Newsletter

The Archive

Every issue takes one real question about trusting AI and works it through: the oversight, governance, and reliability of autonomous AI agents that act on their own. No news roundups, no hype cycle, no recycled takes. Read a few. If the thinking is useful, subscribe and the next one comes to you.

  1. 022 Proof of cost vs proof of value: how agentic AI pilots become money pits

    Proof of cost shows up on its own. Proof of value only shows up if you built it in.

  2. 021 Why an AI's confidence is the least reliable signal you have

    An AI sounds equally sure right or wrong, so its confidence tells you nothing. Here are the signals that do, ranked by how far from the model they come.

  3. 020 The audit trail is the product: logging what, why, and on whose behalf

    An AI agent's work is gone in seconds. The log is all you keep. If it can't say what the agent did, why, and on whose behalf, you don't have an audit trail.

  4. 019 Where to spend your scrutiny: triaging AI oversight by stakes

    Stakes live in what an output can reach.

  5. 018 The 40% that get canceled: why most agentic AI projects die

    The project is not killed because the agent failed. It is killed because nobody can prove it did not, at a price anyone can state.

  6. 017 Self-learning AI breaks the one thing that made automation safe

    Learning and permanent validation are mutually exclusive.

  7. 016 AI guardrails that prevent, not just detect

    An AI agent deleted a production database in 9 seconds with its safety rules in plain view. Here's what separates AI guardrails that stop from ones that watch.

  8. 015 AI compliance, read by a builder: the hard part isn't the paperwork

    A log you didn't write in 2026 doesn't exist in 2027.

  9. 014 A reviewer who can't reach the answer another way isn't oversight

    A reviewer without an independent path to the answer records presence, not judgment.

  10. 013 You can see what governance costs. You can't see what it saves.

    Governance carries a visible price and an invisible return, which is why it loses budget arguments it should win.

  11. 012 The principles AI governance is "discovering" are old automation rules

    The principles AI governance keeps discovering are decades-old automation rules. If the principles are old, they're proven, there's a playbook, and the genuinely new problems get sharp.

  12. 011 Verification you can afford: making checks cheap enough to happen

    Verification dies on cost, not intention: a check priced in human attention stops happening. Checking is cheaper than producing, so push every check down the cost ladder and spend judgment last.

  13. 010 Who's accountable when an autonomous agent makes the call?

    When an autonomous agent makes the call, the deploying company answers for it. "The AI decided" is not a defense anywhere it has been tried.

  14. 009 Human-in-the-loop is becoming human-on-the-loop. What you give up

    On-the-loop is not in-the-loop with less work. It's a different control model. The stakes pick the loop, not the workload.

  15. 008 NIST's agent standards are forming now. Design to them early

    The standards aren't final, but the direction already is. A standard's final text is a cleanup of its drafts; the direction almost never reverses.

  16. 007 What the EU AI Act's August 2026 obligations actually demand

    August 2026 didn't get canceled. It got split.

  17. 006 How to prove meaningful control to a regulator

    Meaningful control is not a claim you make. It's evidence you can produce.

  18. 005 Governance isn't slowing your AI down. It's what lets you go faster

    Governance was never what slowed AI down. The absence of it was.

  19. 004 What 25 years making systems trustworthy taught me about AI oversight

    An AI agent with no named owner doesn't have oversight. It has hope and a logo.

  20. 003 I spent two days proving my own thesis on an AI

    Confident is not correct. Complete-looking is not complete. Make it prove it.

  21. 002 You can't verify everything, so stop pretending you will

    The verification is the work now.

  22. 001 They sound exactly as sure when they're wrong

    Stop asking whether the answer sounds right. Start asking how you'd know if it weren't.

That's the catalog so far. New issues land in your inbox before they land here.