Sania Petrov, the newly appointed chief AI officer for a state government, was handed a mandate and a mess. The governor wanted "an AI advisory board" announced within ninety days. By the time Sania inherited it, four agencies had already created their own informal AI working groups, three vendors had offered to "advise pro bono," and a well-meaning legislator had drafted a bill creating a separate commission. Nobody could say which body actually decided anything. At the first meeting, twenty-two people argued for two hours about the agenda and adjourned without a single decision. Sania realized the problem was not a lack of advice. It was that the advisory structure had no design.

Advisory structures are how a government brings outside expertise and public legitimacy into decisions it still owns. Done well, they sharpen judgment, surface risks early, and build trust. Done badly, they become talking shops that slow everything and decide nothing, or, worse, captured channels where vendors shape the rules they will later be paid under. Most AI governance failures in government are not failures of policy text. They are failures of the structures that were supposed to interpret, apply, and enforce the policy.

Structure Is the Decision Engine

The frameworks that govern public-sector AI do not make decisions. The NIST AI Risk Management Framework, released in 2023 as a voluntary standard, provides a functional vocabulary. OMB Memorandum M-24-10, issued in 2024, sets minimum practices and requires each covered agency to designate a Chief AI Officer and stand up an AI Governance Board. Executive Order 14110, issued in 2023, set direction during its period of effect. The National AI Initiative Act of 2020 created a national advisory committee. Every one of these documents defines who must decide and how. None of them decides anything.

The case record shows how decisive that gap is. Michigan's MiDAS unemployment system produced roughly forty thousand false fraud accusations and more than twenty million dollars in settlements, and the primary absence was not a rulebook but a governance board that would have refused to approve minimum-review automated fraud determinations. The Internal Revenue Service paused its facial-verification rollout in February 2022 after public outcry and letters from members of Congress; the root cause was that no advisory structure carrying civil-rights, privacy, accessibility, and tax-administration authority had been engaged before the rollout.

The positive cases have the same shape. When the Department of Veterans Affairs integrated AI into clinical decision support through its national AI institute, the disciplined part was the advisory process: clinicians, ethicists, veterans' service organizations, and academic affiliate medical schools in the room before deployment. When the Centers for Medicare and Medicaid Services considered AI-assisted prior authorization, the review structure included agency leadership, the department's inspector general, patient advocates, and the quality improvement organization network, and the resulting CMS-0057-F final rule issued in 2024 reflects that review.

Look further afield and the pattern holds. The Dutch childcare benefits scandal, the UK Post Office Horizon matter, Houston Federation of Teachers v. Houston ISD in 2017, and the COMPAS litigation reaching State v. Loomis in Wisconsin in 2016 all trace their worst harms to missing or captured advisory structures at least as much as to technical inadequacy. Structure matters because people matter, and structure determines whose expertise, interests, and lived experience are in the room when the decision is made.

Panels, Commissions, and Task Forces Are Not the Same

Leaders use these words loosely, and the confusion is fatal. Each form does a different job. Pick the one that matches your need.

• Advisory panel. A standing group of experts that advises an existing decision-maker. It does not decide; it informs. Best for ongoing technical and ethical input. Sania needs this for continuous guidance on model risk.
• Commission. A formal, often legally chartered body with a defined mandate and a public report. Heavier, slower, more legitimate. Best for one-time, high-stakes questions, such as recommending a statewide policy on facial recognition.
• Task force. A time-boxed team assembled to solve a specific problem, then disbanded. Best for delivery, like standing up an AI procurement playbook in six months.

Sania's original mess mixed all three into one body, so it had a commission's heaviness, a panel's vagueness, and a task force's impatience all at once. Her first move was to split them: a standing advisory panel for ongoing input, plus a six-month task force to deliver the procurement playbook the agencies actually needed now.

The Legal Taxonomy Leaders Must Get Right

Underneath the plain-English labels sits a legal question with real consequences. Federal advisory structures fall into four legally distinct categories, and knowing which box you are in determines what obligations attach.

1. Committees chartered under the Federal Advisory Committee Act. A group established or utilized by the executive branch to obtain advice, and that includes non-federal members, is generally subject to FACA. That means a charter, a Designated Federal Officer, open meetings subject to limited closure exceptions, published minutes, balanced membership, oversight by the General Services Administration's Committee Management Secretariat, and biennial renewal.
2. Intragovernmental bodies. Councils composed entirely of federal employees are generally not FACA-subject, but remain subject to federal records law and internal ethics rules, and can drift into FACA territory if they begin consulting non-federal stakeholders systematically.
3. Agency governance boards. The AI Governance Boards required by M-24-10 are internal decision bodies chartered by agency order, with decision authority rather than advisory-only scope.
4. De facto advisory bodies. Civil-society coalitions, academic panels, and industry groups whose recommendations the government considers but which are not utilized in the FACA sense. Not formally regulated, but they still create accountability expectations.

Getting the box wrong has been litigated. Running what is in substance a FACA committee without chartering it has produced federal litigation, the executive-branch task force dispute in Cheney v. United States District Court in 2004 being the best-known example. The practical guidance is unglamorous: before you convene anyone who does not work for you, ask counsel which category you are creating. That conversation is cheap. Getting it wrong is not.

The Four Questions That Define Any Advisory Body

Every advisory structure, whatever its legal form, is defined by four answers. Skip any of them and the body drifts.

1. Authority. What can this body actually do? Advise, review, approve, or only comment? Be brutally explicit. The most common failure is a body that believes it decides while leadership believes it only advises.
2. Membership. Who sits on it, and how are they chosen and balanced? Who appoints, who can remove, how long do they serve?
3. Reporting. Who does it report to, in what form, on what schedule, and is the report public?
4. Resources and cadence. Who staffs it, who prepares materials, how often does it meet, and how long does it exist?

An advisory body without explicit authority will either seize power it was never given or quietly become irrelevant. There is no stable middle. The same is true of the other three: an unstaffed body cannot prepare, an unreported body cannot be held to anything, and a body with no end date outlives its question.

Getting Authority Right

Authority is the question leaders most want to dodge, because being clear feels like either over-promising or insulting the members. Dodge it anyway and you guarantee conflict. Sania wrote a single sentence into the charter: the Panel reviews and advises, final decisions rest with the Chief AI Officer, and the Chief AI Officer must respond in writing to any formal Panel recommendation within thirty days. That one sentence did two things. It told members honestly that they advise, not decide, so no one felt ambushed later. And the written-response requirement gave their advice real teeth, because leadership could no longer ignore a recommendation silently.

The thirty-day figure is Sania's charter choice, not a legal requirement; federal practice on documented response windows commonly lands at thirty or sixty days, and either works provided it is written down and honored. Advisory power without a response obligation is theater; advisory power with one is influence. Publish both the recommendation and the response, and the obligation becomes self-enforcing, because a missing response is visible to everyone who cared enough to read the recommendation.

Chartering a Governance Board That Actually Works

The internal decision body is a different animal from the external advisory panel, and agencies routinely conflate them. Under M-24-10, covered agencies stand up an AI Governance Board chaired by the Chief AI Officer. Composition determines capacity. A working board seats the Chief AI Officer as chair, the General Counsel or a senior delegate, the Chief Information Officer, the Chief Information Security Officer, the Senior Agency Official for Privacy, the civil-rights or equity lead, the mission owner relevant to each agenda item on a rotating basis, the senior procurement executive for contract items, and the Inspector General liaison as an observer.

Less functional boards fail in recognizable ways. An engineering-only composition treats AI as a technical question and never sees the rights implications. A policy-only composition never sees operational reality and issues guidance nobody can implement. A rotating cast with no continuity cannot build institutional memory, so every decision is relitigated. Sania used exactly this list as a diagnostic against her own body and discovered she had no privacy officer and no procurement executive, which explained why every vendor question had stalled.

Decision rights should be written, not assumed. A functioning board approves new AI use cases classified as safety-impacting or rights-impacting, approves waivers with the required notice, approves impact assessments and monitoring plans, reviews incidents and corrective actions, oversees the annual use case inventory submission, and reviews procurement items above defined thresholds. Cadence should balance urgency against deliberation: emergency items reviewable within business days, standing meetings monthly, a strategic review quarterly. Minutes kept, decisions memorialized, and an annual public summary published alongside the inventory. A board that meets rarely, decides little, and produces no public record is a liability in an audit.

Chartering a Committee for External Input

When an agency needs structured input from non-federal experts, from industry, academia, civil society, and subject-matter specialists, a chartered advisory committee is typically the right vehicle. The national AI advisory committee created under the National AI Initiative Act of 2020 is the prototype at the top of the federal government, operating with subcommittees including one focused on law enforcement. Within agencies, the pattern appears in the Commerce Department's telecommunications and standards advisory structures, the human research protections advisory committee at the health department, the energy department committees that review AI work in the national laboratories, and the homeland security advisory council, which has run AI-related subcommittees.

Chartering is a discipline with fixed parts. The charter states purpose, scope, membership, reporting, and sunset. The General Services Administration reviews it. The Designated Federal Officer manages meetings, notices, minutes, and member ethics briefings. Meetings must be announced in the Federal Register at least fifteen calendar days in advance, with narrow exceptions, and requests to close a meeting must be supported under the Government in the Sunshine Act exemptions. Balanced membership is not optional; it is statutory. Products must be published, because a committee that does not publish never produces durable value.

At the state level, equivalent open-meetings statutes impose similar obligations, including California's Bagley-Keene Act, the Texas Open Meetings Act, the Illinois Open Meetings Act, and the New York Open Meetings Law. The safest working assumption for a leader in any jurisdiction is that advisory work is public work and should be designed as though every draft will eventually be read. That assumption is also the cheapest one, because it costs nothing at the start and everything to retrofit.

Designing Membership Against Capture

Membership is where advisory bodies are won or lost on legitimacy. Two failure modes dominate. The first is homogeneity: a board of technologists that never includes the people the systems affect. The second is capture: vendors and their allies populating the body that will shape the rules governing their own products. A committee of only industry voices will be seen, correctly, as captured. A committee of only critics will be seen as symbolic. Credible bodies mix industry practitioners, academic researchers, civil-society advocates, mission operators, and representatives of affected communities.

Sania designed her panel of eleven seats with explicit balance: three independent technical experts drawn from universities rather than bidders, two ethics and civil-rights voices, two frontline practitioners who actually run the affected programs, two community or affected-population representatives, and two from oversight or legal. Those five groups sum to the eleven seats her charter authorizes. She wrote a conflict-of-interest rule borrowed from procurement integrity practice: no member may serve while their organization holds, or is bidding on, an AI contract with the state, and all members file annual disclosures.

Federal conflicts practice is more formal and worth borrowing from. Special Government Employees file confidential financial disclosures on OGE Form 450, recuse themselves from matters affecting their personal financial interests, and operate under the federal criminal conflict-of-interest statute at 18 U.S.C. 208 and its implementing regulations. Agencies should also manage institutional conflicts, where an employer relationship creates bias even without a direct financial interest in the specific matter. The NIST framework's Govern function calls for diverse perspectives and stakeholder input; a balanced, conflict-screened membership is how you operationalize that rather than merely cite it.

The third pillar is durability, and it is the one nobody plans for. Appointing members to fixed terms without a renewal or replacement pipeline means the body loses its institutional memory at exactly the moment it matures, because the people who finally understand the work rotate off together. Build the succession plan into the charter alongside the appointment process, and stagger terms so that no single cycle empties the room.

The Advisory Network Above Your Agency

No agency advisory structure operates alone, and a leader who does not map the surrounding network routes every question to the wrong place. At the federal level five bodies do distinct work. The Chief AI Officer Council, chartered under OMB guidance issued in 2024, is the interagency body that produces shared templates, harmonizes impact-assessment practice, and coordinates implementation across agencies. The CIO Council's AI Community of Practice is a practitioner network where the people doing day-to-day implementation trade patterns, failures, and tooling.

Alongside them, the national AI advisory committee supplies external advice and public legitimacy, the national AI initiative office coordinates policy at the executive level, and the General Services Administration's AI Center of Excellence offers hands-on implementation help. Route accordingly: policy interpretation to the Chief AI Officer Council, shared tooling to the community of practice, implementation help to the center of excellence, external validation to the advisory committee. Standing bodies change names and charters between administrations, so confirm the current status and remit of any of them before you cite it in a document.

Internationally, comparative structures are worth studying without importing wholesale. The European Union AI Act establishes a board-level coordination structure for member states. Singapore has developed advisory work around its model AI governance framework. The OECD convenes AI committees whose outputs are influential but non-binding. Comparators are useful for design ideas and useless as authority; nothing an international body recommends binds your agency unless your own law makes it so.

State, Local, and Tribal Structures

Advisory innovation is not a federal monopoly, and state and municipal models are often the more practical template for an agency of moderate size. Texas established an AI advisory council through HB 2060 in 2023, with defined membership and reporting requirements producing recommendations to the legislature. New York City runs an algorithmic management policy established by mayoral executive order, which grew out of an earlier automated decision systems task force and has evolved into a permanent review structure. California has run generative AI advisory work under executive order N-12-23, issued in 2023. Washington State's AI work sits under RCW 43.386.

City-level structures tend to be the sharpest, because the decisions are concrete. Seattle's Surveillance Advisory Working Group operates under the city's surveillance ordinance at Seattle Municipal Code 14.18 and reviews specific acquisitions before council approval, which is a materially stronger position than reviewing policy in the abstract. Boston, Oakland, Santa Clara County, and others run comparable structures. Tribal governments, including the Navajo Nation and confederated tribes bodies, have begun establishing AI advisory work aligned to tribal sovereignty principles, which is a distinct posture rather than a smaller version of state practice.

Cross-state associations of governors and state chief information officers, and the state-government working groups that sit alongside them, provide peer learning and model policy that will save you months. Sania's most useful early call was not to a consultant. It was to the equivalent officer in a neighboring state who had already chartered a panel and could tell her which clauses in her draft charter would cause trouble.

A One-Page Charter Template

Sania reduced everything above to a charter that fits on one page. Any agency can adapt the fields below. If you cannot fill in every line, the body is not ready to launch.

• Name and type. Panel, commission, or task force, chosen deliberately, with the legal category confirmed by counsel.
• Purpose. One sentence on why it exists and what decision it serves.
• Authority. Exactly what it can do, and the leadership response obligation with its window.
• Membership. Number of seats, the balance of perspectives, appointment and removal process, term length, and the staggering and succession plan.
• Conflict-of-interest rule. Disclosure requirement, bar on members tied to active or bidding vendors, and treatment of institutional as well as personal conflicts.
• Reporting. To whom, how often, in what form, and whether public.
• Cadence and lifespan. Meeting frequency, notice requirements, and for task forces, a hard end date.
• Staffing and budget. Who prepares materials and runs the body day to day, and the protected time and budget line that make it possible.
• Transparency. Whether agendas, minutes, and recommendations are published, meeting Section 508 accessibility standards so the public can actually read them.
• Effectiveness measures. The outputs you will judge it by, such as decisions supported, policies promulgated, audits completed, and harms avoided, rather than meetings held and members appointed.
• Sunset and review. When the charter itself is reviewed or the body wound down.

Anti-Patterns to Refuse

• Advisory purgatory. A committee that meets, produces reports, and receives no action. Require a documented agency response to each recommendation within a stated window, and publish both the recommendation and the response.
• The rubber stamp. A governance board that approves everything placed in front of it. Require a written staff analysis for each decision, a vote record in the minutes, and a routine dissent pathway. If the minutes never record a disagreement, the board is not reviewing anything.
• The captured committee. Membership skewed toward regulated or vended interests. Statutory balance requirements and disclosure forms constrain this; they do not cure it. A filed conflict disclosure records an interest, it does not neutralize one, so pair disclosure with recusal that is actually enforced and with appointment processes that are open to people outside the industry.
• The invisible structure. Work done in silent internal groups with no public record. Publish charters, meeting schedules, and annual reports, and remember that a chartering document creates obligations rather than legitimacy. Legitimacy comes from who is in the room and what happens to what they say.
• The performative committee. A body designed to absorb criticism rather than generate decisions. Tie it to real authority, whether approval, veto, or at minimum a mandatory response.
• The unfunded mandate. An advisory body with no staff and no budget. Protect the staff officer's time and give the body a budget line, or accept that it will produce nothing.
• The immortal committee. A body that outlives its mission. Set sunset dates and make renewal contingent on demonstrated value.
• Everyone at the table. The twenty-two-person body that cannot decide. Breadth belongs in consultation; a decision body needs a size at which a real argument can happen.
• Free vendor advice. Unscreened pro bono help from firms that will bid on the resulting work. It is lobbying with a better name, and accepting it is how the capture anti-pattern begins.

Practice Prompts

1. Draw the current AI advisory map for your organization: every formal and informal body that provides AI-related input or makes an AI-related decision. Mark each as panel, commission, task force, or internal governance board, and circle the ones whose category nobody has ever confirmed.
2. Audit your governance board against the membership list in this lesson. Name every seat that is empty, then trace one stalled decision back to the empty seat that caused it.
3. Take one advisory body in your orbit and read its charter, its most recent renewal, and its last report. Draft the single amendment that would most improve it, and identify who has authority to approve that amendment.
4. Write the authority sentence for a body you run: what it can do, who decides, and what leadership owes it in response. Show it to seated members and ask whether it matches what they believed they were doing.
5. Map the interface between your governance board and the external bodies above you. Identify the single communication channel that is most broken and fix that one.
6. Design the full advisory structure for a new high-stakes program, such as AI in unemployment insurance fraud detection: the internal decision body, the external advisory committee, the community consultation mechanism, and the oversight interface. Write the charter for each, and check that you can staff all of them.

Reflection

Think about the last recommendation an advisory body gave your organization. What happened to it? If you cannot trace it to a decision, a written response, or a documented refusal, then the body is producing advice that evaporates, and the members know it even if nobody has said so. Ask yourself what they would tell a reporter about how seriously their work is taken.

Then consider the composition question honestly. If you listed everyone who advises you on AI, how many of them sell AI, and how many have ever been on the receiving end of an automated government decision? A structure that consults only the people who build systems will produce advice that is technically excellent and civically blind, and you will not notice, because everyone in the room will agree.

Glossary

• Advisory panel. A standing body of experts that informs an existing decision-maker without holding decision authority of its own.
• Commission. A formal, often legally chartered body with a defined mandate and a public report, suited to one-time high-stakes questions.
• Task force. A time-boxed team assembled to deliver a specific product and then disband.
• Federal Advisory Committee Act. The federal statute governing advisory committees that include non-federal members, imposing chartering, a Designated Federal Officer, open meetings, published minutes, balanced membership, and periodic renewal.
• Designated Federal Officer. The federal employee who manages a chartered committee's operations, including meetings, notices, minutes, and member ethics briefings.
• AI Governance Board. The internal agency decision body required under OMB Memorandum M-24-10 and chaired by the Chief AI Officer.
• Chief AI Officer Council. The interagency body composed of agency Chief AI Officers, chartered under OMB guidance issued in 2024.
• Special Government Employee. A committee member or short-term appointee subject to federal ethics rules, including confidential financial disclosure and recusal from matters affecting personal financial interests.
• Institutional conflict. A bias arising from a member's employer relationships rather than from their own financial interest, which disclosure of personal finances will not surface.
• Capture. The condition in which a body meant to constrain an interest is populated and effectively steered by that interest.

Related Lessons

• Ethics Boards and Advisory Structures for the ethics-specific body, its distinct remit, and how it relates to the general advisory panel designed here.
• Establishing an AI Governance Board for the full chartering process of the internal decision body.
• Your Agency's AI Governance Structure for how these bodies sit within the wider agency governance picture.
• Community Engagement in Government AI for reaching affected populations beyond seated representatives.
• Public Consultation on AI Policy for the formal consultation mechanisms that sit alongside standing bodies.
• Cross-Agency AI Coordination for working the interagency network described here.
• Federal/State/Local AI Alignment for reconciling structures across levels of government.
• State and Local Government AI Policy for the jurisdiction-specific models this lesson surveys.
• AI Use Case Inventory and Documentation (OMB M-24-10) for the inventory your board is expected to oversee.

Closing

Sania's rebuilt structure was less impressive than the one she inherited. It had fewer people, a narrower mandate, and a charter that fit on a page. It also produced a procurement playbook in six months, a written leadership response to every recommendation, and a public record that a legislator could read without asking for a briefing. The twenty-two-person body had produced two hours of argument about an agenda.

The discipline is not complicated, which is why the failures are so consistent. Choose the form deliberately. Confirm the legal category before you convene anyone. Write down what the body can do and what leadership owes it. Balance the membership and screen the conflicts. Fund the staff work. Publish the output. Set the end date. Advisory structures built this way govern; the ones built without these decisions merely meet, and the difference shows up in public the first time something goes wrong.

Key Takeaways

• Structure decides, documents do not. Frameworks and memoranda define who must decide and how; the advisory structure is the engine that actually makes the call.
• Match the form to the job. Panels inform, commissions legitimize high-stakes one-time questions, task forces deliver and disband; mixing them produces a body that does none well.
• Confirm the legal category first. Chartered advisory committee, intragovernmental council, internal governance board, and de facto body carry different obligations, and misclassifying has produced litigation.
• Define authority explicitly. State whether the body advises, reviews, or approves, and never leave members and leadership holding different beliefs about who decides.
• Give advice teeth with a response obligation. Requiring leadership to answer recommendations in writing, then publishing both, turns advisory power from theater into real influence.
• Composition is capacity. Governance boards fail predictably when privacy, civil rights, procurement, or mission owners are missing from the room.
• Engineer membership against capture. Balance requirements and disclosure forms constrain capture but never cure it; pair them with enforced recusal and open appointment.
• Plan for durability. Stagger terms and build a succession pipeline, or the body loses its memory exactly when it matures.
• Assume advisory work is public work. Notice, minutes, and published products are cheap at the start and impossible to retrofit.
• Write a one-page charter before launch. If you cannot fill every field, including staffing, budget, and sunset, the body is not ready.

Frequently Asked Questions

Do we need a chartered external committee, or is an internal governance board enough? They do different jobs and most agencies need both. The internal board decides: it approves use cases, waivers, impact assessments, and incident responses, and it can be composed entirely of your own officials. The external committee supplies expertise and legitimacy you cannot generate internally, which is precisely why it carries chartering, notice, and balance obligations. Using one where you need the other is the most common structural error in this space.

How large should an advisory body be? Small enough that a real argument can happen and everyone speaks. Sania's eleven-seat panel is one workable answer, and the twenty-two-person body she inherited is a clear example of the failure mode. Breadth of perspective belongs in the consultation and outreach layer, where you can hear from hundreds of people, not in the body that has to reach a position.

Can we accept free help from a vendor who knows this technology better than we do? Take the technical education, in writing, on the record, and available to anyone. Do not seat them on the body that will shape rules they later bid under. If the expertise genuinely exists nowhere else, buy it under a contract with conflict terms rather than accepting it as a favor, and disclose the arrangement. Free advice from an interested party is the cheapest thing in your budget and the most expensive thing in your record.

What if leadership simply ignores the body's recommendations? That is a design failure, not a personality problem, and it is fixed in the charter. Require a written response within a stated window and publish the recommendation and the response together. Leadership retains the right to disagree, which it should, but silent non-response stops being available. Bodies whose advice disappears lose their best members first, and those members are the hardest ones to replace.

Should minutes and recommendations really be public? For chartered committees, publication obligations generally attach as a matter of law, and state open-meetings statutes impose comparable duties. Even where you have discretion, publish. The body exists partly to lend public legitimacy to decisions your agency owns, and legitimacy that is never visible does no work. Publish accessibly as well, so that the disclosure reaches people who use assistive technology.

How do we know the structure is working? Measure outputs, not inputs. Decisions supported, policies promulgated, audits completed, and harms avoided are meaningful; meetings held and members appointed are not, and a body that reports only the second set is telling you it has nothing in the first. Add one qualitative check: ask whether the minutes have ever recorded a disagreement, because a body that never disagrees is not reviewing anything.