Your delivery metrics describe the system. This scores the shape holding it up, which is a different question and usually gets a worse answer. One engineering organisation, out of a hundred points, across accountability and authority, the platform and the path to production, the evaluation and review function, and competence and continuity.
Twenty rows, each naming the artefact that settles it, and three bands. It is not a replacement for lead time or change failure rate, and nothing in it tries to be.
Four domains of unequal weight will not average into anything useful, so do not try to make them. Read the heaviest row scoring zero, then place the total in one of three bands. Each band is a forecast of what breaks next rather than a grade for what you have built so far.
Seventy-four to one hundred. The structure survives a departure, a reorganisation and an audit without anybody improvising on the day. The points still missing usually sit in the evaluation function, where an organisation this capable often still has the people doing the measuring reporting to the people being measured. Those are the expensive rows to buy back.
Forty-five to seventy-three, which is where the worked organisation lands. It works, and it works because four or five people know things that are written down nowhere. Treat the score as a forecast of what breaks when one of them changes team. Open the heaviest zero in accountability first, since authority is the one thing that cannot be pushed downwards after the fact.
Under forty-five. The appointments exist, the policy exists, and nothing underneath either has ever been exercised. Start at the top of accountability and stop scoring the rest until it clears twenty. No other domain can be read honestly while the question of who may stop a system is still unanswered.
Accountability carries thirty points because every other domain is decorative where it fails. Ask for a dated record of a production system taken offline by one person, with no second approval anywhere in the chain.
Take the last fifty and count how many needed a request raised on somebody else. Then measure the working days from a new team starting to its first AI feature carrying real traffic.
On the current chart it should terminate outside delivery, with the manager named. Then ask for one release stopped on evaluation evidence in the last two quarters, and what happened in the week afterwards.
Pick a system whose original engineers have all left, open the pack and have the current owner confirm it works. Continuity is the cheapest domain to repair and the one left latest.
No, and it is built not to. Lead time, deployment frequency, change failure rate and time to restore measure what the system does. This measures whether the organisation is shaped to keep it doing that.
Run both, since delivery figures held up by four people who have not resigned look identical to healthy ones until they stop.
Once per organisation. What is being scored is the shape: who answers for a system, who may stop it, who measures it, and what survives the next reorganisation. Scoring team by team produces an average of local arrangements and conceals the fact that only one team has ever written any of it down.
Every other domain is decorative where it fails. Where nobody may take a system offline without asking somebody who is not in the room, the platform rows, the evaluation rows and the continuity rows together describe a well-organised inability to act.
Continuity carries twenty for the opposite reason, being the cheapest to repair.
An organisation that has written everything down and exercised none of it. Nineteen teams, eleven AI systems in production, four of them with a natural person named. Fifty-two points go unawarded and eighteen of those sit on rows no document can satisfy.
The figures illustrate a plausible shape rather than report an observed organisation.
It asks for one natural person per system, by name, and a function or a rota or a committee counts as a blank. Human oversight under Article 14 still requires competence, training and authority in an individual, and Regulation (EU) 2026/1744 left that sentence alone when it came into force on 27 July 2026.
Take the heaviest zero first, then check whether the rest share a cause. Engineering Team Design for AI: An Engineering Reference is the catalogue you build from, with the organisational control families and the artefact that proves each one.
This instrument says where to start, and that one says what to put there.
A VP of engineering, plus whoever holds the oversight appointment on at least one production system. The second person settles arguments quickly, since most rows ask whether an event happened rather than whether a policy exists, and only an appointee knows which Tuesday it was.
Drop your details and we'll send Benchmarking Engineering Team Design for AI straight to your inbox - no spam, unsubscribe anytime.
Bring the system register and the name sitting against each oversight appointment. We will score two domains with you and point at the heaviest zero that an event, rather than a document, would close. SECTION 7 - FAQ - 5 to 8 questions
Book a scoring session