Example engagement showing how we approach this kind of project. The client, quotes, and figures are illustrative, not a real client record.
Customer support · AI agent
An agent that took 38% of ticket volume in week three.
Their support team was buried under 12,000 tickets a month. We deployed an agent wired to their order, refund, and CRM systems — with evals against six months of real transcripts and a kill-switch wired into the loop.
Outcome · week 3
shipped38%
Of support tickets deflected by week three — zero false-positive escalations
38%
Ticket deflection in week three
$0.04
Median cost per resolved chat
0
False positive escalations to date
The starting state
The support queue was growing faster than headcount. Hiring a fourth tier of agents was on the table — and so was sunsetting the support SLA. We pitched a third option: an agent that handles the top 12 ticket categories, hands off cleanly when policy says so, and gets evaluated against real transcripts every release.
Ticket deflection · 8-week ramp
54% by W8
What we shipped
We treated the agent like a regulated system, not a chatbot. The eval suite ran on every prompt change against six months of real customer transcripts. Tools were narrowly typed and bounded by per-call budgets the agent could not escape.
- 12 tools wired to the real APIs —
refundOrder,lookupShipment,updateAddress— with signed schemas. - Eval suite of 1,400 graded transcripts; 92% pass rate required before any prompt change shipped.
- Per-conversation budget cap of $0.50, hard kill-switch wired into the loop, full trace observability.
- Human handoff with full conversation context piped into Linear — agents see what the bot saw.
Tool registry · last 30 days
Where it landed
Week three: 38% of incoming tickets handled end-to-end by the agent, with zero escalations marked as false positives. Median cost per resolved chat: $0.04, all tokens in. The hiring plan was scrapped; the support team now owns escalations, complex cases, and the eval suite.
Their agent took 38% of our ticket volume in week three and didn't escalate a single false positive. The eval suite alone was worth the engagement.
More case studies
Series B SaaS platform
Two stalled agencies, one rebuild — shipped in eight weeks.
Inherited a stalled rebuild from two prior agencies. We diagnosed the architecture, scrapped the half-finished migration, and shipped a multi-tenant platform that took the first five customers live before the original timeline's discovery phase would have ended.
Professional services firm
WordPress to Next.js, zero-downtime cutover, +247% organic clicks.
Their WordPress site couldn't pass Lighthouse, broke in design review, and required a contractor for every change. We replatformed onto Next.js with a headless CMS — and the SEO compounded the perf wins into 247% more organic clicks over 90 days.
Ready when you are
Got a support queue you can't hire out of?
A 30-minute discovery call. We'll tell you which ticket categories an agent can take — and which need a human.