Team design is the layer of an AI programme that ships no artefact by default, which is why nobody audits it. This catalogue fixes that by admission rule: an item earned a place only if a reviewer could ask for the object that proves it and be handed one.
A register entry qualifies. A directory group, an access control list, a dated drill log and a signed approval all qualify. A principle does not, and neither does a completion percentage. Twenty-eight items, seven families, from the on-call boundary to the legal interface.
The catalogue runs to seven families. These three are where a review stalls first, and not one of them is settled by a statement of intent. Each resolves to a document with an owner, a date and a last-modified timestamp, which is why item numbers stay stable and are worth citing in a ticket.
ORG-1 wants one named owner per production system recorded as a directory identity rather than a job title, with a named deputy and a date the entry was last confirmed. The owner holds written authority to suspend without a further approval, and that authority is drilled on a schedule. The artefacts are a register entry, a delegation record and a timed suspension log.
ORG-3 is a reporting line, and a reporting line is a structure with a diagram. Write access to the evaluation set and its labels belongs to a function that does not report to the delivery lead, evaluation is a funded headcount rather than a rotation that lapses when delivery is behind, and a threshold change records its approver, its date and its reason.
ORG-2 treats the platform as a product with a named owner, a published roadmap and consumers allowed to decline it, not as a ticket queue. DORA put developer independence at a 5% productivity gain in 2024, and a queue is how that gets spent. Measure from the consumer side: time to a new team's first successful production deploy.
A title moves with a reorganisation and a directory group does not. Add a named deputy and the date the entry was last confirmed, so a register nobody has touched in a year declares itself.
Groundedness, refusal rate or task success with named thresholds, mapped in an alert routing table to the person they wake. Availability monitoring will stay green through the entire incident.
Prompts, system instructions, decoding parameters and retrieval configuration go under version control and the same review rule as code. Behaviour edited in a production console has no diff and no reviewer to name.
Trace, prompt, retrieved context, model version and release bundle captured under a retention class before service is restored, held by a named preservation role with its own place on the rota.
ORG-1 and ORG-6, in that order. Ownership decides whether anybody can stop a system, and incident roles decide whether the evidence survives the fix. Both are cheap to write down and expensive to discover missing, and the other five families are easier to fund once those two are settled.
A value changes nothing at three in the morning when a system is answering fluently and incorrectly. An item made the list only if a reviewer could ask for the object proving it and be handed one, which rules out charters and training completion rates and leaves registers, rotas, access control lists and drill logs.
DORA identified seven archetypes across nearly 5,000 technology professionals in 2025, and a team simultaneously building the platform, shipping features and holding the pager matches none of them. Declaring the archetype and the capacity split is how that gets named out loud rather than absorbed quietly by three people.
Probably more, not less. IBM found 63% of organisations with no policy at all in 2025, which is the easy half. The hard half is an organisation that wrote one and still cannot name the individual it binds, the group that individual belongs to, or the last time they did anything it describes.
Exercise. A delegation record says somebody may suspend a system; a timed drill log says they did, on a date, against something that was serving traffic. ORG-1.2 asks for both, and the second one is what separates an owner from a name in a spreadsheet.
Only where a duty turns into a structure. Engineering Team Design for AI Under Regulation sets out why Regulation (EU) 2026/1744 assigns duties to people rather than to architecture, which duty lands on which role, and what evidences it.
Take that first if a supervisory deadline is driving the work, and this catalogue to build.
VPs of engineering and the people who hold their org chart, plus the risk and legal counterparts named in ORG-7. It assumes you have AI systems in production and want item numbers to cite in a review note rather than a case for why team structure matters.
Drop your details and we'll send Engineering Team Design for AI: An Engineering Reference straight to your inbox - no spam, unsubscribe anytime.
A two-week trial sprint pointed at ownership and incident roles tells you whether anyone named in your register could suspend a system today, and what evidence would still exist afterwards. Two weeks, your repository, zero commitment. SECTION 7 - FAQ - 5 to 8 questions
Book a team design review