The Policy Operating Model Was Built for a Different Era
A simplified traditional policy process often resembles this:
There is nothing inherently wrong with this sequence. Many important policy questions require deliberate investigation, expert consultation and time.
The problem arises when this becomes the default operating model regardless of the speed or nature of the issue. Modern policy environments increasingly exhibit three characteristics.
1.1 Change arrives before institutional cycles are complete
Emerging technologies can move from experimentation to widespread adoption in months. International regulatory approaches change rapidly. Public behaviours can shift before official statistics reveal them. New business models can cross multiple regulatory domains at once.
In these environments, the policy question itself may change during the research cycle.
1.2 Policy problems are increasingly interconnected
Artificial intelligence is not simply a technology-policy issue. It can simultaneously raise questions about labour, education, cybersecurity, competition, public services, intellectual property, data governance and national competitiveness.
Similarly, climate adaptation is inseparable from infrastructure, finance, housing, public health and urban planning. Linear research processes struggle when the system being studied is not linear.
1.3 Information abundance creates an interpretation problem
AI and digital information systems have made enormous volumes of knowledge accessible. The constraint increasingly shifts from finding information to determining:
- what matters
- what is reliable
- what is changing
- what is missing
- what it means for a specific institution
- what decision it should inform
This is fundamentally an intelligence problem.
OECD research on anticipatory governance similarly identifies the growing need for governments to integrate foresight, experimentation and innovation into ordinary policymaking rather than leaving them as disconnected specialist activities.
The Wrong Response: Automating the Old Model
The simplest application of AI to policy work is to insert AI into existing tasks. A researcher uses AI to summarize documents. A team generates a first draft more quickly. A department introduces a chatbot. A policy unit searches legislation using a language model.
These applications can be valuable. But they do not necessarily constitute institutional transformation.
Consider a policy process that takes twelve months because:
- evidence is fragmented
- responsibilities are sequential
- data arrive periodically
- consultation begins late
- options are not tested
- lessons from implementation do not return to policy design
If AI simply writes the final report faster, the operating model remains unchanged. The deeper opportunity is to ask:
That question leads to a different model.
From Policy Research to Policy Intelligence
The transition can be described through five shifts.
None of these transitions eliminates traditional research. Rigorous research becomes one component of a broader system.
The objective changes from producing knowledge about a problem to building the institutional capability to:
The AI-Native Policy Operating Model
The Institute's model separates three layers of responsibility.
AI provides scale
AI can extend the range of analysis through:
- Search — locating relevant material across large information environments.
- Sensing — continuously monitoring defined signals and indicators.
- Structuring — organizing unstructured evidence into usable forms.
- Synthesis — comparing and summarizing large bodies of material.
- Scenario generation — producing alternative future conditions for examination.
- Simulation — helping explore potential effects and interactions.
- Drafting — rapidly producing alternative formulations, options and communication products.
People provide judgment
Human policy professionals remain responsible for:
- Purpose — determining what problem deserves attention.
- Framing — defining the question correctly.
- Context — understanding institutional, political, legal and cultural conditions.
- Values — determining what outcomes are desirable.
- Trade-offs — balancing competing legitimate objectives.
- Verification — challenging evidence and model outputs.
- Accountability — standing behind consequential recommendations and decisions.
The operating model provides discipline
Technology and expertise still require an institutional process connecting them. That process must provide:
- Workflow — when different types of intelligence enter.
- Controls — what requires review and authorization.
- Traceability — how claims connect to evidence.
- Feedback — how outcomes return to future analysis.
- Learning — how knowledge persists beyond an individual project.
At the Institute, this interface is the Policy Intelligence Cycle.
The Policy Intelligence Cycle
Then sensing begins again. Every decision changes the environment from which the next cycle starts.
What This Looks Like in Practice
Consider a government assessing whether and how autonomous AI agents should enter public services.
Under a conventional model
Government might commission a study of international approaches, receive recommendations and then begin considering implementation.
Under a policy-intelligence model
- Sense tracks agent capabilities, emerging incidents, adoption and regulatory signals.
- Understand maps service opportunities, cybersecurity risks, accountability questions, workforce implications and citizen expectations.
- Design creates alternative governance models — for example, restricted assistance, supervised autonomous action or bounded pilot deployment.
- Simulate examines failure scenarios, model manipulation, erroneous service decisions and escalation mechanisms.
- Test deploys a limited pilot with defined services, users and human override.
- Evaluate measures accuracy, efficiency, user trust, incidents and operational burden.
- Advise provides leadership with evidence-backed options for expansion, restriction or redesign.
- Communicate explains the selected approach and its safeguards.
The value comes not from one AI-generated report. It comes from connecting intelligence to controlled action and action back to learning.
Human-in-the-Loop Governance
The more capable AI becomes, the more important governance becomes. The Institute proposes several baseline safeguards.
Named accountability
Every consequential product should have an identifiable human owner.
Evidence traceability
Material claims should be recoverable to their sources wherever practicable.
Contestability
AI-generated analysis should be challengeable and revisable.
Proportional oversight
Higher-consequence, less reversible and more rights-sensitive uses should receive stronger review.
Uncertainty disclosure
Policy advice should communicate what is known, what is uncertain and which assumptions materially affect the recommendation.
Clear delegation boundaries
Institutions should distinguish between what AI may:
These are not equivalent permissions.
NIST's AI Risk Management Framework emphasizes governance and organizational accountability throughout the AI lifecycle, while UNESCO calls for oversight, auditability, traceability and ultimate human responsibility.
Institutional Requirements
An AI-native policy operating model cannot be created by purchasing a model licence. At least eight institutional capabilities are required.
1. Trusted information access
Teams need structured access to relevant legislation, data, research, decisions and institutional records.
2. Secure AI infrastructure
Sensitive government work requires controlled environments, access management and security appropriate to the information involved.
3. Multidisciplinary teams
Policy intelligence requires combinations of: policy expertise, domain expertise, data and AI capability, foresight, behavioural insight, design, evaluation and legal understanding.
4. Institutional memory
Insights generated in one policy cycle should be recoverable in the next.
5. Governance
Organizations need policies defining acceptable AI use, accountability, validation requirements, data handling and escalation.
6. Experimentation capability
Institutions need mechanisms to test policy ideas before broad implementation where appropriate.
7. Evaluation capacity
A system that experiments without learning is merely producing activity.
8. Leadership sponsorship
Workflow redesign often crosses organizational boundaries. Without senior sponsorship, AI can remain confined to isolated productivity tools.
Recent OECD evidence similarly identifies data governance, infrastructure, skills and organizational capacity as important enabling conditions for scaling government AI.
A Practical Implementation Pathway
Governments do not need to redesign the entire policy system at once. A disciplined transition can begin with one domain.
Stage 1 — Select
Choose a policy area where:
- the environment changes quickly;
- evidence is distributed across many sources;
- decisions recur;
- uncertainty is significant; and
- better intelligence could create visible value.
Stage 2 — Instrument
Establish continuous sensing. Define the sources, signals, datasets and developments the team needs to monitor.
Stage 3 — Integrate
Connect sensing with research, option design and decision support. Do not allow the AI component to become a separate technology exercise.
Stage 4 — Govern
Establish:
- human approval points;
- evidence standards;
- audit trails;
- access controls;
- acceptable-use rules;
- testing requirements; and
- escalation procedures.
Stage 5 — Test
Run one complete Policy Intelligence Cycle on a real government question.
Stage 6 — Measure
Compare the new model with the previous approach.
Stage 7 — Scale
Expand only where evidence demonstrates value.
AI-native government should itself be developed according to the logic it advocates:
How Success Should Be Measured
The model should not be judged by the number of AI tools deployed. A policy intelligence capability can instead track measures such as:
Responsiveness
- time from meaningful signal to analyst awareness;
- time from policy question to decision-ready advice.
Analytical quality
- breadth and quality of evidence reviewed;
- source traceability;
- number of credible alternatives considered;
- explicit treatment of uncertainty.
Anticipation
- emerging risks identified before materialization;
- scenario coverage;
- frequency of policy adjustment in response to new evidence.
Decision usefulness
- decision-maker use of intelligence outputs;
- clarity of trade-offs;
- recommendation adoption or modification.
Experimentation
- proportion of suitable policies tested before scale;
- lessons incorporated after pilots.
Organizational productivity
- analyst time shifted from repetitive information processing toward interpretation and judgment.
Governance
- compliance with human-review requirements;
- auditability;
- incidents, corrections and escalation patterns.
Not every indicator should be optimized. Speed, for example, is valuable only if quality remains sufficient.
The objective is better institutional performance, not maximum automation.
Risks and Limitations
A credible AI-native model must address where it can fail.
Hallucination and factual error
Generative AI can produce confident but incorrect information. Material claims require verification.
Weak evidence at scale
Processing more sources does not guarantee better evidence. AI may scale poor-quality information just as effectively as good information.
Automation bias
People may defer to plausible machine output even when they should challenge it.
Bias and representation
Historical data and source environments may embed distortions or exclude important perspectives.
Security and privacy
Policy intelligence can involve sensitive government information. AI architectures therefore require appropriate technical and institutional controls.
False precision
Simulations and forecasts can create an illusion of certainty. Scenario outputs should be understood as structured explorations, not guaranteed futures.
Deskilling
Excessive delegation can weaken the very analytical capabilities needed to challenge AI.
Institutional dependency
Governments must understand critical systems sufficiently to avoid becoming unable to operate, scrutinize or replace them.
These risks do not argue against AI-native policy work. They argue for designed governance rather than unstructured adoption.
A New Division of Labour
The future of policy work is unlikely to be either fully human or fully automated. It will be built around a new division of labour.
Machines are increasingly good at:
- scale
- speed
- retrieval
- structuring
- comparison
- pattern recognition
- generation
- repeated analytical operations
Humans remain essential for:
- purpose
- judgment
- interpretation
- values
- legitimacy
- negotiation
- empathy
- institutional context
- accountability
The objective is not to decide which side wins. It is to design institutions that combine both deliberately.
From Faster Research to Continuous Intelligence
Artificial intelligence creates an obvious opportunity to make existing government work faster. But speed is the smaller opportunity.
The larger opportunity is to redesign how governments sense, understand, design, test, decide and learn.
Traditional policy research will remain indispensable. Long-form analysis will remain valuable. Human expertise will become more rather than less important. What changes is the architecture surrounding them.
- Research becomes connected to continuous sensing.
- Foresight becomes connected to present decisions.
- Policy design becomes connected to simulation and testing.
- Implementation becomes connected to evaluation.
- Evaluation becomes connected to the next cycle of intelligence.
- And AI becomes an enabling layer across the system rather than an isolated tool.
That is the transition from AI-assisted policy research to AI-native policy intelligence.
The policy institutions that succeed in this transition will not necessarily be those that deploy the most AI. They will be those that establish the strongest relationship between machine capability, institutional design and human judgment.
The next transformation of government policy work will not come from adding AI tools to old institutions. It will come from redesigning how institutions turn change into intelligence and intelligence into action.
Selected References
- OECD, Towards Anticipatory Governance Guidelines for Public Sector Organisations (2025).
- OECD, Building Anticipatory Capacity with Strategic Foresight in Government (2025).
- OECD, Governing with Artificial Intelligence: The State of Play and Way Forward in Core Government Functions (2025).
- OECD, Digital Government Outlook 2026 — Adopting and Governing AI in Government.
- NIST, Artificial Intelligence Risk Management Framework 1.0.
- UNESCO, Recommendation on the Ethics of Artificial Intelligence.