Blog

How we build this

We write about the work itself: command center architecture, measuring agent quality, and our own mistakes.

30 October 2026

What sits behind the 98.8% number

AI agent accuracy is easy to measure in a way that looks good. Here is how we count it and why public benchmarks did not help.

Quality7 min read

Agent findingsData refreshed a minute ago

Risk2 min

Deal “Turkey, 4 adults” has been waiting 40 minutes. The client asked about a surcharge and the manager is offline.

Money11 min

An invoice for 480,000 ₸ has been unpaid for three days. I sent a reminder and will wait until 6 pm.

Finding1 h

Instagram leads close 1.7× worse than WhatsApp. The reason is a reply that comes 12 minutes later on average.

Blog

An agent completes a task. An employee owns an area. The difference is in boundaries, memory and reporting rather than model size.

Four lead sources, a CRM with a thin API, PDF bank statements and analytics in a spreadsheet. What we built in six weeks and where we got it wrong.

We spent two years building analytics for small businesses and kept seeing the same thing: beautiful charts nobody opens. Here is what we changed.

Media coverage, podcasts and videos about MoonAI, in a separate feed.