Field Notes
Operational writing for AI in production.
Governance, supply-chain security, agent ops, observability. Written for the IT executive whose org is past the AI demo and into the audit review.
- 4 min readAI DistributionMarketing
Your website's next reader is a model.
Buyers increasingly meet your company through an AI assistant's answer, not your homepage. The assistant read your site — or failed to — and composed the first impression you never got to make. Machine readability has quietly become a distribution channel, and almost nobody owns it.
- 4 min readAI EnablementWorkforce
License count is not adoption.
"We rolled out AI to two thousand employees" is a procurement metric wearing an adoption costume. Usage clusters into a small power-user tail, a large occasional middle, and a silent bottom — and the ROI of the whole program lives in moving the middle. You can't move what you don't measure, and logins don't measure it.
- 4 min readAI GovernanceSecurity
You don't need an AI platform. You need ten boring controls.
The procurement reflex says: big new risk category, so issue an RFP for a big new platform. Meanwhile the actual gaps in most AI operations are small, specific, and closable in days each. Here are the ten controls that matter — none of which requires a nine-month rollout or a seven-figure commitment.
- 4 min readAI GovernanceOperations
Stop writing AI policies. Start writing AI runbooks.
Every organization now has an AI policy — a PDF full of principles about responsible use and human oversight. Ask what happens at 2am when an agent deviates from its workflow, and the PDF goes quiet. Policies describe intentions. Runbooks describe actions. Right now the industry is drowning in one and starving for the other.
- 4 min readAgent OpsAI Governance
Your agents are already talking to each other. Nobody provisioned the channel.
Multi-agent coordination didn't arrive as an architecture decision. It arrived as a markdown file in a shared repo, a ticket one agent writes and another reads, a scratch document two teams' assistants both edit. The messages are flowing today — over transports with no identity, no access control, no retention, and no monitoring.
- 4 min readAI GovernanceKnowledge Management
Prompt sprawl is the new shadow IT.
The prompts that draft your customer emails, summarize your contracts, and screen your inbound leads live in personal notes apps, Slack DMs, and browser bookmarks. Unversioned, unreviewed, unowned — and gone the day their author resigns. That's not a knowledge-management quirk. It's business logic running outside every control you have.
- 4 min readAI ActCompliance
The AI Act's high-risk obligations land August 2. Your agents qualify more often than you think.
On 2 August 2026, the EU AI Act's obligations for high-risk AI systems start applying. Most of the coverage is legal analysis aimed at model providers. If you deploy AI inside hiring, credit, or essential-service workflows — including agents you assembled in-house — the operational checklist is yours, and it's shorter than the legal memos suggest.
- 4 min readAI GovernanceLeadership
"Are we doing AI right?" deserves a better answer than a demo.
Every board is asking the question. Most CIOs answer with adoption numbers and a well-rehearsed demo — which answers "are we doing AI" and says nothing about "right." Here are five operational dimensions you can actually score, and what red, yellow, and green look like on each.
- 3 min readAgent OpsAI Governance
Your agents pass their evals and still break your processes.
Eval scores measure whether the model produced a good answer. Production incidents are almost never about bad answers — they're about skipped approvals, missing audit entries, and steps executed out of order. Task success and process conformance are different properties, and only one of them is on your dashboard.
- 4 min readMCPAI Supply Chain
The MCP server you approved is not the MCP server you're running.
Security review signed off on the MCP server in March. It's July, and the server has shipped four updates since — new tools, widened schemas, rewritten descriptions — none of which anyone re-reviewed. A point-in-time approval of a surface that mutates continuously is not a control. It's a memory.
- 5 min readIAMMCP
Every MCP connection is a service account no one provisioned.
MCP connections behave like long-lived service principals — broad scope, no owner, no expiry, no review. IAM teams spent a decade learning that lesson once. The agent stack is unlearning it in real time.
- 4 min readAI Supply ChainSecurity
The AI supply chain stopped being about model providers.
When your team connects an MCP server to Claude or ChatGPT, third-party tools, prompts, and resources land directly in a model's working memory. That's a software supply chain — and almost no one is inventorying it yet.
- 5 min readAI GovernanceAgent Ops
Your AI agents act in production. The audit trail does not.
Agents now ship code, file tickets, send email, move money. The systems they touch keep their usual logs — but the agent's reasoning, the tools it considered, and the version of the prompt it was running rarely make it anywhere durable. That gap is where the next class of incidents will be diagnosed, after the fact, with no data.