Seven principles. None is negotiable.
I did not invent them for AI. I took them from twenty years of infrastructure, where a mistake ends in a regulator inspection or a laboratory standstill. Today I run people, servers and AI agents the same way. Each principle comes with two pieces of evidence: one from infrastructure, one from work with AI.
-
01
The frame comes from me. Execution from the team. Accountability stays with me.
I set the class of solution and the boundaries. I hand over design and execution. I never hand over the result.
InfrastructureSixteen people in three countries and five locations. Two voluntary departures in five and a half years.
AII run a team of 24 agents like a team of people: each has an owner, a scope and a ban on changing someone else's work. The plan goes on screen before the first line of code is written.
-
02
Regulation from the first decision, not at the end.
Security is a property of the design. Asking “don't touch this” is not a safeguard.
InfrastructureFDA inspection without a single finding for IT. ISO 27001 programme from 28 non-conformities to one. The business continuity model met NIS2 without changes on the day the directive took effect.
AII type passwords and keys myself, the model does not see them. Personal data and money sit behind a gateway. 29 of 30 data extraction attempts stopped measured, the acceptance condition is 30 of 30 target.
-
03
Verified or “unverified”. There is no third option.
“It should work” is not a system state. A backup nobody has restored is not a backup.
InfrastructureEvery year a full restore test of around twenty critical systems. One hundred per cent, signed off by the QA department, not by my own declaration.
AIAn agent may write “works” only after checking in target conditions. Otherwise it writes “UNVERIFIED” and what is missing. 25 of 25 figures in the report reconciled with the ledger measured.
-
04
A mistake gets written down, not hidden.
A mistake nobody wrote down will come back. Written down, it becomes a procedure.
InfrastructureI counted the experiments that failed because of IT: over two thousand a year. Dedicated lab support was set up. A few dozen remained.
AIA shared failure log for all my agents: what failed, what to do next time, what it cost. Every agent reads it at the start of work.
-
05
Change without stopping the company.
A maintenance window is a luxury. Usually there is none, so I design as if there were none.
InfrastructureThree acquired companies in two countries on one standard. Company split into three companies in three months. Server room move with no maintenance window and no outage.
AIEvery change in my environment has a record of how to roll it back. Nothing is deleted permanently: the agent moves it to a backup and says where.
-
06
Sized to the need, not oversized.
I raise the standard where regulation, scale or risk requires it. I do not change a working environment without a reason.
InfrastructureFrom networks on Linux routers for ten-person companies to a cluster of eight hosts and around 250 machines for a group. Same approach, different scale.
AIA small local model where data must not leave the building. A large one where thinking is needed. I match the model to the task, not the task to the model.
-
07
On myself first.
I do not roll out for anyone something I do not use myself every day.
InfrastructureMy own IT practice and around fifty companies in my care before I ran group infrastructure. My own idea taken to a European patent application, later cited by Cisco.
AISince May 2026 AI has been running my own infrastructure, mail, phone and car with me. Everything I show ran on my own setup first.