Insights

Notes on operations, AI, and running the place.

Essays on what actually breaks inside an operation, what AI changes about it, and what it takes to lead through either. Most of these appeared first on LinkedIn.

AI · October 2026 · 2 min read

Start with the task you do forty times a month

Most agents start AI with the listing description, the least repeatable thing they write. The better first task is the one you do forty times a month, written down first.

AI · June 2026 · 2 min read

Asimov wrote the Three Laws as a trap

Asimov didn't write the Three Laws as a solution — he wrote them to show how a clean framework fails in practice. Anyone deploying AI now faces his real question: what should the rules be?

AI · June 2026 · 2 min read

The Three Laws never required the truth

None of Asimov's Three Laws require a machine to tell the truth. The same gap shows up in AI tuned to agree with you — and a tool that can't disagree isn't a thinking partner.

AI · July 2026 · 2 min read

Why a ranked list of rules can't govern AI

A ranked list of absolutes breaks in the real world, as Asimov showed. A workable AI framework holds principles in tension — and lives in your people and policies, not just the machine.

AI · July 2026 · 2 min read

AI safety isn't the vendor's job or your people's job alone

Businesses bet AI safety on the vendor or on their own people. Both bets fail in opposite directions, and the failures land in the gap between them that nobody owns.

AI · July 2026 · 3 min read

The three layers of an AI governance framework

Responsible AI has to live in three layers at once: the vendor's, your people's, and your organization's. The middle one is the only one you own, and the one most companies skip.

Operations · July 2026 · 3 min read

Automation is an amplifier

Layer AI onto a flawed workflow and it doesn't find the flaw — it runs it faster and at volume. Before choosing an AI tool, find out what your process does when it breaks.

Operations · August 2026 · 3 min read

The unexamined process

Writing an undocumented process down exposed steps nobody would choose; redesigning it cut repair turnaround 75%. Put AI on an unexamined process and you make it permanent.

Operations · September 2026 · 2 min read

Exceptions don't leave a trace: why you can't map an operation by instrumenting it

The real operation lives in its exceptions, and exceptions leave no trace. They go undocumented because it was never anyone's job, which no amount of observation fixes.

Operations · September 2026 · 2 min read

In 2018, the FDA said yes

The FDA cleared the first autonomous AI diagnostic in 2018. Adoption is still limited, and everything that sat in between was operations work, not regulation.

AI · July 2026 · 3 min read

The Governance Gap

Private AI conversations turned up in Google results because users assumed the platform handled sharing and the platform assumed they did. That gap is governance, not security.

AI · July 2026 · 3 min read

The Gates Nobody Showed You

Most people building business software with AI never saw the checks a developer used to carry. Six questions cover what went missing when procurement stopped gating the build.

Operations · August 2026 · 3 min read

The workaround that worked

Every screen that failed got restarted by hand, and it worked every time. That's why nobody found the ceiling underneath — the tell is labor. The network went from 500 screens to more than 12,000.

AI · September 2026 · 3 min read

Read the source, not the thread

Two viral AI stories said more than their sources did, and one got the direction backwards. Before you decide on an AI claim, read the original and find what the author says they don't know.

Operations · September 2026 · 3 min read

The one-page AI rule

Most small-company AI policies are a feeling, not a rule. One page that names the approved tools, the account, what never goes in, and who to ask does more than twelve pages in a drive.

AI · September 2026 · 3 min read

The middle layer, in a chart

When OpenAI connected ChatGPT to Epic, the coverage covered the vendor's limits and the clinician's judgment. The middle layer — the settings a hospital owns — barely came up.

AI · September 2026 · 3 min read

Who checks? Microsoft's AI code of conduct and its missing half

Microsoft's new AI code of conduct sets a real behavioral standard. It doesn't say who enforces it, and it doesn't cover most of the AI Microsoft sells — the same gap as Asimov's laws.

Operations · September 2026 · 3 min read

Who notices? OpenAI's six AI incidents, read as an operator

OpenAI disclosed six cases of its models misbehaving. Inside a business they'd be ordinary incidents, and only your incident response policy decides whether anyone finds or reports them.

Operations · September 2026 · 3 min read

You onboarded the person. Not the agent.

An AI runs the therapy session and a clinician watches. The selling point is a supervision ratio, and it is an operations question your staff onboarding already answers.

AI · October 2026 · 3 min read

A serial number and a version number

A bed kept one serial number for life while its parts changed. An AI agent needs the same two things: which agent, and what version it was running at the time.

AI · October 2026 · 3 min read

Who has to report it

Medical devices split the reporting duty by who can see what. AI has the maker's half starting and the operator's half missing, and nobody is obligated to write anything up.

Operations · October 2026 · 3 min read

Usage dashboards measure the wrong layer

A usage dashboard counts what the vendor can instrument: seats, sessions, tokens. Whether the work got faster or cleaner lives in your process, and nobody sells a report for that.