The National AI Governance Playbook · Part III · Chapter 6
The five design questions
The design phase resolves to five questions: who carries each function, in what order the regime arrives, where the evaluation capability lives and how it is funded, which instrument binds the design, and what keeps sectoral rules coherent. This chapter defines the five, treats the two that have no later chapter of their own, and shows how the answers interlock.
By Myriam Ayada · MindXO · July 2026
In brief
Parts I and II supplied the vocabulary: four functions, two regimes, and the six-link chain that carries a principle down to a control. Part III is the design phase itself, the work that produces the second state of Chapter 2, and everything that phase must settle compresses into five questions: who carries each function (allocation), in what order the regime arrives (sequencing), where evaluation and assurance live and how they are funded (capability), which instrument binds the design (legal anchoring), and what keeps sectoral rules aligned (cross-sector coherence). Answered together, on the record, the five produce the designed policy. Answered separately, or implicitly, they produce the overlapping mandates and unowned functions of Chapter 2.
From architecture to decisions
Part II closed with a complete architecture: the four functions of Chapter 3, the two regimes of Chapter 4 with their interface, and the six links of Chapter 5 running from principle to control. An architecture describes what a finished system contains. It leaves open every decision about a particular country: which institutions carry the functions, in what order the machinery arrives, with what money, and under which law. Those decisions are the design phase, the stage Chapter 2 identified as the second state of a national program, and they form the subject of Part III.
The agenda of that stage is shorter than it appears. Everything the design phase must decide compresses into five questions, and each of them acts on named links of the Chapter 5 chain. Allocation assigns the owners of links three to six. Sequencing sets the timing of links three and four. Capability builds links five and six. Legal anchoring selects the instrument at link three. Coherence governs the citations that join links four and five. A design stage that has answered all five, in writing, has produced the designed policy at link two. A design stage that has answered them separately, or left them implicit, has produced the overlaps and gaps that Chapter 2 catalogued.
The five questions
The five questions are set out below in the order in which they tend to arise. Each entry records what the question decides, the option space, the failure that follows when it is skipped, and where the playbook develops it. Two of the five, allocation and coherence, have no dedicated chapter of their own; the two sections that follow treat them in this chapter.
Q1. Allocation: who carries which function
Decides: which named entity carries each of the four functions of Chapter 3, and therefore who owns links three to six of the chain. Options: a single home for the three horizontal functions, separate bodies for each, or hosted units inside existing institutions; in each variant, use-regulation remains with the sector supervisors. If skipped: two programs claim the same function while another function has no owner, the overlapping mandates of Chapter 2. Developed in: this chapter, below, and Chapter 9 for the design stage itself.
Q2. Sequencing: in what order the regime arrives
Decides: the order in which the parts of the regime take effect: what binds first, what waits on capability, which sectors move early. Options: capability before obligations, obligations phased by sector risk, or general obligations with deferred application dates. If skipped: everything arrives at once, and obligations bind before the capability that evidences them exists. Developed in: Chapter 7.
Q3. Capability: where the engine lives, and on what budget
Decides: the institutional home of the safety and security engine, evaluation, testing and assurance, and the funding that sustains it. Options: a national institute, a unit hosted inside an existing body, or procured external capacity, funded by appropriation, levy or fees. If skipped: the engine defaults to nobody, and rulebooks demand evidence no institution can produce. Developed in: Chapter 10.
Q4. Legal anchoring: which instrument binds
Decides: the binding vehicle at link three of the chain, the point at which the design acquires legal force. Options: a dedicated statute, a decree, targeted amendments to existing law, or an existing statutory plan. If skipped: anchoring defaults to a future comprehensive law, and every obligation waits on its passage. Developed in: Chapter 8.
Q5. Coherence: how sectoral rules stay aligned
Decides: the shared frame of definitions, risk thresholds, reporting formats and standard citations that every sectoral rulebook draws on. Options: a named coordination body, the design authority of Chapter 9, or a published frame that rulebooks are directed to cite. If skipped: rulebooks drift, firms in several sectors face divergent duties for the same system, and the technical regime is reinvented per sector. Developed in: this chapter, below; the citation itself is the interface of Chapters 4 and 5.
Sheet 06 of 13, the five questions and the policy they produce: five cards, allocation, sequencing, capability, legal anchoring and coherence, each recording what it decides and the failure that follows when it is skipped, feed a single output, the designed policy, the second state of Chapter 2. Tags map each question onto the links of the Chapter 5 chain it acts on: allocation names the owners of links three to six, sequencing times links three and four, capability builds links five and six, anchoring selects the instrument at link three, and coherence governs the citations between links four and five.
Allocation, the first among the five
Allocation comes first because every other answer presupposes it. A sequencing decision needs to know whose readiness is being sequenced. A capability decision needs a home to fund. An anchoring decision needs to know whose powers the instrument must create. The unit of allocation is the function, and the working rule is symmetry: an owner per function, and a function per owner. A function with two owners produces the duplicated programs of Chapter 2; a function with no owner produces a gap that becomes visible only when its output is first needed.
Chapters 3 and 4 supply the rules of thumb. The three horizontal functions, standard-setting, evaluation and assurance, can share a single institutional home, since they attach to the same object and draw on the same scarce technical talent. Use-regulation stays with the sector supervisors, who already hold the inspection powers, the sanction regimes and the sector knowledge the function requires. The third rule is separation: the body that produces the evidence must sit apart from the bodies that act on it, the independence condition of Chapter 3, since an evaluator that also licenses ends up assessing its own decisions.
Mature designs on the public record allocate by name. Regulation (EU) 2024/1689 assigns the oversight of general-purpose models to an AI Office, the supervision of deployed systems to national market surveillance authorities, and conformity assessment to notified bodies, an explicit owner for each function the regulation creates. The United Kingdom's white paper allocates with the opposite instrument: existing regulators apply five cross-cutting principles within their remits, and a central function inside government monitors coverage and coherence. The two designs differ in almost every choice except the discipline itself: each function has a named carrier, recorded in a public document that can be produced when ownership is contested.
Coherence, the quiet question
Coherence is the quietest of the five because nothing visible depends on it at design time. It is a frame: a shared set of definitions, a shared scale of risk thresholds, shared reporting formats, and a shared list of the standards that sectoral rulebooks cite, the citation of Chapter 5 made uniform across sectors. The content is modest, often a few pages: what counts as an AI system, which uses fall in which risk tier, what an incident report contains, and whether the rulebooks reference ISO/IEC 42001, the NIST AI Risk Management Framework, or national profiles of them.
The frame needs a named holder. The natural candidates are a standing coordination body, where one exists, or the design authority of Chapter 9 for the duration of the design phase, with custody passing to a permanent home when the phase closes. Which body holds the frame matters less than that exactly one does: a frame held by nobody is maintained by nobody, and each rulebook then fixes its own vocabulary at the moment of drafting.
Without the frame, three costs arrive on predictable schedules. Sectoral rulebooks drift, since each supervisor drafts against its own precedents. Firms operating in several sectors face divergent duties for the same system, with an incident reportable in one sector and unrecognised in another. The safety and security regime is reinvented per sector, which duplicates the scarcest capability in the system. The supervisory work that makes drift possible is already under way: the Monetary Authority of Singapore's consultation on AI risk management guidelines shows a sector supervisor writing model-risk expectations in its own terms, which is what supervisors are for; whether such rulebooks converge across a jurisdiction depends on whether a coherence frame exists for them to cite.
How the questions interlock
The five questions are answered together or they are answered badly, because each constrains the others. Capability constrains sequencing: an obligation that binds before the capability able to evidence it exists cannot be supervised, which is why Chapter 7 places the engine early in every workable order. Allocation and sequencing together size the anchoring: the instrument at link three must create exactly the powers the allocation assigns and carry exactly the application dates the sequencing sets, and an instrument drafted before those answers exist either exceeds what the design needs or returns for amendment. Coherence depends on allocation twice, once for the body that holds the frame and once for the supervisors directed to cite it.
A useful early step is to open a one-page decision record for each question at the start of the design phase: the decision, its owner, the instrument that carries it, and the date it takes effect. Blank fields are informative, since a blank names work that remains while a missing record hides it. Five pages then form the auditable core of the design phase, the readiness assessment of Chapter 13 checks little more than their existence and their consistency, and the record protects the design against the alternative described below, in which the questions are answered without anyone deciding anything.
Common failure mode. Five questions, answered by default. A design stage that never asks the questions still answers them. Allocation defaults to whoever moved first, sequencing to everything at once, capability to nobody, anchoring to a future law, and coherence to none. Each default is an answer chosen without a decision record, and an answer without a record is discovered rather than reviewed, typically at implementation, when the overlaps and gaps of Chapter 2 present themselves as institutional facts.
The five questions set the agenda for the rest of Part III. Chapter 7 takes up the one with an order inside it, sequencing the rollout: which parts of the regime arrive first, which wait on capability, and how a timetable is written so that it can be kept.
Three questions for every government
- For each of the five questions, does a dated decision record exist naming the decision, its owner and its instrument?
- Which functions have more than one owner, and which have none?
- Which body holds the coherence frame, and which definitions, thresholds and citations does it actually share?
Selected public sources
- Regulation (EU) 2024/1689 (AI Act), European Union, 2024
- A pro-innovation approach to AI regulation, white paper and government response, United Kingdom, 2023 and 2024
- Consultation Paper on Guidelines on Artificial Intelligence Risk Management, Monetary Authority of Singapore, 2025
- ISO/IEC 42001, AI management systems, ISO/IEC, 2023
- AI Risk Management Framework and Generative AI Profile, NIST, 2023 and 2024