TL;DR: Build a transparent AI governance structure around named human decisions that leave a record:
- Name who decides each type of AI-assisted decision and what authority that person holds.
- Set checkpoint triggers and escalation routes before work begins, and place checkpoints where outcomes carry consequence.
- Record each checkpoint decision (accept, modify, or reject) with its evidence, so authorized reviewers can see what was decided, by whom, and why.
- Check that the people at each checkpoint can actually govern the AI they use.
Who should hold decision authority?
Transparency starts with naming the people who decide. Basil C. Puglisi, MPA, makes the case in What AI Transformation Actually Requires, where he observes that AI blurs decision language: “A system recommends, a platform selects, and an agent executes, while a person reviews and a manager approves.” His answer is to separate the steps: “Organizations should distinguish between generating information, recommending an action, initiating an action, approving it, stopping it, reversing it, and accepting responsibility for the result.” One distinction carries the most weight: “Approving an action lets the process move forward, while accepting an outcome at a checkpoint assigns ownership of it.”
The NIST AI Risk Management Framework points the same way. GOVERN 3.2 calls for “Policies and procedures” that “define and differentiate roles and responsibilities for human-AI configurations and oversight of AI systems,” and GOVERN 2.3 states that “Executive leadership of the organization takes responsibility for decisions about risks associated with AI system development and deployment.” (NIST AI 100-1)
Who qualifies? The Checkpoint-Based Governance page answers: “Qualification is the ability to apply governance, not subject-matter expertise.” It adds that “Placing someone at decisions beyond their experience is a design failure rather than a failure of their authority,” and that “A specialist enters through escalation, which is a form of modify.” The checkpoint itself is “where a named human is required to be present, documented, and accountable.” Basil’s book Governing AI: When Capability Exceeds Control sets out the formal version in Appendix B, “Checkpoint-Based Governance (CBG v5.0),” under its four constitutional properties.
Which decisions need a checkpoint?
Decisions with consequences need a checkpoint, and a team should define the triggers in advance rather than leave them to instinct. The transformation essay lists the kinds of moments that should stop normal work: “A model may contradict a known source, two systems may reach materially different conclusions, or a generated recommendation may exceed the reviewer’s expertise.” Its rule for teams is direct: “The organization should identify the conditions that stop normal work and require another level of review, and it should make clear who receives the escalation and what authority that person or function holds.”
Escalation does not leave a decision hanging. It “routes the work to someone with greater authority, and the receiving decision-maker then closes the checkpoint through a separate decision.” The CBG page closes a common shortcut: “No platform count, confidence score, or level of agreement substitutes for human arbitration.”
Checkpoint density should match consequence. Governing AI reports in Chapter 11, “Governance in Practice,” that checkpoint overhead stayed manageable at the scale of a single manuscript while raising a scalability concern for enterprise deployment, which is why a team should place checkpoints where outcomes matter rather than on every output.
How does the team show its work?
A team shows its work through decision records that authorized reviewers outside the decision can inspect to see what was decided, by whom, and why, with access controls for confidential information and personal data. A documentation protocol the author calls CARCS produces such a record: “a ten-section document that captures what happened, who decided, what the evidence was, and what comes next,” finalized only after “a named human has signed off on it.” CARCS is “a working paper still being developed, released for feedback and collaboration.”
Transparency also needs a rule about facts. The CBG page states it in a heading, “Authority over what happens is not authority over what is true,” and explains: “The arbiter controls acceptance, modification, rejection, and publication.” Yet “A factual assertion made by the arbiter still answers to evidence.” A team lead who can overrule a colleague’s decision still has to show the evidence behind a factual claim.
Instructions need checking too. The CBG page notes that “The confirmation is a comprehension check, not a receipt,” so when a decision enters a shared AI workflow, ask the person receiving it to restate the instruction in one line and resolve any mismatch before work continues.
The Minds That Bend the Machine explains why records matter beyond the team. Chapter 14 reviews Helen Toner and asks whether any oversight body can stay independent when the organization it oversees controls the information it sees, a question that applies to any team whose leaders review AI work through summaries alone.
How do you know people can govern the AI they use?
Assess governance competence through how people actually direct, check, and own AI work. An instrument for that purpose, which the author calls HEQ, is introduced in Measuring the Governance Competence at the Center of AI Literacy as “a behavior-anchored framework that measures the governance competence at the center of AI literacy: the ability of a human to direct, challenge, verify, and own the work done with AI.” It is “a deployed, research-grounded instrument at diagnostic stage, not yet validated on an independent cohort.” Its score “is an indicator inside a human-supervised process and never the decision itself.”
The structure also builds the people inside it. The CBG page describes a developmental mechanism: “The checkpoint develops the person who exercises it, which is what keeps it from decaying into ritual.” In The Minds That Bend the Machine, Chapter 22 reviews Carter Cousineau, whose work embeds controls in the workflows people already use, so governance becomes part of the job rather than a separate task.
For a check by someone who did not build the structure, Puglisi Consulting’s AI Strategy and Readiness Review tests the plan “against the facts it rests on and the outcomes it promises to measure.” Puglisi Consulting states that its advisory reviews are not a legal opinion, a compliance determination, or a substitute for qualified counsel.
Questions Readers Ask
What makes an AI governance structure transparent?
A transparent structure names who holds decision authority, publishes the triggers that require a checkpoint, and keeps a record of each decision with its rationale. Authorized reviewers outside the decision should be able to see what was decided, by whom, and on what evidence, without relying on memory or summaries, subject to normal access restrictions.
Does every AI output need human review?
No. Checkpoints belong where outcomes carry consequence, such as client deliverables, financial figures, or public statements. Routine low-risk work can run with lighter controls, as long as the team defines in advance which conditions stop normal work and send it to a named person for a decision.
Who should be the human decision maker on an AI team?
Choose someone qualified to apply governance to the decision’s scope, not necessarily a specialist. Checkpoint-Based Governance treats placing someone beyond their experience as a design failure. When a decision exceeds the reviewer’s expertise, escalation brings in a specialist, which counts as a form of modify.
How should escalation work in an AI governance structure?
Define the conditions that trigger escalation, such as an AI output contradicting a known source or two systems disagreeing. Name who receives the escalation and what authority they hold. The receiving decision maker then closes the checkpoint by accepting, modifying, or rejecting the work and recording that decision. Track any escalation that stays open past the team’s deadline.
Can an AI system approve another AI system’s output?
Not under Checkpoint-Based Governance. No platform count, confidence score, or level of agreement substitutes for human arbitration, and an output that passes without a completed human decision is not a governed decision. AI can check AI as a control, but a named person still decides.
Related Questions
Sources
National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
Puglisi, B. C. (2025). Governing AI: When capability exceeds control (2nd ed.). https://basilpuglisi.com/governing-ai-when-capability-exceeds-control/
Puglisi, B. C. (2026, April 23). CARCS: Compliance accountability record & case study. basilpuglisi.com. https://basilpuglisi.com/haia-carcs-compliance-accountability-record-case-study/
Puglisi, B. C. (2026, September 13). Measuring the governance competence at the center of AI literacy. basilpuglisi.com. https://basilpuglisi.com/measuring-governance-of-ai-literacy/
Puglisi, B. C. (2026, September 21). What AI transformation actually requires. basilpuglisi.com. https://basilpuglisi.com/ai-transformation-requirements/
Puglisi, B. C. (2026). The minds that bend the machine. https://basilpuglisi.com/minds-that-bend-the-machine/
Puglisi, B. C. (2026, September 11). What is Checkpoint-Based Governance? Human authority. basilpuglisi.com. https://basilpuglisi.com/cbg/
Puglisi Consulting. (n.d.). Advisory reviews. https://puglisiconsulting.com/advisory-reviews/
#AIgenerated using the HAIA Ecosystem | CC BY-NC-SA 4.0
Free for personal, educational, and noncommercial research use with attribution. Commercial exploitation, paid productization, and enterprise commercialization require separate permission and licensing.
Suggested citation: Puglisi, B. C. (2026). Provide advice on how to build a transparent AI governance structure for my team. basilpuglisi.com.
Leave a Reply