You cannot automate what you have not standardized.
The frameworks below are the operating system behind our work: why most transformations fall short, the sequence that fixes them, and where AI actually earns its place.
The plain-English guide to using AI you can trust
You do not need to know what AI is to read this page. That is the point of it.
The Governed AI Lexicon
Forty-three terms you will meet in an AI governance conversation, each with what it means, what an institution actually worries about, and what it costs you.
The Governed Agent Stack
Nine controls in three bands. Each one named with the failure it prevents, what it costs, and how you would know it had stopped working.
The ESSA principle
Four gates every process must pass before technology touches it.
The Process-to-AI Pyramid
A three-layer model for sequencing transformation so AI lands on a foundation that holds.
The Process Maturity Model
A five-level diagnostic that tells you whether a process is ready to automate, before the budget is committed.
The Four Checks
Four checks every AI output must clear before anyone acts on it.
Articles
Longer-form writing from the practice, published with sources and dates.
First-pass yield for AI systems
Manufacturing settled the measurement problem AI teams are currently arguing about. The metric transfers without modification.
9 minReadYour AI program has a vocabulary problem
Four frameworks, four different jobs. Most organizations pick one, use it for everything, and discover the gap in a conversation they cannot reschedule.
9 minReadThe retrieval questions nobody asks until it is too late
Everyone argues about which vector database to use. Nobody asks what happens to the embeddings when the document is deleted.
10 minReadEntitlement is not provenance: the perfectly cited breach
The answer named its source, linked the document, and gave the page number. The person reading it was never cleared to see that document.
9 minReadThe blindfolded judge: what our eval harness was actually measuring
The first real run scored 0.227. The agents were fine. The harness had never shown them the code.
8 minReadWhat the format ate: when a better prompt starts deleting things
The cap was three themes. The fourth was air conditioning, two guests, summer coming, and nothing in the output said it had been cut.
8 minReadReliable, and still wrong: testing AI like a measuring instrument
Thirty for thirty on the closed sort. The ranking flipped between runs, and no run said the top two were tied.
7 minReadWhere LLMs actually fail: composition, not retrieval
Retrieval never failed. Every verified error was an addition the source never made.
8 minReadWhere AI earns its place
Five questions locate the work. One rule places the human gate.
8 minRead