DECISION ARCHITECTURE | TRUSTED ARCHITECTURE | OPEN CAPABILITY
Government should use AI to expand options and accelerate understanding while deliberately protecting the moment of irreversible choice.
By Vincent Sullivan, Founder, Cyber Strategy Institute
Federal AI policy often treats speed and safety as competing goals. One side argues that government must move at commercial tempo. The other responds with approval layers, testing gates, and restrictions. Both approaches can miss the key distinction between intelligence and commitment.
The wrong debate
AI should make sensing, search, translation, summarization, option generation, simulation, and drafting faster. Those activities expand the decision space. The risk begins when the same system narrows the options, recommends one path, approves it, and triggers execution before accountable officials have evaluated consequence and reversibility.
The federal doctrine should therefore be simple: fast intelligence, slow commitment.
Tempo should follow consequence
A low-stakes drafting task should not require a board. A reversible software configuration may need automated rollback and monitoring. A procurement exclusion, benefit denial, cyber response action, personnel decision, or physical-world operation deserves stronger review because error is costly, rights may be affected, or reversal may be difficult.
The correct control variable is not whether AI was used. It is the authority and consequence attached to the decision. As consequence rises, government should add source verification, alternative analysis, independent challenge, separation of duties, documented dissent, and recovery planning.
What agencies should require
Every consequential AI-assisted process should identify the decision owner, authorized roles, evidence threshold, approved sources, required alternatives, escalation criteria, and rollback path. The system should preserve a decision receipt containing the model and version, material inputs, sources, assumptions, confidence, dissent, approval, and execution status.
No single workflow should generate, recommend, approve, and execute a high-impact action without explicit legal authority and controls proportional to the consequence. Continuous behavioral assurance should monitor deception, manipulation, sycophancy, verification, and control bypass. Yellow conditions should trigger added review. Red conditions should suspend commitment or execution and enable substitution.
Procurement can make the doctrine real
Agencies should buy decision quality rather than document volume. Solicitations should require a Decision Architecture Plan and a Decision Evidence Package. Model-neutral interfaces, portable policy, standardized logs, and substitution rights should prevent one provider from owning the mission and the frame.
This approach supports both innovation and accountability. It allows broad experimentation, local models, open-weight systems, and rapid generation while concentrating friction where the exercise of authority can cause serious harm.
Government does not need to choose between moving fast and governing well. It needs an architecture that knows which part of the process should be fast.
Sources and further reading
[1] U.S. Department of Defense, Data, Analytics, and AI Adoption Strategy (2023).
[2] U.S. GAO, AI Accountability Framework, GAO-21-519SP.
[3] U.S. GAO, GAO-25-107653.
[4] NIST AI RMF and AI RMF Playbook.
Author note: Vincent Sullivan is the founder of Cyber Strategy Institute and a retired U.S. Air Force lieutenant colonel. His work focuses on engineered security, runtime governance, federal acquisition, and AI decision sovereignty.
Accelerate sensing and analysis. Slow only the commitment that cannot be easily reversed.
