Skip to content

Outcomes,not decks

Four recent engagements. The clients stay anonymous; the numbers are exactly as measured.

Case 01 · Healthcare SaaS

Patient intake, rebuilt

−62%intake time

The problem. A multi-clinic healthcare platform was losing patients halfway through a fourteen-screen intake full of duplicate questions and manual insurance checks.

What changed. After shadowing front-desk staff, we rebuilt the flow around the few questions that actually gate triage and moved insurance checks to a background job. It shipped clinic by clinic behind a feature flag.

Before and after: Patient intake, rebuilt
MeasureBeforeAfter
Screens to complete145
Insurance checkManual, blockingAsync, with fallback
Median intake11m 20s4m 18s
Abandonment31%9%

Case 02 · Logistics

Ops that run themselves

31 hrssaved every week

The problem. A logistics firm's back office ran on tribal knowledge: forty steps and seventeen handoffs, held together by spreadsheets.

What changed. Seventy percent of approvals turned out to follow predictable rules, so they now route automatically with a human on the exceptions. The rebuild took eight weeks and has run without an error since launch.

Before and after: Ops that run themselves
MeasureBeforeAfter
Process steps4012
Manual handoffs170
Approval routingTribal knowledgeRule-based
Weekly hours44h13h

Case 03 · B2B analytics

From factory to focus

2.3×activation in a quarter

The problem. A B2B analytics startup shipped fast but kept few of the customers it won. The roadmap had eighty items and no spine.

What changed. Usage data showed one workflow drove nearly every retained account. We cut the roadmap by sixty percent to serve it, and rebuilt onboarding so new users reach it in under five minutes.

Before and after: From factory to focus
MeasureBeforeAfter
Roadmap items8032
Time to core value3 sessions< 5 min
Activation8.4%19.3%
90-day retention41%58%

Case 04 · Customer support

A support copilot

78%of tickets resolved without a human

The problem. Support volume was growing faster than the team could hire to meet it.

What changed. An LLM layer now reads each ticket and drafts a reply grounded in the help centre. Anything low-confidence goes to a person, and every correction feeds the eval set. Concept to production took six weeks.

Before and after: A support copilot
MeasureBeforeAfter
Tickets touched by a human100%22%
First response3h 40m1h 42m
Quality reviewAd hocWeekly eval set
Escalation pathNoneConfidence-gated

Bring the fuzzy problem

Thirty minutes, no pitch deck. Tell me what's going wrong and you'll get an honest read on whether I can help.

Start a conversation

2 client seats open for Q4 2026