One boundary, two readings · Arc II
Arc II · Governance
Create accountable records when commitments take effect.
Governance asks who authorized the consequence and what record proves it. Authority Control governs the same boundary security relies on, read as accountability.
The decision record
Every determination produces a record.
Approvals, logs, tickets, chat threads, model outputs, and workflow events rarely produce one durable authority record. The commitment boundary emits one when the commitment takes effect.
Decision record · emitted at the boundary
- ActorWho or what acted, with identity linked to the commitment as a property of the record.
- Authority basisThe documented authorization path that supported the action.
- ScopeWhat was permitted, and the limits the evaluation applied, including configured cumulative limits across actions.
- DeterminationPermit, Defer, or Block, with the basis preserved either way.
- OutcomeOutcome information preserved with the decision history.
Retention and privacy
Decision records are designed so the integrity and existence of a determination remain verifiable over time while sensitive record content can be retired under the organization’s retention policy.
Evidence and oversight
Decision records provide control-operation evidence of the kind regulators and auditors ask organizations to produce, and they feed defined escalation and review paths.
What governance gains
Warning becomes constraint. Record becomes memory.
Attribution
A measurable unit exists for every governed decision. The committing actor is tied to the decision when the action becomes official.
Accountability
Who acted, on what authority, within what limits. A substantive evaluation, not a procedural approval.
Trace
Durable evidence that an action was properly authorized, and blocked attempts recorded with the same fidelity.
Memory
Records accumulate into organizational intelligence. Thresholds, escalation paths, and authority scopes become calibratable.
In the 2012 JPMorgan London Whale loss, the board had visibility, reporting, and escalation. It lacked a control that turned warning into constraint before the next trade. Read the case →
The cascade
Seven consequences of an ungoverned commitment.
When a commitment takes effect without a governed authority decision, the consequences do not stay contained. The cascade traces seven ways one ungoverned commitment propagates through the organization.
Recommendation
no gate
no record
no verification
no record
no verification
The Commitment Boundary
Ungoverned commitment
Unmeasured ROI
+
Limited Attribution
+
Ambiguous Authority
+
Unobservable Judgment
+
Invisible Overrides
+
Limited Learning Substrate
+
Unenforced Policy
+
Select a consequence
Unmeasured ROI
Adoption climbs. Measurable value does not follow.
Without decision-level records, value measurement operates at program level, where AI influence is diffuse and individual impact is invisible.
88% of organizations use AI. 6% capture meaningful enterprise value.
McKinsey 2025
Benefits are entangled with broader transformation. Investment expectations are widespread while only a minority report positive returns. Nearly half of business leaders say proving GenAI business value is the single biggest hurdle to adoption.
← Back to cascade
Limited Attribution
Outcomes cannot be traced to AI involvement.
When value is measured at program level, individual outcomes cannot be traced to AI involvement. Attribution requires a record linking a specific decision to a specific result. That record does not exist.
Wells Fargo, 2002-2016
3.5 million unauthorized accounts were created through the same credential path as authorized accounts. The account-opening commitment boundary had no structural mechanism distinguishing authorized from unauthorized creation. Attribution required a years-long forensic reconstruction because no record at the moment of creation captured who authorized each account.
The structural gap predates AI. Without a commitment-level record, attribution depends on forensic reconstruction rather than evidence captured at the point of commitment.
42% of companies abandoned most AI projects in 2025, up from 17% the prior year.
S&P Global 2025
We know we're using AI more and spending more. We cannot isolate which decisions improved because of AI and which did not. Projects are discontinued because value is unattributable.
← Back to cascade
Ambiguous Authority
The organization cannot trace what shaped the recommendation.
When outcomes cannot be attributed, the decision chain itself becomes opaque. The question shifts from "did AI add value?" to "who is actually deciding?" The pattern predates AI. AI accelerates it.
Boeing 737 MAX, 2015-2019
The FAA delegated certification authority to Boeing employees through the Organization Designation Authorization program. Delegates held formal authority in policy. At the certification commitment boundary, no structural mechanism verified that authority was exercised independently of management pressure.
Authority existed in the delegation structure. Whether that authority was genuinely exercised at the commitment boundary was unobservable.
40% of employees received AI-generated content passed off as substantive work, comprising 15% of all work content received.
BetterUp / Stanford 2025 · 1,150 U.S. employees
Evaluators "fall asleep at the wheel", retaining formal decision authority while effectively outsourcing cognitive effort. The output performs authority while lacking the judgment that would warrant it.
← Back to cascade
Unobservable Judgment
Considered review and routine acceptance are indistinguishable.
When authority is ambiguous, the quality of judgment exercised under that authority becomes unmeasurable. The organization cannot separate deliberate evaluation from routine acceptance. The pattern predates AI. AI makes it pervasive.
Enron, 1999-2001
The Finance Committee formally approved the SPE transactions. Board minutes recorded the approval. The approval relied on management representations rather than independently verified facts. Judgment was formally exercised: committee members voted, the decision was documented. Whether that judgment engaged with the substance of what was being committed was structurally unobservable.
Formal approval occurred. Substantive judgment was unverifiable at the commitment boundary. The record captured that a decision was made. It did not capture whether judgment was exercised.
82% of leaders use AI weekly, yet 43% report declining skill proficiency across their teams.
Wharton-GBK 2025 · Year Three Enterprise AI Adoption
We cannot tell who is genuinely evaluating AI output and who is rubber-stamping. The interaction leaves no observable trace. Usage is universal. Engagement is unmeasured.
← Back to cascade
Invisible Overrides
Acceptance, challenge, and override leave no structural record.
When judgment itself is unobservable, the specific interactions (acceptance, challenge, override) disappear entirely. The organization has no structured record of what humans did with AI recommendations.
Physicians used AI primarily to confirm existing hypotheses rather than expand their differential diagnoses.
Goh et al. 2024 · JAMA Network Open · Randomized Clinical Trial
When people push back on AI, we have no record of it. When they accept without review, we have no record of that either. The pattern, confirmation rather than challenge, is invisible at scale.
← Back to cascade
Limited Learning Substrate
Process improvement has no evidence base.
When override and acceptance patterns leave no trace, organizations have no feedback data from which to improve processes. Workflow redesign requires knowing which decisions benefit from AI involvement, and that knowledge does not accumulate.
Organizations with mature process governance are 3.3× more likely to scale AI successfully.
Accenture 2024 · 2,000 senior executives
We want to redesign our processes around AI, but we have no data on which decisions benefit from AI involvement and which need more human judgment. The 6% figured this out through custom effort. The rest are guessing.
← Back to cascade
Unenforced Policy
Governance describes what should happen. Enforcement is absent.
When process improvement has no evidence base, governance policies remain unenforced, describing intended behavior with no mechanism to verify compliance at the moment it matters. The pattern predates AI. AI widens the gap between policy and enforcement.
Wirecard, 2015-2020
Audit policy required independent verification of cash balances. At the commitment boundary where the auditor accepted evidence as verified, nothing enforced that the evidence was independent. Management controlled the evidence path. €1.9 billion in purported cash balances were confirmed through intermediaries the company controlled.
The policy described the requirement. The commitment boundary had no mechanism to enforce it. The auditor's acceptance proceeded without structural verification of evidence independence.
Only 1 in 4 organizations have fully operational AI governance controls.
AuditBoard 2025
We have policies. We've drafted frameworks. We reference NIST and ISO. Between the policy document and the moment someone acts on an AI recommendation, there is no enforcement mechanism.
← Back to cascade
Design partners
Accountability, at the point of commitment.
Every governed commitment traces to an actor, authority basis, scope, and determination.