Case studies and how we measure

Why this page is empty of names

The e-commerce AI category is full of case studies with implausible numbers and anonymous attribution. They are easy to write and worth nothing to someone trying to make a purchasing decision.

When we publish one, it will carry a named company, a named person, a stated measurement window and a baseline you can sanity-check. Until a design partner agrees to all four, there is nothing to put here.

The metrics we hold deployments to

• Median first-response time

  Measured per channel, before and after. The one customers actually feel.

• Automated resolution rate

  The share of conversations closed with no human reply — reported per topic, because a blended figure hides everything interesting.

• Complaint rate on auto-sent replies

  The counter-metric. A rising resolution rate with a rising complaint rate is a failure, not a success, and this is the number most vendors omit.

• Escalation accuracy

  How often the agent correctly routed something to a human — and how often it should have but did not.

• Marketplace response-time compliance

  For Trendyol and Hepsiburada sellers, whether the response clocks stayed inside the platform's thresholds.

How a measured deployment runs

1. 1. Baseline for two weeks

   Before anything changes, we record your current numbers. Without this there is no result to report later, only an assertion.

2. 2. Shadow mode

   The agent drafts, humans send. Accuracy is measured against what your team actually did.

3. 3. Staged release

   Auto-send enabled per topic, each one measured independently.

4. 4. Review at thirty days

   Against the baseline, with the counter-metrics in the same report.

Become a design partner

We are working with a small number of stores where we can measure properly and publish openly. If you run meaningful support volume across a store plus at least one marketplace, and you would be willing to be named if the results are good, we would like to talk.