance tickets more than 40% of the time. They had also tried a keyword rule engine, which caught obvious cases (the word “payroll” in a subject line) but missed the composite signals that actually predict urgency in this vertical.
What Our Team Diagnosed
The root cause was not a support-tooling problem. It was a signal-extraction problem masquerading as one. HR-compliance urgency does not live in any single field of a ticket. It lives in the intersection of: the customer’s employer size bracket, their state of operation (California and New York have stricter clocks), the calendar proximity to a compliance event (open enrollment, quarterly ACA reporting, W-2 season), and the semantic content of the ticket body.
Our diagnostic pass over 90 days of resolved tickets showed that a properly weighted classifier using those four signal families could have correctly triaged 91% of the volume that the incumbent AI add-on had gotten wrong. The company did not need a new support platform. They needed a purpose-built triage layer sitting in front of the one they already owned, and a redesigned lifecycle content program that would deflect the ticket categories that did not need human eyes at all.
Strategy MV3 Shipped
We engaged on our AI and Automation program with a paired content operations workstream. Three deliverables anchored the plan.
The first was a custom triage classifier built on the company’s own historical ticket corpus, deployed as an inbound webhook that scored every ticket on a 0-100 urgency index and routed based on the score. The classifier used a small fine-tuned language model rather than a general foundation model, because latency mattered and the taxonomy was tight. Every scored ticket was still visible to the human triage analyst, but the analyst’s role shifted from classifier to auditor.
The second was a self-serve knowledge program targeted at the 12 ticket categories our analysis showed were fully deflectable with better content. HR admins were repeatedly asking the same questions about ACA codes, dependent eligibility rules, and open enrollment communications, and the existing help center content was written in product language rather than compliance language. Our SEO and content team rebuilt those 12 articles with structured schema, plain-language answers, and embedded decision trees.
The third was a real-time dashboard for the VP of Customer Experience showing ticket flow, SLA risk by segment, and content deflection rate, so the internal team could see the impact of the automation on the metrics the CFO cared about.
Implementation
Vance oversaw the engagement. Our automation team built the classifier and the routing logic; our analytics team stood up the dashboard and the deflection attribution model; our content team executed the knowledge-base rebuild against a briefed editorial calendar. Kickoff to first production deployment took 6 weeks. The knowledge program ran on a rolling weekly cadence for the next 90 days.
Tools used included the client’s existing support platform, a fine-tuning pipeline for the classifier, our standard content brief and QA workflow, and a BigQuery-backed reporting layer that the client’s own analyst could query after handoff.
Outcomes
- Time-to-first-response dropped from 4h 18m to 47 minutes, a 82% reduction, holding steady over the 12-week measurement window after go-live.
- Ticket deflection rate on the 12 targeted knowledge categories reached 38%, meaning nearly four in ten tickets that would have hit the queue in the prior quarter now self-served through the rebuilt help content.
- SLA breach rate on urgency-1 tickets fell from 14% to under 2%, the metric the CRO had escalated on.
- Support cost per customer dropped 31% across the mid-market segment, which unlocked the CS team’s willingness to onboard 500+ employee logos again.
- NPS in the mid-market segment recovered from 22 to 51 over the same window, and net revenue retention on the segment improved by 9 points at the next quarterly board reporting cycle.
Timeline
Kickoff to first classifier deployment: 6 weeks. Full knowledge program shipped and measurable deflection: 14 weeks. NRR and NPS outcomes measured at the client’s next full quarterly board cycle, approximately 5 months from engagement start.
What The Client Said
“The triage classifier alone would have paid for the engagement in the first quarter. The compounding effect once the deflection content was live is what actually changed how we forecast the year.”
— Rachel, VP of Customer Experience
NDA Framing
Client identity is protected under NDA. Company name, product screenshots, and verbatim customer quotes are unavailable in the public case study. Full details, references, and a technical architecture walkthrough are available under mutual sign-off during a discovery call.
Ready to run this play on your support org?
If your CS team is drowning in ticket volume that scales linearly with your customer base, and your existing AI support tooling is misclassifying the tickets that actually matter, our AI and Automation program can diagnose the signal-extraction gap and ship a purpose-built triage layer in weeks, not quarters.
Book a discovery call or explore our AI and Automation service.
