Skip to main content

The Working Group That Outlives Its Founder

The short answer

A working group is four written design decisions, not a calendar invite with a name

Decision scope and escalation, role-based membership, release-matched cadence, and a readable record. A group missing any one part fails in a specific, predictable way; Axon's AI Ethics Board had three of the four for years and still collapsed the moment the fourth, escalation, was tested.

What you will be able to do

  • Design a working group charter with all four load-bearing parts: decision scope and escalation, role-based membership, release-matched cadence, and a readable record, tailored to your own organization's AI decisions.
  • Distinguish decisions a working group can make on its own authority from decisions it must escalate, and name the specific committee or role the escalation goes to, so "we recommend" never quietly becomes "we decide" or vanishes into a founder's private judgment.
  • Write a membership rule that names a role (Head of Data Science, Legal counsel for AI, a rotating product-team representative) rather than a person, so a departure changes who sits in the chair without changing whether the chair exists.
  • Set a meeting and decision cadence tied to your organization's actual release or deployment cycle, not to a fixed calendar interval that drifts out of sync with the decisions it exists to catch.
  • Produce a record format that every member, and every person the group's decisions affect, can read without requesting access, and that states not just what was decided but why, so a decision can be reconstructed months later.
  • Diagnose a working group that has stopped producing decisions the organization can point to, using the four-part charter as a checklist, and identify which specific part is missing rather than concluding vaguely that "the group is not working."
  • Defend the charter you produce against the hardest challenge a working group ever faces: the moment its own founder or sponsor wants to go around it, exactly as Axon's board discovered in June 2022.
  • Write the charter in plain, checkable sentences a member who was not in the room can repeat back accurately, applying the same glance-test discipline this program applies to a change narrative.

The ideas, one by one

The escalation path is the leg that only gets tested once, and it is the one that matters most

A working group can look completely healthy for years on the strength of a founder's or sponsor's personal commitment to consult it, and that commitment tells you nothing about whether the group survives the one decision where the founder's own judgment disagrees, which is exactly what happened at Axon in June 2022.

Name a specific escalation body, not a specific person's discretion

"We will consult the group" is not an escalation path if the only check on ignoring that promise is the same person's own willingness to keep it. A real escalation path names a body distinct from the decision-maker, with a stated turnaround time.

Membership defined by role, with a deputy rule, survives departures; membership defined by name does not

A charter that names "the VP of Engineering" inherits its chair automatically when someone new takes that role. A charter that names a person by name has to be rewritten, or simply stops functioning, the day that person leaves.

Cadence has to be derived from the real release rhythm, not copied from a generic governance template

A quarterly cadence against a weekly release cycle does not produce careful oversight; it produces a rubber stamp, because by the time the group meets, reversing a shipped decision has become expensive.

A record with no reasoning preserves the vote and destroys the point of keeping records at all

"Approved, 6 to 2" tells a future reader nothing about whether the decision still holds given new facts. Capture the evidence and the reasoning, not only the outcome.

Diagnose the specific missing leg; do not conclude vaguely that "the group is not working."

A struggling working group almost always has one identifiable broken part among the four, and the fix is different for each one. Naming the leg precisely, using the four-part checklist, is the transfer skill this topic builds.

The charter is the structural counterpart to the change narrative

The narrative is what you told the people affected by an AI decision (see Topic 9.6); the charter is who keeps deciding, and how, after you have told them. Both belong in the same dossier, because both are tested the same way: by whether they hold up under a hostile, honest read months later (see Topic 10.1).

A working group's authority should not depend on any one person's continued presence, however good that person is

Herbert's rewrite of Halyard's charter did not distrust Sonia; it acknowledged that a structure built to run on one person's instincts is not a structure at all, it is that person, temporarily wearing a committee's name.

A charter you have never checked against real behavior is an untested hypothesis, not a proven success

Time to decision, override rate, seat continuity across a departure, and whether an outsider can reconstruct a record are the observable signals that tell you whether the four parts function in practice, not merely on paper.

Listen to it

The same lesson, talked through by two hosts as episode 74 of the podcast.

Real cases

These examples show working groups whose charters held under pressure and working groups whose charters failed, with the specific missing part named. The Axon AI Ethics Board is the anchor case for this topic; the others are referenced for contrast and are owned in full by other topics where noted.

Example 1: Axon's AI Ethics Board, the escalation leg that was never there (United States, 2018 to 2022). A thirteen-member outside board, formed in 2018, exercised real influence for years, steering the company away from facial recognition and away from a social-media scraping tool. First raised to the board around a year earlier, the drone concept was rejected by the board 8 to 4 only weeks before the announcement below. In June 2022, following the Uvalde shooting, the founder and chief executive announced the company would build the drone, expanded to include persistent surveillance, without the board's approval. Nine of thirteen members resigned within days, stating that the company had "bypassed" its own commitment to consult the board. The company then halted the project. The board's judgment was vindicated; its charter was not built to make that judgment binding, because there was no named body the disagreement escalated to before the founder's own choice became final. (Policing Project, 2022; The Register, 7 June 2022; NBC News, 2022; TechCrunch, 6 June 2022.)

Example 2: A works council with statutory teeth, membership and escalation built into law (European Union, ongoing). Where a works council has a legally defined consultation or co-determination right over a workplace AI decision, the escalation leg of the charter is effectively written by statute: management cannot simply proceed over a properly exercised objection without following a defined legal process, and the body the objection escalates to (a labor court or an equivalent tribunal, depending on jurisdiction) is external to the company entirely. This is the strongest possible version of "a named body above the founder's own discretion," and it is one reason consultation done right survives leadership turnover in a way an informal advisory promise never can (see Topic 9.5).

Example 3: NIST's Govern function, the scope question asked at the level of the whole organization (United States, ongoing). The NIST AI Risk Management Framework's Govern function asks an organization to define, among other things, who has authority over which AI risk decisions and how accountability is structured across roles, which is the organization-wide version of the same scope question a single working group's charter has to answer at a smaller scale (see Topic 6.2). A working group charter without a clear decision scope is, in miniature, an organization that has not completed its own Govern function.

Example 4: A financial-services model risk committee, cadence built around a regulatory calendar (financial sector, general pattern). Regulated financial institutions commonly run model risk governance committees whose cadence is set by validation and reporting cycles required by their supervisors, not by an arbitrary quarterly habit. Because the cadence is derived from an external, non-negotiable clock (an examination cycle, a validation deadline), the committee cannot quietly drift out of sync with the decisions it exists to catch the way an internally chosen "quarterly" cadence can. The general lesson transfers even where no regulator sets the clock: derive the cadence from a real external rhythm, not from what looks appropriately serious on a calendar.

Example 5: A departed-founder handoff done right, the membership leg tested directly (general pattern, illustrative of the mechanism, not a single cited case). When a working group's charter names a role (the Head of Platform Engineering, say) rather than a name, a founder's departure changes who currently holds the chair without requiring anyone to notice, debate, or re-found the group; the successor simply inherits the seat as part of inheriting the role. Contrast this with a group whose charter, written or unwritten, effectively reads "chaired by whoever started this," where a departure leaves the seat empty until someone with enough standing decides to claim it, and the group's authority is genuinely uncertain in the interval.

Example 6: A rubber-stamp committee, the cadence and scope legs both broken at once (general pattern, common in practice). A working group that meets after the product has already shipped, and whose only real power is to recommend rather than decide, tends to converge on approving everything it sees, not because its members lack judgment, but because reversing a shipped decision is expensive and the group was never given the authority or the timing to catch the decision before that cost existed. The fix is not better members; it is moving the cadence earlier in the release cycle and, where the decision category warrants it, moving the scope from recommend to decide.

Example 7: The EU AI Act's governance chapter, scope written into law rather than left to a company's discretion (European Union, ongoing). Article 53 of the EU AI Act, in force for general-purpose AI model providers since 2 August 2025, requires quality-controlled technical documentation of a model's training and testing, kept current and retained as evidence of compliance rather than reconstructed after the fact. The article does not itself dictate a company's internal committee structure, but it does something structurally identical to what this topic asks of a working group charter: it forces the question of who can produce the evidence behind a claim, in writing, in advance of anyone asking. Where a legal instrument requires that evidence to exist on demand, an organization's own working group charter has less room to leave decision scope vague, because a regulator, not only an internal auditor, will eventually ask the same question this topic asks: who had the authority to make this call, and can you show it. (see Topic 6.2)

Example 8: A charter that survived its own founder's departure by design (general pattern, illustrative of the mechanism). Where a working group's chair is defined by role and a handover between a departing chair and their successor includes a walkthrough of open escalations, the group's authority transfers with no gap, because nothing about its standing depended on the outgoing person's continued presence to begin with. The observable difference from a founder-dependent group is not dramatic on the day of the departure; it is the complete absence of a crisis, an empty seat, or a pause in decisions, which is precisely the boring, unremarkable outcome a well-designed charter is built to produce.

Example 9: A model risk committee's decision record produced under regulatory examination (financial sector, general pattern). When a banking supervisor examines a regulated institution's model governance, the examiner routinely asks not only what was decided but why, requesting the documented rationale behind a model's approval or a risk exception. An institution whose committee kept only approval dates and vote counts finds itself reconstructing reasoning under examination pressure, months or years after the fact, which is a materially worse position than producing an existing record; this is the readable-record principle tested by an outside party rather than an internal audit.

Example 10: A trade association's technical standards committee, cadence tied to a publication cycle rather than a calendar (general pattern, common in standards bodies). Voluntary technical standards bodies commonly set working group meeting schedules around draft-publication milestones (a public comment period, a ballot deadline) rather than a fixed monthly or quarterly habit, because a vote that misses a publication window has real, costly consequences for the standard's release schedule. This illustrates the release-matched-cadence principle in a domain with no software release cycle at all: the cadence is still derived from the real external rhythm the group's decisions must intersect, whatever that rhythm happens to be.

Where people go wrong

  • "We have an AI committee, so we have AI governance." A recurring meeting with a name is not a governance structure until decision scope, role-based membership, release-matched cadence, and a readable record are all written down. Axon had thirteen genuinely qualified people meeting regularly for four years and still lost the one decision that mattered most, because the charter under the meeting had a missing leg.
  • "The founder's personal commitment to consult us is enough." A personal commitment is real and valuable right up until the moment the founder's own judgment wants something different, and on that day a personal commitment offers nothing a written, named escalation path would have offered instead. Write the commitment into a structure that does not depend on any one person continuing to keep it.
  • "Naming people instead of roles is fine, we trust the people we chose." Trust in the current people is not the design question. The design question is what happens to the group's authority the week those specific people are gone, promoted, or on leave. A charter that answers "nothing changes" without anyone having to decide that is a role-based charter; a charter that answers "we would have to figure that out" is a name-based charter waiting to fail quietly.
  • "Quarterly is the standard cadence for a governance committee." There is no standard cadence; there is only a cadence matched, or not matched, to how fast decisions actually need to be made in your organization. A quarterly cadence against a weekly release cycle does not produce careful governance, it produces a group that rubber-stamps things that already shipped, because reversing them has become expensive by the time the group meets.
  • "Minutes are a formality; what matters is that we decided." A decision with no recorded reasoning is not reconstructable. Six months later, no one, including the people who made the call, can reliably explain why a threshold was set where it was, which means the organization cannot defend the decision to an auditor, learn from it if it turns out wrong, or even confirm it still makes sense given new facts.
  • "An advisory board that can only recommend has no real power, so the escalation path does not matter much." Axon's board could only recommend, and it still steered two major decisions successfully for years, because for those two decisions the founder's own judgment agreed with the board's. The escalation path is exactly what matters on the decision where the founder's judgment does not agree, which is the only decision the whole structure was ever really tested against.
  • "If we write down that the group has final authority, we are giving away too much control." A working group with genuine final authority on a defined, narrow slice of decisions is not a loss of control; it is a designed limit on how much any single person's judgment, including a founder's, can override on that specific slice without going through a named body first. The alternative, as Axon discovered, is that the informal control the founder thought they had turned out to be a decision the whole organization then had to answer for publicly.
  • "A working group that never blocks anything is doing its job well, because it means nothing bad is being proposed." A working group that has never once exercised its scope-defined authority to block, or has never had a recommendation escalated and tested, has an unproven charter, not a healthy one. The absence of conflict may mean the group's authority has never been needed, or it may mean, as at Axon for years before 2022, that no one has yet proposed the one thing the founder's own judgment would want to override.
  • "Once the charter is signed, the work is done." A charter with no review date or review trigger goes stale exactly as quietly as it once went unwritten: a release cadence changes, an escalation body is reorganized, a new jurisdiction's law creates a consultation right the charter never anticipated, and none of it shows up as a dramatic failure until the day it is tested. Write a review mechanism into the charter itself, the same way a change narrative names when it will be revisited (see Topic 9.6).
  • "A charter this formal only makes sense for a large organization; we are too small for this." The four parts scale down to two people as readily as they scale up to a multinational; a small organization that skips them is not avoiding unnecessary formality, it is answering the same four questions informally, which means answering them differently depending on who is in the room that day. Simplify the charter's language for a small team's reality; do not skip the structure.
  • "Dense, formal policy language makes a charter sound more authoritative." A member who cannot repeat back what the charter says in a plain sentence has not actually absorbed it, and a charter no one can repeat back accurately will not be followed under real pressure any more reliably than a euphemism-laden change narrative is believed. Write the four parts in plain declarative sentences, the same glance-test discipline that governs every learner-facing artifact in this program (see Topic 9.6).
  • "A vendor's AI model is the vendor's responsibility to govern, not ours." Once your organization deploys a vendor's model in a real decision that affects people, the governance obligation is yours regardless of who built the underlying system; a charter silent on how the group obtains vendor information is a charter that will eventually make a real decision on information it never had the standing to demand.
  • "The charter's job is finished once every part is written; enforcement will take care of itself." A written scope, a named escalation body, role-based seats, and a record format are the design; whether teams actually route disagreements through the escalation path, whether departures actually get filled by role, and whether the record actually gets used are separate, observable facts checked through the health signals in Section 3K, not guaranteed by the document's existence.
  • "An escalation body's turnaround time is a nice-to-have, not something worth pinning to a number." A stated turnaround time is what separates a real escalation path from a hopeful phrase; without it, an escalated disagreement can sit unresolved past the point where the underlying decision has already become irreversible, which produces the same practical outcome as never having named a body at all.
  • "Governance work is separate from the AI work itself, so it can wait until after the system is built." A charter designed after a system has already shipped is designed under exactly the pressure this topic warns against, with a real decision already made and a group being assembled to ratify it rather than to have genuinely reviewed it in time; design the charter before the first real decision it will need to catch, not after.
  • "The working group's scope covers this, so we are fine." A working group's scope only covers a decision if the decision is still open when the group sees it. A vendor contract signed, or a procurement commitment made, before the group ever reviews the model it binds the organization to is a decision the charter's scope never actually reached, whatever the charter says; the group is left rubber-stamping a choice procurement already made. Name procurement and contracting as a trigger for working group review in the charter itself, not only the model's own release.

Questions people ask

What is AI working group?
A standing body (also called a council, board, or committee) that makes or escalates the recurring decisions an AI program generates, such as whether a feature ships past a risk threshold. In this program, distinguished from a mere recurring meeting by whether four specific design decisions have been written into its charter.
What is working group charter?
The written artifact specifying a working group's decision scope and escalation path, its role-based membership, its meeting cadence, and its record format. Judged, like the change narrative, by whether it holds up under the hardest test it will ever face, not by how it reads on the founding day.
What is decision scope?
The precise statement of what a working group can decide finally, on its own authority, and what it can only recommend. A scope stated as a vague aspiration ("advises on AI ethics") resolves in favor of whoever holds informal power on the day of a disagreement.
What is escalation path?
The named body, distinct from the decision-maker whose choice triggered the disagreement, that a working group's recommendation or decision escalates to when a sponsor, founder, or executive wants to proceed differently, together with a stated turnaround time for that body to resolve it. The specific charter part missing from Axon's AI Ethics Board. More on Escalation path
What is role-based membership?
A membership list that names each seat by job function (for example, the VP of Engineering) rather than by the person currently holding it, paired with a designated-deputy rule specifying who covers the seat when the primary role holder is unavailable. Survives departures that name-based membership does not.

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →