Field notes on enterprise agentic AI — governance, workflows, memory, and running AI inside your own tenant.
November 5, 2026
An answer in a chat window isn't a deliverable. How we rebuilt document generation on Microsoft Graph and the tenant's own templates — including an honest account of what was wrong before.
November 2, 2026
An approve button is not a safety feature. What enterprise AI actually needs before an agent writes: effect tiers, frozen arguments, and a rule for the answer that never comes.
October 22, 2026
A wrong AI answer doesn't throw an error — it quietly loses you a user. Why the feedback loop, tied to source attribution, is the only instrument that tells you a deployment is drifting.
October 19, 2026
Copilot inherits your tenant's permissions — and most large tenants aren't ready for that. Curated, ACL-trimmed chat is a faster route to safe AI than a two-year permissions cleanup.
August 10, 2026
Every SaaS AI tool asks you to ship data to someone else's cloud. An Azure-native platform with zero data egress changes what your security team has to approve.
August 3, 2026
Enterprise AI that forgets everything between conversations isn't enterprise-ready. How a 6-scope memory model makes AI context durable and governed.
July 27, 2026
LangGraph is a strong developer framework. But enterprise AI workflows need governance, human checkpoints, and visual design. Here's the honest comparison.
July 20, 2026
Most human-in-the-loop AI is an approve/reject button. Here's what enterprise HITL actually requires - structured forms, routing, and delivery where people work.