The National AI Governance Playbook · Part III · Chapter 9

Who leads the design phase

The design phase needs an owner, and the owner's properties matter more than its address: authority to convene, neutrality among future deliverers, access to technical evidence, and an end date. This chapter surveys the candidate homes, writes the mandate on one page, and closes Part III at the handover.

By Myriam Ayada · MindXO · July 2026

In brief

The design phase of Chapter 2 needs an owner. Four properties matter more than the owner's address: authority to convene ministries, regulators and the standards body; neutrality among the entities that will later deliver; access to technical evidence; and an end date with a handover, because a design phase that does not end becomes a rival to the system it designed. Candidate homes include a centre-of-government unit, the digital ministry, an interministerial committee with a staffed secretariat and an external commission with a reporting date. Each supplies some properties and lacks others, and a one-page mandate states the agenda, the artefacts owed, the access rights, the end date and the destination of each artefact at handover.

A phase, and therefore an owner

Chapter 2 followed governance programs through four states and located the recurring failure in a skipped design stage: delivery assigned straight from strategic intent, with scope, roles and funding resolved piecemeal inside the delivery entities. An adjacent failure produces the same symptoms at a slower tempo. The design stage exists, a committee meets, papers circulate, and nobody owns the result. The five questions of Chapter 6 are necessarily answered by many hands, since ministries hold sector knowledge, regulators hold enforcement practice, the standards body holds the technical catalogue and the finance ministry holds the budget. The answers can be distributed; the pen cannot.

One entity must hold the pen, and holding the pen has a specific content. The owner convenes the table and sets the calendar. It drafts the artefacts and controls their versions. It keeps the decision records, each choice with its evidence and its dissent, so that a later dispute reopens a file rather than a negotiation. And it hands over cleanly when the phase ends, with every artefact delivered to a named recipient. This chapter treats the entity that carries those duties: the properties it needs, the candidate homes that supply them, and the mandate that makes the arrangement explicit. It closes Part III.

The four properties

Jurisdictions have run the phase from cabinet offices, ministries, committees and commissions. Outcomes track four properties more closely than they track the address.

Convening authority. The owner can put ministers, sector regulators and the standards body at one table and set deadlines they keep. The authority is usually borrowed from above, through a head-of-government mandate, a cabinet decision or a budget line the participants need. An owner limited to invitations runs a correspondence exercise, and the phase advances at the pace of its least interested participant.

Neutrality. An owner that will later run a delivery program has an interest in every boundary the design draws, and boundaries drawn by an interested party tend to settle around that party. This is the failure mode of Chapter 2 arranged in advance, inside the stage meant to prevent it. Where full separation is unavailable, the compensation belongs in the mandate: published decision records, an external reviewer and ratification of boundaries above the owner's level.

Access to technical evidence. Design decisions about capability thresholds, testing obligations and cited standards rest on claims about what systems can do. The owner therefore needs a working line to the evaluation capability of Chapter 10 or, where none exists yet, to its precursor: a supervisor's testing unit, a national laboratory, a standing technical panel. Without that line, thresholds are set by analogy and revised by negotiation.

A deadline and a handover. The phase produces the five artefacts of Chapter 2, delivers each to a named recipient and dissolves into the governance of delivery. A design phase that does not end becomes a rival to the system it designed: entities that should be delivering keep referring decisions back to it, and it keeps accepting them. The end date and the handover plan belong in the mandate from the first day.

Sheet 09 of 13, the owner and the handover that ends the phase: the design authority sits at the centre carrying the four properties, convening authority, neutrality, technical evidence and an end date, with spokes to the ministries, the sector regulators, the standards body, the evaluation capability and the legislature. Below, a timeline runs from the mandate to a handover marker where the five artefacts transfer, the allocation of functions to the delivery entities, the legal instrument to the legislative calendar and the measurement baseline to the measurement owner, and the owner dissolves.

Candidate homes

Four homes recur in practice. None supplies all four properties, so the choice turns on which property the jurisdiction can import through the mandate and which it can obtain only from the address.

H1. A centre-of-government unit: cabinet office or equivalent

Strengths: the highest convening authority and the cleanest neutrality. It delivers no program, and it speaks with the authority of the head of government. Limits: the thinnest technical depth. Evidence arrives through the evaluation function, secondments or commissioned studies, and the unit's attention is contested by every other national priority. Fits when: ministries are many and rivalrous, and the principal design risk is that no boundary drawn between them would otherwise hold.

H2. The digital or AI ministry: where the strategy already sits

Strengths: the strongest technical adjacency and the deepest staffing. The teams that know the systems, the vendors and the standards work already sit there. Limits: the weakest neutrality where the ministry will also deliver. An owner in that position draws boundaries around itself, the pattern Chapter 2 identifies. Fits when: one ministry clearly leads on AI and delivery sits elsewhere, with the sector regulators and agencies rather than inside the ministry itself.

H3. An interministerial committee with a real secretariat: the secretariat is the owner

Strengths: legitimacy spread across every ministry with a stake. Decisions arrive pre-negotiated, and the design survives turnover at the top of any single ministry. Limits: it moves at the speed of its slowest member, and a committee cannot hold a pen. The secretariat is the actual owner, so it must be staffed, funded and named in the mandate. Fits when: the constitution distributes power and no single ministry can convene the others, so ownership needs a collective umbrella with a working core.

H4. An external commission or taskforce: reports by a date

Strengths: independence from every future deliverer, undivided attention and a reporting date built into its terms. Limits: no standing authority once it reports. Its recommendations bind no one until a government owner receives them and carries them forward. Fits when: the diagnosis and the design need distance from incumbent interests. It fits the analysis and the drafting, never the handover alone.

Public practice illustrates the range. The European Commission ran the design through a white paper and impact assessment before proposing legislation, holding the pen at the centre and commissioning the technical input. The United Kingdom ran the phase from a department, the Department for Science, Innovation and Technology, through a white paper and consultation cycle, with delivery assigned to existing regulators, which kept the owner at one remove from enforcement. Singapore ran finance-sector design from the supervisor itself: the FEAT principles and the 2025 consultation on AI risk management guidelines both came from the Monetary Authority of Singapore, while the digital ministry held the horizontal pieces. Each choice reflected where convening authority and technical depth already sat.

The mandate on one page

Whatever the address, the arrangement holds only when it is written down, and one page suffices. Six elements belong on it.

The agenda. The five questions of Chapter 6, stated as the questions the phase exists to answer.
The artefacts owed. The five artefacts of Chapter 2: drawn boundaries, an institutional model, a funding decision, a legal vehicle and a measurement baseline, each with a due date.
Authority over decision records. The owner keeps the record of every design decision, with its evidence and its dissents, and the record survives the phase as the reference for later disputes.
Access rights to evidence. A stated line to the evaluation capability of Chapter 10 or its precursor, including the right to commission tests and to receive results in full.
An end date. The date the phase dissolves, extendable only by the authority that issued the mandate.
The handover. Where each artefact goes: the allocation of functions to the delivery entities, the legal instrument to the legislative calendar of Chapter 8, and the measurement baseline, the future scorecard, to the measurement owner of Chapter 12.

Common failure mode. The permanent design committee. A committee without a secretariat produces minutes; a committee without an end date produces a standing veto. Both outlast their usefulness. Delivery entities wait for guidance that never lands, then draw their own mandates, and the overlaps of Chapter 2 return with a committee attached.

The handover closes Part III. The phase has its questions from Chapter 6, its sequence from Chapter 7, its legal vehicle from Chapter 8 and, with this chapter, its owner and its end date. Part IV turns from deciding to building. Chapter 10 opens with the capability every earlier chapter has leaned on, the safety and security engine: the evaluation and testing function that supplies the evidence the design phase consumed and the assurance the rulebooks will cite.

Three questions for every government

Which entity owns the design phase, and does its mandate state the five questions, the five artefacts and an end date?
Is the owner separate from the entities that will deliver, and where it is not, what compensates?
What, precisely, transfers at handover, to whom, and what dissolves?

Selected public sources

White Paper on Artificial Intelligence: a European approach to excellence and trust, European Commission, 2020
A pro-innovation approach to AI regulation, white paper and government response, United Kingdom, 2023 and 2024
FEAT Principles, Monetary Authority of Singapore, 2018
Consultation Paper on Guidelines on Artificial Intelligence Risk Management, Monetary Authority of Singapore, 2025