Conversations on enterprise AI

Adoption is the first of three stages, not the finish line. The audio companion to the post: Adopt, Operate, Govern as a ladder you cannot skip rungs on — why operating is the stage everyone improvises from memory, why governing is the one that notices when an agent drifts rather than when it stops, and where to start if you have adopted AI and stalled at "we think it's fine."

Related reading: From Wrangling to Governing: Adopt, Operate, Govern →

Player not loading? Listen on Substack →

Most teams running AI agents are wrangling them — watching the output, remembering roughly what is set up, fixing what they happen to notice. The audio companion to the post: why the quiet, competent, wrong failures are the ones that should worry you, how a scan of one real environment went from roughly 270 to 384 moving parts in a few weeks, and the three things — see, understand, act — that staying in control actually requires once wrangling hits its ceiling.

Related reading: Wrangling AI Agents in the Wild →

Player not loading? Listen on Substack →

The first move toward Anthropic's Claude Certified Architect exam was not a textbook — it was handing an agent the broken plumbing on the study tools themselves. A field note on preparing in the open: the gate that finally lifted for a firm this size, what the exam actually rewards over rote recall, the two tracks running at once, and the auth-fix loop that turned out to be the whole lesson in miniature.

Related reading: Studying for an AI Architecture Exam, With AI →

Player not loading? Listen on Substack →

A dashboard built for a pre-AI company answers one question: how did we do? Wire Claude into the work itself and that dashboard quietly stops being the one you need. The four instruments an AI-integrated Mission Control adds that a lagging-indicator dashboard structurally cannot hold — decision latency, operational friction, computed maturity, and synthesis on the surface — and why bolting an AI tab onto the old dashboard just measures the old company faster.

Related reading: Mission Control for an AI-Integrated Company: Why the Dashboard Changes Shape →

Player not loading? Listen on Substack →

NIST tells you what governance is supposed to achieve. Configuration management tells you how the work gets done day to day. A crosswalk for compliance teams — Map to Baseline, Govern to Change Control, Measure to Drift Detection — and the operational artifacts that close NIST's "Manage" gap.

Related reading: Mapping NIST AI RMF to Configuration Management: A Crosswalk for Compliance Teams (Part 4) →

Player not loading? Listen on Substack →

Two production wipes in two months, same root cause both times. Instead of another retrospective, a four-perspective council — engineer, architect, security reviewer, operator — each with standing to object. Structured disagreement produced a four-rule protocol (WIP=1, pre-flight gate, stop conditions, never edit on main) that lives in the harness, not in human memory.

Related reading: I Was Done Trusting the Process. So I Convened a Council. →

Player not loading? Listen on Substack →

The most common failure mode in production AI is treating a Key Performance Indicator as if it were a Key Risk Indicator. KPIs measure whether an agent is producing value; KRIs measure whether it is still operating within approved boundaries — monitored continuously, paired with thresholds, and escalated through a three-tier model borrowed from 30 years of financial-services operational-risk discipline.

Related reading: KRIs vs KPIs for Autonomous Agents (Part 3) →

Player not loading? Listen on Substack →

The operational record behind the post — the four-phase build that turned Claude from present-but-unused into staff: map the work into Claude zones and human zones, write the rules every session inherits, wire a control plane the agents read at session start, then measure what works and prune what does not.

Related reading: From Chaos to Cadence: How We Built Prism's Claude Operating System →

Player not loading? Listen on Substack →

The missing twin of agentic governance. Four disciplines ported from enterprise IT and financial-services operational risk — Baseline, Change Control, Drift Detection, Audit and Remediation — and the Key Risk Indicator tier model that lets governance run at the speed of the agent.

Related reading: Configuration Management for AI Agents (Part 2) →

Player not loading? Listen on Substack →

The latency-gap argument and the observability-theater framing — why the disciplines that close the gap already exist in financial-services KRIs and enterprise IT configuration management.

Related reading: Observability Isn't Governance (Part 1) →

Player not loading? Listen on Substack →

A first-person field case — how a self-imposed, fabricated constraint moved the Prism Chief of Staff agent outside its baseline for a full session with no governance signal.

Related reading: A Hallucination on Our Own Stack (Bonus) →

Player not loading? Listen on Substack →

Orientation for the series. The central problem in agentic AI and why configuration management discipline is the operational layer that closes it. Start here if the CM-AI Framework is new to you.

Player not loading? Listen on Substack →