Your model will fail. The question is what catches it.
No part of an airliner is as safe as the airliner. Aviation stopped trying to perfect the component fifty years ago and built the system around it instead. AI safety is still trying to perfect the component.
Crox is a safety management practice for deployed AI. We bring aviation’s methods, confidential reporting, just culture, layered defences and investigation that asks why a decision made sense, to the model you have already shipped.
Guardrails, safety training, evaluations, red teams. Almost the whole safety effort is aimed at making the model fail less. The numbers improve. They never reach zero, because a filter on a probabilistic system is itself probabilistic.
Meanwhile the harm arrives through the system around the model. Nobody watching the output against policy. Nobody with the standing to stop a release. No channel for the engineer or the clinician who saw it nearly go wrong and fixed it quietly.
“The model hallucinated” is this decade’s “pilot error”: usually true, never sufficient, and a verdict that closes the inquiry instead of opening it.
You own a model that is already in front of patients, customers or underwriters, and one of these is on your desk this quarter:
- An LLM feature about to ship into a regulated workflow.
- A customer, insurer or diligence process asking what controls exist around the model.
- A board, regulator or clinical safety officer asking who is accountable when it gets one wrong.
If none of these are true yet, the paper is the right place to start. If one is, two days with one team and one model finds out which of your defences are real.
White paper
Loss of Situational Awareness
Why fixing the model will not make AI safe, and what aviation did instead. What a half-century of aviation safety built, what transfers to deployed AI, what does not, and the one thing AI lacks that aviation built first, in 1976. Free to read, free to forward.
Read the paper →Observe how the model is actually run. Then build the system around it.
Not an audit against a standard. Aviation’s line operations safety audit puts a trained observer in the jumpseat on routine flights and counts what nearly went wrong. We do the same for one team and one model, under terms that make it safe to tell us.
Start here
The Safety Diagnostic
One team, one model in production · Two days on site · £12,000 fixed
Aviation audits normal operations, not paperwork. We map the five layers around your model, interview the people who have caught it nearly going wrong, and hand you the holes that line up. Non-jeopardy terms in writing: nothing is attributed to a person, nothing goes to HR.
Learn more →Then
The Standing Review
Monthly · For teams that have had the diagnostic · From £8,000 a month
A monthly near-miss review and incident investigation on call. The part of a safety system that decays first when nobody owns it. Offered at the readout, not before.
Learn more →Also
The Talk
Product and engineering teams · 45 minutes
What pilots do about mistakes that AI teams don’t. Four habits that transfer, no pitch in it. Free for a handful of companies while it is being sharpened.
Learn more →Most people writing about AI safety have read about safety cultures. Adam has shipped the AI and flown the plane.
A decade shipping AI where wrong answers have consequences: the UK’s first BSI-audited clinical chatbot at Babylon Health, built under ISO 13485, MDR and DCB 0129. NHS risk stratification that cut unplanned readmissions by 35%. Automated underwriting. Today, Head of Product at SideLight AI, delivering AI work for Tier 1 and Tier 2 law firms.
Before that, five years flying 737s for Ryanair: filing confidential reports on his own errors, working under a regulator, inside the safety system this method comes from. He has sat on both sides of the safety case.
We build what we argue for.
The flight data recorder does not ask the pilot what happened. Neither should the system around a model. These are the working tools behind the method.
BBX
Product · bbx.crox.io
A flight data recorder for AI. Every request and response captured, so when someone asks why it did that, the answer comes from the record, not from memory.
Take a look →rogerwilco
Open source · PyPI
Before an agent does something it cannot undo, check what it says it is doing against what is actually on disk, and refuse if they differ. Built from about five thousand experiments on agents handing each other payments.
Take a look →DODAR
Research · dodar.crox.io
A structured-reasoning benchmark across eight models. The cockpit decision loop applied to language models, with the method and results published in full.
Take a look →“Adam cares deeply about the quality and impact of his work. Never shirks the tough conversations or sharing an unpopular truth.”
— Jasel Mehta, COO
“He was professional, holistic and inclusive in his approach. I would highly recommend on projects which need his eye for detail.”
— Scott Beat, Quality Manager
Which of your layers are lined up this quarter?
Evals, guardrails, human review, monitoring, rollback. Most teams have two of the five and believe they have four. Two days with one team and one model finds out which. The intake call is thirty minutes: pick the team and the model, agree the terms, fix the dates.