Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
Deadline Compliance Is Not AI Implementation¶
Revised and substantially expanded July 17, 2026. Executive Order 14110, discussed as the historical catalyst for the original post, was revoked in January 2025; the operating lessons remain relevant to later federal AI policy.
In February 2024, federal officials described early progress implementing President Biden's executive order on artificial intelligence. The Office of Science and Technology Policy (OSTP) occupied a central coordinating role, and the public conversation emphasized completed deadlines, forthcoming guidance, and agencies “leading by example.”
The original version of this post reproduced that progress narrative without interrogating its unit of measure. It treated executive action, policy development, and agency implementation as though they were successive names for the same thing.
They are not.
A mandate can be issued but not translated. A deliverable can be published but not adopted. A process can be adopted but fail to change technical behavior. A system can be deployed but fail to improve a public outcome. Each transition requires different evidence and different ownership.
The management problem is therefore one of implementation traceability: can leaders follow a policy intent all the way through accountable deliverables, changed operating practices, fielded capabilities, and measured outcomes—and can they see where that chain breaks?
The Implementation Chain¶
The contemporary Federal News Network interview described the tempo generated by Executive Order 14110. That tempo was real. The order assigned many actions to agencies with deadlines and coordinating responsibilities.
For program management, however, “action completed” should be only the first link in a longer chain:
- Intent: What public or mission outcome is the policy trying to produce?
- Mandate: Which authority directs or permits action, and to whom?
- Deliverable: What guidance, standard, service, contract vehicle, dataset, evaluation method, or organizational mechanism must be produced?
- Adoption: Which programs must change behavior, by when, and what support do they need?
- Capability: What can an agency or operator now do reliably that was not possible before?
- Outcome: What changed for the mission, workforce, public, market, or affected population?
- Learning: What evidence should revise the policy, implementation, or technical system?
Most executive-order tracking concentrates on steps two and three because they are easy to assign and report. The public value lies primarily in steps five through seven.
Why Deliverables Fail to Become Capability¶
Federal organizations are adept at producing policy artifacts. Capability requires those artifacts to cross multiple institutional boundaries.
A new AI risk framework may need to be interpreted by mission programs, integrated into acquisition, converted into engineering requirements, supported by evaluation tools, reviewed by privacy and security teams, and operated by personnel with clear decision authority. If any participant receives only a high-level memorandum, each will create a local interpretation. The organization achieves formal diffusion without coherent execution.
Common breaks in the chain include:
- Authority without resources: an office receives responsibility but no staff, budget, data, or contracting support;
- guidance without executable criteria: teams are told to use AI “responsibly” but not what evidence is required for a decision;
- standards without integration: a technical publication exists but does not appear in development pipelines, solicitations, acceptance tests, or monitoring;
- pilots without operational owners: a prototype demonstrates value but no program owns production funding, authorization, support, or user adoption;
- inventories without discovery: annual reports are built through email rather than from living technical and acquisition records;
- governance without decision rights: boards discuss risk while no identified official can accept, mitigate, or reject it;
- metrics without causality: dashboards count activities that are only weakly connected to outcomes.
Implementation leadership should be organized around finding and repairing these breaks.
Build an AI Policy Control Tower¶
“Control tower” can sound like centralized command. The useful analogy is shared situational awareness across a distributed enterprise. Agencies retain mission authority, while a coordinating function maintains a traceable view of dependencies, decisions, evidence, and obstacles.
An AI policy control tower should combine five capabilities.
A requirement graph¶
Represent mandates as structured requirements linked to legal authority, responsible organizations, deadlines, definitions, dependencies, and expected outcomes. Policy text remains authoritative; the graph makes relationships queryable.
This is particularly important when multiple policies overlap. An AI system may be subject to acquisition guidance, privacy law, records requirements, cybersecurity controls, civil-rights obligations, sector regulation, and agency-specific policy. Programs need a resolved implementation path, not a stack of hyperlinks.
A deliverable and adoption registry¶
Track not only whether guidance was published, but which organizations have incorporated it into their processes. For a new evaluation standard, adoption evidence might include updated acquisition templates, pipeline checks, trained evaluators, completed system assessments, and exception records.
A capability map¶
Identify the services agencies need in order to comply and deliver: secure model access, data environments, red-team support, privacy review, test harnesses, model registries, contract language, monitoring, and incident response. Map where those capabilities exist, their capacity, and which agencies can reuse them.
An evidence layer¶
Attach artifacts and metrics to claims of progress. If an agency reports that it can assess high-impact AI, the evidence should show the method, people, systems assessed, decisions made, time required, unresolved gaps, and observed outcomes.
An impediment and escalation process¶
Implementation problems should be categorized and routed: statutory ambiguity, budget, acquisition lead time, hiring, data access, security authorization, vendor terms, standards maturity, or cross-agency dependency. Leadership can then address systemic barriers instead of asking individual programs to improvise around them.
Use Leading and Lagging Evidence Together¶
Waiting for public outcomes can take years. Counting documents is too shallow. A useful implementation system combines leading indicators with lagging evidence.
For an agency attempting to improve responsible AI acquisition, the evidence might progress as follows:
- Leading: standard contract clauses approved; acquisition workforce trained; evaluation service funded; model and data-rights questions added to market research.
- Behavioral: solicitations use the clauses; evaluators are engaged before award; vendors provide required artifacts; programs document risk decisions.
- Capability: agency can compare systems on mission-relevant evidence, detect unacceptable terms, preserve exit options, and monitor deployed performance.
- Outcome: fewer failed pilots, faster responsible deployment, lower switching cost, improved service or mission performance, fewer unmitigated incidents.
This sequence prevents two errors. Leaders do not dismiss enabling work merely because outcomes are not yet visible, and they do not mistake enabling work for the outcome itself.
Coordinate Definitions Without Centralizing Every Decision¶
Cross-government AI policy requires enough consistency to support accountability and reuse. It also must accommodate enormous variation in mission, authority, data, and consequence.
The coordinating layer should standardize the grammar:
- core system and lifecycle identifiers;
- minimum inventory fields;
- evidence and incident schemas;
- high-level risk and impact dimensions;
- interfaces for reporting and shared services;
- methods for documenting exceptions and residual risk.
Agencies should retain the ability to apply mission-specific thresholds, controls, and decision authorities. A weather-research model, benefits-eligibility support tool, cyber-defense agent, and internal writing assistant should not pass through an identical review. They should be describable within a common evidence system.
This is federated governance: centralized visibility and reusable infrastructure paired with distributed, accountable decisions.
Policy Turnover Makes Traceability More Valuable¶
The 2025 policy change did not make implementation traceability obsolete. It made it more important.
Executive Order 14179 revoked the 2023 order and directed agencies to review actions taken under it. To perform that review intelligently, leaders needed to know which policies, contracts, systems, roles, and controls derived from the prior order; which also rested on statute or independent authority; which produced useful capability; and which conflicted with the new direction.
Without traceability, policy transition becomes indiscriminate deletion or laborious rediscovery. With it, officials can distinguish:
- an executive-order-specific requirement from a generally useful engineering control;
- a rescinded definition from a statutory obligation;
- a paperwork burden from a service that accelerates responsible delivery;
- a governance role that needs a new charter from a function that no longer has authority;
- a technical asset worth preserving from a process that should be retired.
OMB's successor M-25-21 guidance continued to require agency strategies, Chief AI Officers, use-case inventories, governance, high-impact AI risk management, and public accountability while revising priorities and definitions. An agency with structured implementation knowledge could adapt. An agency with scattered compliance artifacts had to begin again.
OSTP's Most Important Role Is Translation¶
OSTP does not operate every federal AI system. Its highest-value function is not issuing technical instructions to every program. It is helping translate among policy, science, standards, agency operations, industry capability, and public consequence.
That role includes identifying where a problem requires:
- a cross-government standard rather than agency-by-agency invention;
- scientific research rather than premature regulation;
- procurement leverage rather than voluntary guidance;
- legislation or appropriations rather than executive direction;
- shared infrastructure rather than another policy memorandum;
- domain-specific agency judgment rather than centralized prescription.
Translation is a knowledge-management function. It depends on preserving implementation evidence and making it usable across organizational boundaries.
The Novel Inference: Policy Is a Product With Users¶
Federal policy is often written as though publication completes delivery. A better model treats policy as a product whose users are agency leaders, program managers, engineers, acquisition professionals, evaluators, oversight bodies, and members of the public.
Those users need different interfaces. An engineer needs testable requirements. An acquisition official needs enforceable terms. An agency head needs risk and portfolio visibility. An inspector general needs reconstructable evidence. A member of the public needs meaningful transparency. The authoritative policy can remain one document while its implementation is expressed through role-specific tools, workflows, and data.
Product thinking also implies feedback. Where do users misunderstand the requirement? Which review steps add delay without better evidence? Which shared services are oversubscribed? What harms or failures were not anticipated? What behavior is being gamed? Policy teams should observe these signals and revise implementation just as a technical product team improves a system.
Deadline compliance creates tempo. Implementation traceability turns tempo into accountable change.
The enduring test for any federal AI initiative is therefore straightforward: can a leader move from the announcement to the evidence of what changed, for whom, under whose authority, at what cost, with what residual risk? If not, the government has progress reporting but not yet an implementation system.
This intersection of policy, knowledge architecture, engineering delivery, and organizational change is central to my work on public-sector AI. If your organization needs to make a complex transformation traceable from intent through mission outcome, connect with me on LinkedIn or send a direct inquiry.