The Decision the Board Owns Before the Budget Locks

The Q4 approval of 2027 AI spend is the only AI decision the board makes in its own name, and a board that signs it without a record of what it relied on has converted management's uncertainty into the board's signature.

The board is not being asked to approve artificial intelligence. It is being asked to sign. Every other act in a company's AI program is management's: the vendor selection, the deployment, the monitoring, the metrics, the policy, the committee charter. The approval of next year's spend is the one act that carries the board's name and no one else's, and the one act a director will be asked about, by name, if the question is ever put. The question in front of the board in the fourth quarter of 2026 is not whether the company should invest in AI in 2027; management has answered that, and the request on the agenda is the answer. The question is what the board will require to be true, and shown to be true, about the systems it has already funded before it attaches its signature to the next tranche. A board that approves on a briefing has relied on a characterization. A board that approves on evidence has relied on a record. The statute that protects directors who rely in good faith distinguishes between those two things, whether or not the board does.

The board's language, not management's

The vocabulary matters because the decision is easy to misdescribe as an operating matter. Management speaks of programs, roadmaps, use cases, and adoption; the board's vocabulary is narrower and older: approval, reliance, reservation, condition, and record. In that vocabulary the decision has four parts. The board approves spend. The board relies on what it is shown. The board reserves certain decisions and delegates the rest. And the board records what it did. Each of the four is entirely the board's, none requires a director to understand a model, and each has a failure mode visible only after the fact. The approval can be unconditional when it should have been conditioned; the reliance can rest on an assurance rather than an artifact; the reservation can go unmade, so that nobody knows who may stop a system; and the record can describe a briefing rather than a decision. This paper takes the position that the four share one cause: the board did not say, in advance and on the record, what it would need to see. (Position: that framing, and the ninety-day checkup it leads to, are Touch Stone Publishers frameworks. Nothing in this paper states or implies that any law, regulator, or court requires a board to run a checkup.)

The checkup, as Touch Stone Publishers proposes it, is the board's own test, run before the budget locks, of five things it can inspect rather than hear about for each material AI system already funded: a current inventory entry; a named accountable owner; an outcome measured against a defined denominator and baseline; a monitoring record; and a route to pause or decommission. (Position: the Touch Stone Publishers Evidence-of-Execution standard.) It asks three questions of each system: what changed, what is still exposed, and what must be decided. (Position: the Touch Stone Publishers three-question checkup.) It is proportionate to consequence, so that the full standard falls on the systems whose failure would reach a person, a customer, or a regulated obligation, and inventory and ownership fall on everything. (Position: Touch Stone Publishers consequence tiering.) And it is not a legal requirement. The reason to run it is that the board's own approval, reliance, reservation, and record depend on it, an argument this paper makes at the true strength of the law and no stronger.

Reliance, and what a briefing cannot carry

Under Section 141 of the Delaware General Corporation Law (State of Delaware, 8 Del. C. § 141(a), (c), (e), https://delcode.delaware.gov/title8/c001/sc04/, inspected September 17, 2026), the business and affairs of a Delaware corporation are managed by or under the direction of its board of directors except as the statute or the certificate of incorporation otherwise provides; the board may designate one or more committees; and a director is fully protected in relying in good faith upon the records of the corporation and upon such information, opinions, reports or statements presented to the corporation by any of the corporation's officers or employees, or committees of the board, or by any other person as to matters the director reasonably believes are within that person's professional or expert competence and who has been selected with reasonable care by or on behalf of the corporation (C-0020). Three things about that provision should be read exactly. The reliance it protects is reliance on records and on reports from identified people. The "selected with reasonable care" qualifier attaches only to the third category, outside persons, not to officers, employees, or committees. And the provision is general corporate law with nothing about artificial intelligence; it creates no AI oversight duty and no audit mandate (C-0020).

What the provision does supply is the shape of protected reliance, and that shape is the whole of the board's interest in the checkup. Reliance on a record presupposes that a record exists; reliance on an officer's report presupposes that the officer said something within that officer's competence. A director who has seen the inventory entry for a system, dated and confirmed, is relying on a record of the corporation. A director who has seen an executive's written acceptance of accountability for that system is relying on a statement by an officer within the officer's competence. A director who has heard that appropriate governance is in place is relying on a characterization, and whether that is the reliance the statute protects is a question this paper does not answer, because no Delaware court has decided an AI oversight claim of any kind (C-0021). The argument is therefore prudential, not doctrinal, and Touch Stone Publishers states it as a position: a board that wants its reliance to rest on something should make sure that what it is shown is a record or a report, not a description of one. A briefing cannot carry that weight, not because briefings are dishonest but because a briefing is management's account of its own work, and a board that never specified what it needs to see cannot refuse the account it is given.

The mechanism the board is inside

The reason boards end up relying on characterizations is structural rather than personal, and it runs in both directions at once. Governance structure is cheap to create and easy to disclose; execution evidence is expensive to produce and easy to defer. Downward, the board asks for governance and management delivers what a quarter allows: a policy, a charter, a committee, a briefing slot. The request is satisfied on paper, no one is assigned to produce the harder artifacts, and the minute records that oversight occurred. Upward, management owns the systems and the metrics, so management defines what the board sees, reporting adoption counts because those exist and outcome against denominator because it rarely does. The load-bearing step, in both directions, is the same: the board never defined the evidence that would satisfy it. (Position: Touch Stone Publishers calls this the Declarative Board Failure Pattern, a board that declares its expectations rather than defining the evidence that would show they were met; the third section develops it.) Both chains end in the same artifact gap, discovered when an acquirer, a regulator, or a plaintiff asks what the board decided about a specific system.

One dated data point shows how rarely the body signing the approval was, when the question was last asked, the body that had accepted the governance. McKinsey's March 2025 report, The state of AI: How organizations are rewiring to capture value (McKinsey & Company, March 2025 edition, https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/2025/the-state-of-ai-how-organizations-are-rewiring-to-capture-value_final.pdf, p. 3 and p. 26), built on an online survey fielded July 16 to 31, 2024 of 1,491 participants in 101 nations, reports that among respondents whose organizations use AI, 28 percent say the CEO is responsible for overseeing AI governance and 17 percent say the board of directors oversees it, with respondents on average naming two leaders as jointly in charge; McKinsey's November 2025 and August 2026 editions do not update these shares (C-0010). The figure is self-reported, its respondent population and question wording are McKinsey's, it is two survey cycles old, it is not a census of boards, and it is not evidence that board oversight causes better outcomes or that its rarity indicates negligence. It is used only to note that joint ownership was the norm, which is the form the unmade accountability conversation most often takes.

The exposure the board is signing: an assumption model

The exposure in this decision is not a dollar figure, and this paper presents none as a fact; the evidence base contains no verified settlement, penalty, remediation, or market-reaction figures. The exposure the board can reason about is a share of its own signature: the fraction of the 2027 request the board would be approving without any later way of knowing whether it succeeded. Touch Stone Publishers models that with every input labeled as an assumption and its basis stated, so that a board can replace each input once the inventory exists.

Index the 2027 AI request at 100 units. Input A is the share of the request attached to top-tier systems. Assumption: base 50, downside 65, upside 35. Basis: none in the evidence; a placeholder the board replaces with the tiered inventory the first thirty days produce. Input B is the share of top-tier spend attached to systems with no denominator defined before the spend. Assumption: base 60, downside 80, upside 40. Basis: in McKinsey's March 2025 edition, from the July 16 to 31, 2024 survey of 1,491 participants in 101 nations, less than one in five respondents whose organizations use AI in at least one business function, with those answering "don't know" removed, said their organizations were tracking well-defined KPIs for gen AI solutions, with Exhibit 4 showing 23 percent at organizations with 500 million dollars or more in revenue and 14 percent at smaller organizations (McKinsey, March 2025 edition, p. 10 and Exhibit 4; C-0011). That is a self-reported, dated, correlational figure from a different question than the one modeled, and McKinsey's later editions report newer figures on related subjects without repeating it; it is the basis for an assumption about the reader's company, not a measurement of it. The output is the share of the request approved without a later test of success: base 30 of 100 units, downside 52, upside 14. The range is wide because the inputs are assumptions, and the point is the denominator, not the number: the exposure is measured against the board's own signature, and the board can shrink it before the meeting by requiring the denominator as a condition of approval rather than discovering its absence afterward. This board-level model is scoped narrower than the portfolio-wide model in the Playbook and the CFO paper, where Input U states the denominator gap across the whole approved request at base 70, downside 85, upside 50; Input B here states the same gap only within the top-tier share Input A has already carved out, and the board's tighter base case (60 against 70) reflects the assumption, stated for this narrower population and unproven until the inventory replaces it, that top-tier spend concentrates in fewer, larger systems more likely to already carry some measure than the average system the portfolio-wide figure includes.

The request that will eventually come

The twelve-month horizon supplies two dates on which the board's record will be read: the Q4 2026 approval itself, and the 2027 proxy season, for which one episode in 2026 shows the form the question takes. Alphabet's 2026 proxy statement, filed April 24, 2026 (Alphabet Inc., DEF 14A, https://www.sec.gov/Archives/edgar/data/1652044/000130817926000342/goog-20260424.htm, Proposal 12, 2026 Proxy Statement pp. 90 to 92), contained a shareholder proposal, filed by the Shareholder Association for Research & Education on behalf of the Pension Plan of The United Church of Canada as lead filer, asking the board to amend the Audit Committee charter to provide formal oversight of responsible AI development and deployment and of AI-related human rights risks, arguing that October 2025 charter amendments had left an accountability gap and that the charter was unclear on whether the board formally oversees material AI-related risk; the company's opposing statement said the full board holds ultimate oversight of AI strategy and risk and that existing committee structures already provide focused, multi-layered oversight. At the June 5, 2026 annual meeting the proposal was not approved, with 461,472,553 votes for, 11,863,462,046 against, 45,493,465 abstentions, and 580,489,723 broker non-votes (Alphabet Inc., Form 8-K filed June 11, 2026, Item 5.07, https://www.sec.gov/Archives/edgar/data/1652044/000119312526267578/d57679d8k.htm) (C-0014). A shareholder proposal is an argument, not a finding; the company's response is self-reported; the proposal is not pending and won no meaningful support. Read one way, the vote says structure satisfied the owners. Read the other, it says a company can describe multi-layered oversight in a proxy while the proponent's question, whether the board formally oversees material AI-related risk, is answered with a description of committees rather than a record of decisions. This paper uses the vote as evidence of neither, only as notice of the sentence a board will be asked to write in 2027, and of the difference between answering with a minute that names a system and a paragraph that names a committee.

The ten-year horizon is different in kind. The systems will be replaced, the executives will move on, and most of the directors who sign the 2027 approval will not be in the room when the question is finally asked. What remains is the record, and the decision before the budget locks is a decision about what it will say: that the board received an update and approved the request, or that it defined the evidence it required, inspected what was produced, conditioned its approval on what was not, reserved the decision to stop, and wrote that down against a named system on a stated date. The rest of this paper is about making the second kind of record in the time the board has.

What the Law Requires and What It Does Not

No inspected authority imposes an AI-specific duty on a board to run a checkup, and the reason to run one is the board's own reliance, approval, and record, stated as a position and not as doctrine.

A board deciding whether to require evidence before it approves should first know that nothing requires it to. This section states the law at its true strength, weaker than the commissioning frame around AI governance usually assumes, because a board that believes it is complying with a mandate will run the checkup as compliance and a board that knows the decision is its own will run it as governance, and the two produce different records. Every statement below is confined to the authorities directly inspected in the evidence base; it is not a survey of all law, and sector regulators, contracts, procurement rules, and other states may impose requirements this section does not address.

Federal instruments bind agencies, and the voluntary framework is voluntary

None of the directly inspected federal instruments imposes an AI governance audit requirement on the boards of private companies (C-0028). OMB Memorandum M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, dated April 3, 2025 (Office of Management and Budget, https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf, pp. 1 to 2 and Appendix § 1(a)), applies to agencies as defined in 44 U.S.C. § 3502(1). GAO's September 9, 2025 report, Artificial Intelligence: Federal Efforts Guided by Requirements and Advisory Groups, GAO-25-107933 (U.S. Government Accountability Office, https://www.gao.gov/products/gao-25-107933, Fast Facts and Highlights), and its June 30, 2021 Accountability Framework, GAO-21-519SP (https://www.gao.gov/products/gao-21-519sp, Fast Facts and What GAO Found), describe federal requirements and an advisory framework for federal agencies and other entities, not a board mandate. NIST's Generative AI Profile, NIST AI 600-1, published July 26, 2024 (National Institute of Standards and Technology, https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence), is a companion to an AI Risk Management Framework that NIST describes as intended for voluntary use, and NIST's AI RMF FAQ (https://www.nist.gov/itl/ai-risk-management-framework/ai-risk-management-framework-faqs, question 5) answers "No" to whether organizations will be required to use the Framework (C-0028). Executive Order 14365 of December 11, 2025, Ensuring a National Policy Framework for Artificial Intelligence (90 FR 58499 to 58501, FR Doc. 2025-23092, https://www.govinfo.gov/content/pkg/FR-2025-12-16/html/2025-23092.htm, section 3), whose section 3 directs the Attorney General to establish an AI Litigation Task Force to challenge state AI laws inconsistent with the order's policy, is directed at state law; and the White House's March 20, 2026 Legislative Recommendations: National Policy Framework for Artificial Intelligence (https://www.whitehouse.gov/wp-content/uploads/2026/03/03.20.26-National-Policy-Framework-for-Artificial-Intelligence-Legislative-Recommendations.pdf, Part VII, p. 4), whose Part VII asks Congress to preempt state AI laws that impose undue burdens and whose other parts propose duties on AI platforms, are directed at Congress. Neither imposes a duty on the boards of private companies (C-0028). That negative finding covers only the inspected instruments; it is not a statement that no federal requirement touches private-company AI use.

The disclosure regulator's posture points the same way. In remarks to the SEC Investor Advisory Committee on December 4, 2025 (U.S. Securities and Exchange Commission, Remarks of SEC Chairman Paul Atkins to the Investor Advisory Committee, https://www.sec.gov/newsroom/speeches-statements/atkins-remarks-iac-120425, final paragraphs), Chairman Atkins favored principles-based, materiality-driven disclosure over prescriptive, technology-specific checklists (C-0024). A Chairman's remarks are not a Commission rule or a binding interpretation, and they are not a rule against AI-specific disclosure. They are a reason a board should not expect a forty-question AI audit to arrive from the SEC, and a reason the checkup is built around materiality rather than a fixed list: the board chooses its top tier by consequence, the principle the Chairman named.

Delaware is general, silent on AI, and narrowing rather than expanding

Section 141 has already been quoted; it is general corporate law with nothing AI-specific in it (C-0020). The directly inspected Delaware authorities, Section 141, the Delaware Artificial Intelligence Commission statute at 29 Del. C. §§ 9041C to 9047C (State of Delaware, https://delcode.delaware.gov/title29/c090C/sc04/), and the Court of Chancery's oversight opinions in SolarWinds (decided September 6, 2022, affirmed May 17, 2023) and B. Riley (decided March 30, 2026), do not establish a universal, AI-specific oversight duty for the boards of private companies, and as of September 2026 no Delaware court has decided an AI-oversight Caremark claim (C-0021). Three cautions travel with that finding. It is a negative finding about inspected authorities, not a survey of Delaware law. It may not be converted into "boards have no AI oversight duty," because Caremark oversight duties for mission-critical compliance risk remain and are fact-specific (C-0021). And the inspected opinions are not the most recent Delaware oversight decisions; a later Chancery oversight ruling of August 2026 surfaced in verification and was not inspected, so SolarWinds and B. Riley are presented as two decisions read directly, not as the current state of the case law.

Delaware's Artificial Intelligence Commission, at 29 Del. C. §§ 9041C to 9047C, is a sunset-limited state-government advisory body created to recommend statewide processes, principles, and guidelines for executive, legislative, and judicial agencies and to inventory generative AI usage in those agencies; it addresses no private company (C-0022). A director who has heard that Delaware has "an AI statute" should know that this is it.

SolarWinds is the opinion that matters most for how a board should think about AI as risk. In Construction Industry Laborers Pension Fund v. Bingle, C.A. No. 2021-0940-SG (Delaware Court of Chancery, decided September 6, 2022, https://courts.delaware.gov/Opinions/Download.aspx?id=337580, PDF pp. 18 to 21, printed pp. 16 to 19; affirmed by order of the Delaware Supreme Court, No. 411, 2022, May 17, 2023, 297 A.3d 1083 (TABLE)), the court dismissed Caremark claims arising from a cybersecurity incident under Rule 23.1. It treated the risk of third-party cyberattack as business risk rather than as compliance with positive law, stated that whether Caremark should be applied to business risk remains an open question in Delaware, required particularized allegations of bad faith, and added that it was possible to envision an extreme hypothetical in which a bad-faith failure to monitor business risk could create liability (C-0022). The sharper "distinction" language a reader may have encountered is a footnoted quotation from an earlier case, not this court's holding. The opinion is fact-specific, not a universal rule for all states, sectors, or entities, and it does not hold that boards have no duty to oversee cyber or AI risk (C-0022). What it tells a board is that most AI deployment risk, like third-party cyber risk, will probably be characterized as business risk, that whether oversight doctrine reaches business risk is a question the court has declined to close, and that the door it left open is marked bad faith.

B. Riley is the opinion that matters most for how a board should think about outcomes. In Marchner v. Riley, C.A. No. 2025-0164-LWW (Delaware Court of Chancery, decided March 30, 2026, https://courts.delaware.gov/Opinions/Download.aspx?id=393800, printed pp. 20 to 21, PDF pp. 22 to 23; the opinion's face misprints the decided year as 2025, and a Sidley Austin commentary of April 21, 2026, https://ma-litigation.sidley.com/2026/04/bad-investments-not-bad-faith-caremark-claims-have-limits/, corroborates a spring 2026 decision without stating the date), the court dismissed a derivative complaint under Rule 23.1 concerning a financial investment. It cautioned that a plaintiff may not "equate a bad business outcome with a breach of the duty of loyalty," and it framed the first-prong oversight inquiry, quoting Marchand, as whether directors made "a good faith effort to implement an oversight system and then monitor it" and attempted "to assure a reasonable information and reporting system exists" (C-0023). B. Riley is not an AI case, it does not immunize boards from oversight claims, and it is a fact-specific decision in a demand-futility posture (C-0023). The board should nonetheless read the framing closely, because it describes the court's question in the language of the artifacts the checkup produces: an oversight system, monitored, resting on a reasonable information and reporting system. A board that defined the information it required, received it, and recorded what it decided has a record of that effort; a board that received updates has a record of updates.

The objection that fiduciary framing is a category error

The objection should be stated at full strength because it is largely right. Delaware imposes no AI-specific duty. SolarWinds places the relevant risk in a category to which oversight doctrine may not extend at all. B. Riley refuses to convert a bad outcome into bad faith. No court has heard an AI oversight claim. On that record, a paper that tells a board it must run a checkup because its fiduciary duties require it would be misdescribing the law, and this paper does not say that. The concession is complete: as to mandate, the framing is a category error.

What survives the concession is narrower and is stated as a Touch Stone Publishers position. First, the board's approval of spend is its own act regardless of whether any duty attaches to the systems the spend buys, and a board that wishes to say what it relied on has a reason to make sure a record exists, independent of doctrine. Second, the open question SolarWinds left is open in both directions; a board cannot know today whether its AI systems will be treated as business risk beyond Caremark's reach or as the extreme hypothetical the court declined to rule out, and the record it keeps now is the record it will have then. Third, where an AI system touches mission-critical compliance risk, Caremark oversight duties remain and are fact-specific (C-0021), and the tiering criteria are how a board finds out which of its systems those are. Fourth, the framing a court has recently used, good-faith effort to implement and monitor an oversight system resting on a reasonable information and reporting system (C-0023), is better answered by a minute that names a system, the evidence inspected, and the decision taken than by a minute that names a topic. None of that is a duty. All of it is a reason. A board has no AI-specific legal obligation to run a checkup, and it has every prudential reason to be able to show what it relied on when it signed.

What the twelve months hold, and what the decade holds

Over the next twelve months the legal picture will move, and the board should expect it to move without producing a mandate. The federal posture is aimed at preempting state AI law rather than adding federal duties on boards (C-0028). NIST has announced that AI RMF 1.0 is under revision pursuant to the July 2025 AI Action Plan, no revised version had been published as of September 2026, and the framework's voluntary status is unchanged (C-0004). The disclosure regulator prefers materiality to checklists (C-0024). More shareholder proposals of the Alphabet kind should be expected in the 2027 proxy season (C-0014). A board waiting for the law to tell it what to require will wait through the approval, the proxy, and the next approval. Over the decade the picture is simpler: the question put to a former director will be what the board knew and decided about the systems it funded in 2026, and the answer will be whatever the minutes and the inventory show. The law does not require the checkup, and it does not supply the record. Only the board can do that, and only before the budget locks.

The Reserved Decisions

The board owns four decisions and nothing else in the AI program, and a board that makes those four on the record has governed while a board that reviews the systems has only been busy.

The board's work in the AI program is four decisions, not five and not three. (Position: the allocation follows the Touch Stone Publishers Governance Boundary Principle, which holds that the board governs and management manages, and that when either crosses into the other's territory the organization begins to fail quietly and then suddenly.) The four decisions are the tiering criteria, the evidentiary standard, the stop authority, and the conditions on approval. They are the board's because only the board can make them with the authority to refuse what does not meet them, and because they are decisions about what the board will accept and reserve rather than about how anything is built. Everything else, the inventory, the owner assignments, the denominators, the monitoring design, the decommission engineering, and the tier assignments themselves, is management's, and the board's proper posture on all of it is to receive the evidence and refuse it when it falls short. A director who finds the board's AI agenda full of the second list and empty of the first has found the boundary crossed.

The first reserved decision: the tiering criteria

The board decides, in consequence language, what makes a system material: what happens to a person, a customer, or a regulated obligation when the system is wrong or unavailable; whether the system's action is reversible; and whether its effect reaches financial reporting, the authorization of expenditures, or the safeguarding of assets. (Position: Touch Stone Publishers consequence tiering.) Written by the board in those terms, the criteria are a risk-appetite statement a director can read and a later board can find. Left to management, they become a technical taxonomy, calibrated to what the inventory tooling captures and expressed in model complexity and data sensitivity, unreadable by the body that should review them. The failure mode is delegation by inattention: the board discovers at the first incident that the system sat in a tier it never knew existed, defined by a rule it never saw.

The evidence that this decision matters is public-sector and is used for the shape of the argument, not as a fact about private companies. The OECD's Digital Government Outlook 2026 chapter, Adopting and governing AI in government, page dated June 15, 2026 (OECD, https://www.oecd.org/en/publications/digital-government-outlook_0496b2bc-en/full-report/adopting-and-governing-ai-in-government_7ef312a9.html, chapter summary, sections 4.5.4 and 4.5.5, and the guardrails passage before Box 4.6), draws on the 2025 OECD Digital Government Index survey of national governments' public-sector AI governance, to which 36 OECD members responded, Germany and the United States not participating, and every count below uses that denominator. It reports that in 2025, 30 of 36 had a regulatory oversight body or an ethical advisory body dedicated to AI, with functions oriented toward guidance and monitoring rather than hands-on auditing or enforcement; that only 14 of 36 require pre-deployment risk assessments and 11 of 36 conduct post-deployment audits; that most governments lack a central repository of AI use cases, with three maintaining mandatory repositories and ten optional ones, a count the source states without an explicit denominator; that only 10 of 36 report any financial or non-financial impact measurement of AI use cases in government; and it warns that without risk-based assessments, audit structures, accountability frameworks, and formal decision paths for high-risk uses, guardrails might remain symbolic rather than operational (C-0006). These are measured facts about national governments and may not be presented as facts about U.S. private companies. Their relevance to the first reserved decision is the phrase "for high-risk uses": the controls the OECD names as the difference between symbolic and operational guardrails make sense only once someone has decided which uses are high-risk, and that decision is the tiering criteria. A board that has not written them has left that difference to whoever fills in the inventory.

The second reserved decision: the evidentiary standard

The board decides what it will accept as proof. The standard is the five evidences, each in one of three states: absent, in which nobody claims the artifact exists; asserted, in which management states that it exists and the board has taken the statement; and inspected, in which a director or the committee acting for the board has seen the artifact itself. (Position: the Touch Stone Publishers Evidence-of-Execution standard.) Adopted by resolution, the standard means that a report on a top-tier system arriving without the five evidences is one the board has said, in advance, it will not accept, so the refusal is the execution of a decision already made rather than an act of hostility. The failure mode is adoption without amendment: a board that adopts management's proposed standard word for word, without the committee having tested each definition against one live system, has adopted a standard it does not own and will find it cannot refuse.

The corporate record shows why the standard must be the board's own, and it should be read with its limits. PayPal's 2026 proxy statement, filed April 7, 2026 (PayPal Holdings, Inc., DEF 14A, https://www.sec.gov/Archives/edgar/data/1633917/000119312526145721/d59508ddef14a.htm, 2026 Proxy Statement p. 38, "Corporate Governance: Board Oversight: AI Governance Framework," see also p. 37), describes an Enterprise AI Governance Framework adopted in 2025 and formalized through an Enterprise AI Policy and Standard and an AI Governance Charter, with an AI Governance Executive Council and AI Governance Working Groups, the Executive Council reporting into the Enterprise Risk Management Committee and ultimately to the board's Risk Committee and the Board (C-0012). Cognizant's 2026 proxy statement, filed April 17, 2026 (Cognizant Technology Solutions Corporation, DEF 14A, https://www.sec.gov/Archives/edgar/data/1058290/000130817926000290/ctsh014861-def14a.htm, 2026 Proxy Statement pp. 36 to 37, "AI governance and oversight"), reports full-board oversight of AI strategy with delegated committee responsibilities, a cross-functional AI Governance Committee, a Responsible AI Office that provides mechanisms for ongoing monitoring and risk assessment, vendor and client ecosystem governance controls, and Finance Committee updates on AI's impact on pricing models, productivity in IT service delivery, and new revenue opportunities (C-0013). Both are company self-descriptions of design; neither is independent evidence of effectiveness, establishes any causal link between structure and returns, or is a benchmark other boards must match (C-0012, C-0013). They are cited for one observation: a director reading either paragraph cannot tell from it what evidence the board required, or whether any director inspected an inventory entry, an owner's acceptance, a denominator, an exception, or a stop route for any named system. The disclosure rules did not ask, and the companies may well have done all of it. A disclosure of structure and a record of evidence are different documents, and only the board can decide which one it is producing.

The third reserved decision: stop authority

The board decides, for each top-tier system, whether the decision to pause or decommission is reserved to the board or delegated to the named owner with a duty to report its exercise within a stated period. Either answer is defensible. The absence of an answer is not, because the moment the question becomes urgent is the moment a customer-facing system has begun producing wrong outcomes, and a board that has not decided who may stop it will spend that moment finding out. The failure mode is discovering the answer under pressure. The decision has a second part boards often miss: whoever holds the authority, the route must have been exercised or tested at least once, with the test on record, because an untested stop route is an assertion, and a top-tier system with no tested route is reported to the board as an exception rather than folded into a summary.

Who owns the route is management's to produce and the board's to require, and the ownership is split in a fixed way so that no part of it goes unassigned. The Chief Risk Officer is first owner of the route and its trigger: the definition of what "wrong" means for the system and the point at which the pause is required. The Chief Operating Officer owns the fallback, what the business does while the system is paused, and the test that proves the route works. The Chief Information Officer or Chief Technology Officer owns the mechanism, the engineering and the contract terms that make the pause executable. (Position: the Touch Stone Publishers first-owner rule for the fifth evidence.) The accepted system owner exercises the pause on the trigger the CRO defined, without prior approval, under whatever delegation the board decided. The decision belongs above the line because it states what the organization is willing to have happen without a person. A board can evaluate whether it is willing for a system to keep making consequential decisions about people while a defect is investigated; it cannot evaluate the engineering of the kill switch. The board decides the first and requires the second.

The fourth reserved decision: the conditions on approval

The board decides what its approval is conditional on, and the Q4 meeting is the last moment at which the five evidences can be required rather than requested. The failure mode is the unconditional approval, which is the default, because a resolution that says nothing about evidence is an approval on whatever was presented. The decision follows the score. Each evidence is scored 0 if absent, 1 if asserted, and 2 if inspected, for a system score out of 10, and the score places the system in a band: Briefed at 0 to 3, Asserted at 4 to 6, Evidenced at 7 to 8, Governed at 9 to 10. (Position: the Touch Stone Publishers checkup diagnostic.) Spend attached to a Governed system is approved. Spend attached to an Evidenced system is approved subject to a single named condition, the one evidence that keeps it from Governed, with an owner, a date, and a consequence. Spend attached to an Asserted system is approved conditionally, with each missing or asserted evidence named as a condition with its owner, date, and consequence. Spend attached to a Briefed system is not approved. And regardless of band, spend attached to a system that lacks an inventory entry or an accepted owner is not approved until it has both, because a system that is not on the list or has no one accountable for it cannot be conditioned and cannot be stopped; that absent-owner rule operates on its own and not through the score. At the ninety-day point, spend attached to systems that missed their conditions is re-presented to the board for an explicit decision to continue, condition further, or stop. The condition is the board's expectation made enforceable at the one moment it carries the weight of the budget.

The scope decision the supervisors handed back

One event in 2026 shows what it looks like when the reserved decisions are handed to an institution from outside. On April 17, 2026 the Federal Reserve, the OCC, and the FDIC issued Revised Guidance on Model Risk Management (Board of Governors of the Federal Reserve System, SR 26-2, https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm, transmittal letter and attachment p. 3, footnote 3; issued as OCC Bulletin 2026-13 and FDIC FIL-15-2026). SR 26-2 supersedes SR 11-7 and SR 21-8, and the OCC bulletin rescinds OCC Bulletin 2011-12. The guidance is non-binding, the agencies expect it to be most relevant to banking organizations with more than 30 billion dollars in total assets, and it narrows the definition of a model to complex quantitative methods. In footnote 3 of its attachment it states that generative AI and agentic AI models "are novel and rapidly evolving" and "are not within the scope of this guidance," expects a banking organization's own risk management and governance practices to determine controls for such tools, and applies its principles to non-generative, non-agentic AI models (C-0030). That is a scope statement by the agencies, not an exemption from supervision and not a statement that such systems are unregulated; it applies to supervised banking organizations only.

Read in the board's language, the footnote is a reserved decision returned to the institution: for generative and agentic systems, the regime that best answered the objection that existing controls already cover this now says that the institution's own governance decides. Where management alone decides how those systems are governed and the board is briefed on the result, the upward crossing has occurred by default, and the board's record on the class of system the supervisors declined to cover is a minute of a briefing. Where the board responds by reviewing the generative systems themselves, the downward crossing has occurred and the criteria remain unwritten. The instruction is the same either way, and not confined to banks: when a regime hands a class of system back to the institution's governance, the board writes the four decisions and management produces the evidence.

The Declarative Board Failure Pattern

The pattern that most reliably prevents the four decisions from being made is not negligence but a misunderstanding of the job. (Position: the Touch Stone Publishers Declarative Board Failure Pattern.) Board members who believe their role is to declare expected behavior, rather than to model it, ask about it, and listen for it, have misunderstood it. Its pathology is structural: senior figures interact primarily with each other, hand down directives, and remain insulated from the organization they govern, so that the declaration is the last the board hears of the matter until something fails. Applied to AI the signature is recognizable. The board declares that it expects responsible AI, adopts a policy that says so, receives a charter that names a committee, and hears quarterly that the program is on track. Every one of those acts is a declaration; none defines the evidence that would show the expectation was met, and none puts a director in contact with an inventory row, an owner's written acceptance, a denominator, an exception, or a tested stop route. The board has said what it wants, not what it would accept as proof, and management, reasonably, supplies what it has.

The pattern and the four reserved decisions are mirror images: a board that skips the criteria has worried about high-risk systems without saying which; one that skips the standard has asked for assurance without saying what it consists of; one that skips stop authority has said the system must be stoppable without saying by whom; and one that approves unconditionally has voiced its expectations in the discussion and left them out of the resolution. The declarative board's minutes read that a topic was received. The board that has escaped the pattern produces minutes that read that a system was decided, on stated evidence, with named conditions.

The objection that this drags the board into management

The objection is the board's own and deserves a direct answer. A board that demands inventory rows, owner acceptances, denominators, exceptions, and stop-route tests can appear to cross the boundary this section has drawn. It would, if the board were producing any of them or specifying how they are produced. It does not, if the board is deciding what it will accept and refusing what falls short. The distinction is the difference between an audit committee that writes the accounting policy and one that declines to accept financial statements without an auditor's opinion. The checkup asks the board to do the second. It writes four decisions in its own language, receives what management produces against them, scores what it sees, and records what it decided. A board doing more has crossed downward; one doing less has been crossed from below. The four decisions are the line.

A Worked Case (Labeled Hypothetical)

The same board, the same request, and the same systems produce two different records depending on whether the checkup was run in the thirty days before the meeting, and only one of those records answers the question that arrives in February.

The company in this section is a labeled hypothetical. Call it Halvard Industrial, a Delaware-incorporated, exchange-listed distributor of industrial components in the upper mid-market, with a nine-member board, an audit committee that has taken AI oversight into its charter, and a technology function that has spent two years deploying systems under a Responsible AI Policy the board adopted in 2024. Neither the company nor any of its systems exists; every fact about it is invented to illustrate a Touch Stone Publishers position. Three systems will matter: a quote pricing engine that sets prices on roughly a third of order volume and adjusts them by customer segment; a procurement assistant, built on a vendor's agentic platform, that drafts and, within limits, issues purchase orders; and an applicant screening tool that ranks candidates for warehouse and branch roles. Management's 2027 request, presented to the committee in October 2026, indexes at 100 units, of which 55 are attached to expanding those three systems.

Halvard in October: what the committee has

At the October committee meeting the Chief Technology Officer presents the AI program update. The deck reports that 31 use cases are live, that adoption of the pricing engine has reached 34 percent of quotes, that the procurement assistant has issued 4,100 purchase orders in the year to date with a 96 percent straight-through rate, that the screening tool has reduced time-to-shortlist by 40 percent, that the AI Governance Council meets monthly, and that "appropriate monitoring and controls are in place for all production systems." The committee asks about vendor concentration and a customer complaint concerning a quote; management addresses both; the committee agrees to recommend the request in November. Every number in that paragraph is hypothetical, and every one is a count of activity. None is an outcome against a denominator, none names an owner, and none says which of the 31 systems make or shape a consequential decision about a person or a customer.

The public analog for what Halvard lacks is instructive because it is not a company. GAO's May 4, 2026 report, Artificial Intelligence: Uses and Risks for Small Business Contracting and Innovation Research, GAO-26-107828 (U.S. Government Accountability Office, https://www.gao.gov/products/gao-26-107828, What GAO Found and recommendation), found that the Small Business Administration paused all AI use in March 2025 while revising its AI policies, that the pause remained in effect as of April 2026 except for seven pilot or pre-pilot projects, and that SBA had not consistently met public AI use-case reporting requirements, publishing its first inventory in March 2026; GAO's one recommendation was that SBA establish policies and procedures for use-case inventory reporting, including defined roles and responsibilities and documentation of key decisions, a recommendation SBA agreed with and that remained open as of September 2026 (C-0009). The recommendation concerns use-case inventory reporting specifically, not agency AI governance generally, and the SBA case is not a prevalence claim about any other organization. It is cited because the three artifacts GAO found missing at one federal agency, an inventory, defined roles, and documented key decisions, are the first three things Halvard's committee did not receive in October, and because an organization can run under a standing reporting requirement for years and produce its first inventory only when someone with the authority to ask finally asks.

Path A: the November meeting without the checkup

On Path A the October recommendation goes to the November board with the same deck, condensed. The minute reads as follows.

Item 7. Artificial Intelligence Program and 2027 Investment Request. The Chief Technology Officer presented an update on the Company's AI program, including deployment metrics for production systems, the activities of the AI Governance Council, and the 2027 investment plan. Management confirmed that the program operates in accordance with the Responsible AI Policy adopted by the Board in 2024 and that appropriate monitoring and controls are in place. Directors asked questions regarding vendor dependencies, customer feedback, and the competitive landscape, which management addressed. Following discussion, on motion duly made and seconded, the Board approved the 2027 AI investment request as presented. No further action was required.

That minute is accurate; nobody misled anyone. It is also a record that contains no system name, no evidence, no owner, no condition, no reservation of stop authority, and no date other than its own: a minute of a briefing, honestly kept.

Path B: the thirty days before the November meeting

On Path B the committee chair, at the October meeting, declines to recommend the request as presented and asks for four things by the November board: an inventory of every system that makes or shapes a decision about a person, a customer, or a regulated obligation, tiered against criteria the committee will write that week; for each top-tier system, the five evidences in whatever state they exist; a named owner for each, with written acceptance where obtainable in time; and a proposed stop rule for each, the route and its trigger proposed by the Chief Risk Officer, the fallback and the test by the Chief Operating Officer, and the mechanism by the Chief Technology Officer, stating who may exercise the pause and whether the route has ever been exercised. The committee's criteria run to two sentences: a system is Tier 1 if, when wrong or unavailable, it changes a price a customer pays, commits the company's funds, or affects a decision about a person's employment; and every system, regardless of tier, must have an inventory entry and an owner. The full board adopts the standard and the criteria by written consent before November. The thirty-day path below sets out who does what in that month; here it is enough to say what the committee found.

The inventory took eleven business days and returned 27 systems, not 31, four having been duplicates or retired. Six met the Tier 1 criteria: the three named above and three others. The committee scored each evidence 0 if absent, 1 if asserted, and 2 if inspected. For the pricing engine, the inventory entry was inspected (2); the owner was the Vice President of Commercial Operations, who accepted in writing but noted she had no authority to pause the engine without the CTO's approval, so the acceptance was conditional and the owner was scored asserted (1); the success measure at launch was quote-to-order conversion, with no baseline for mispriced quotes and no measure of price dispersion by customer segment (asserted, 1); the monitoring consisted of a vendor dashboard whose last exception nobody could produce (asserted, 1); and no route had been defined by the CRO, no fallback or test existed with the COO, and no pause mechanism had been built (absent, 0). The system scored 5 of 10, Asserted band. For the procurement assistant, the entry was inspected (2); the owner was named as the Chief Procurement Officer, who had not yet accepted (absent, 0); the denominator was the straight-through rate, with no measure of erroneous or duplicate orders and no baseline (asserted, 1); the monitoring produced a weekly summary but no exception log (asserted, 1); and the assistant's authority to issue orders had a contractual limit, described by the CTO as the mechanism, that had never been tested by an attempt to exceed it and had no CRO-defined trigger or COO fallback (asserted, 1). It scored 5 of 10, Asserted band, and the committee noted separately that a system that issues purchase orders reaches the authorization of expenditures and asked the audit committee chair to confirm with the external auditor whether it sat within the scope of internal control over financial reporting. (Internal control over financial reporting, as defined for issuers in SEC Rule 13a-15(f) (U.S. Securities and Exchange Commission, 17 CFR 240.13a-15(f), https://www.ecfr.gov/current/title-17/chapter-II/part-240/subpart-A/subject-group-ECFR3a0ea3e2e2e4a9f/section-240.13a-15, source 68 FR 36666, June 18, 2003), is a process designed to provide reasonable assurance regarding the reliability of financial reporting and the preparation of GAAP financial statements, including policies and procedures over record-keeping, the authorization of receipts and expenditures, and the prevention or timely detection of unauthorized acquisition, use, or disposition of assets that could materially affect the financial statements; the rule does not mention AI, and the proposition that an AI system is within ICFR scope to the extent it affects those elements and outside it otherwise is an inference from the definition, not a statement the rule makes (C-0034).) For the screening tool, the entry was inspected (2); the owner was the Chief Human Resources Officer, who accepted in writing with pause authority (inspected, 2); the denominator was time-to-shortlist, with no adverse-impact measure by protected class and no baseline (asserted, 1); monitoring was absent (0); and the route was inspected (2): the CRO had defined the trigger as any vendor-reported model change or outage, the COO's fallback was manual shortlisting, the CTO's mechanism was a switch the CHRO could throw, and the route had been exercised once during a vendor outage, with the record produced. It scored 7 of 10, Evidenced band, which under the board's rule forces a single named condition rather than a list. The three remaining Tier 1 systems scored 9, 9, and 10, Governed band. The committee reported all six to the board by system rather than in aggregate.

Path B: the November minute

Item 7. Ninety-Day AI Governance Checkup and 2027 Investment Request. The Chair of the Audit Committee reported on the Committee's checkup of the Company's Tier 1 AI systems, conducted against the tiering criteria and Evidence-of-Execution standard adopted by the Board by written consent dated [date]. The Committee inspected the inventory of 27 systems, confirmed on [date], and reviewed the five evidences for each of the six Tier 1 systems. The Committee reported the Quote Pricing Engine at 5 of 10, Asserted band (inventory inspected, 2; owner asserted, acceptance conditional on pause authority, 1; outcome measure asserted, no baseline for mispriced quotes, 1; monitoring asserted, last exception not produced, 1; stop route absent, 0), the Procurement Assistant at 5 of 10, Asserted band (inventory inspected, 2; owner absent, 0; outcome measure asserted, 1; monitoring asserted, 1; stop route asserted, mechanism untested and no defined trigger or fallback, 1), and the Applicant Screening Tool at 7 of 10, Evidenced band (inventory inspected, 2; owner inspected, 2; outcome measure asserted, no adverse-impact measure, 1; monitoring absent, 0; stop route inspected, 2). The three remaining Tier 1 systems were reported at 9, 9, and 10 of 10, Governed band. RESOLVED, that 2027 spend attributable to the three Governed systems is approved. RESOLVED, that 2027 spend attributable to the Quote Pricing Engine is approved conditionally, and that the conditions are: (a) written owner acceptance by the Vice President of Commercial Operations including authority to pause the Engine without prior approval, by [date]; (b) a defined denominator and baseline for wrong outcomes, being mispriced quotes by customer segment, by [date]; (c) production to the Committee of a monitoring log and its most recent exception, by [date]; and (d) a stop route comprising a trigger defined by the Chief Risk Officer, a fallback and a recorded test owned by the Chief Operating Officer, and a mechanism delivered by the Chief Technology Officer, by [date]; and that failure to satisfy any condition by its date suspends further disbursement pending an explicit decision of the Board. RESOLVED, that 2027 spend attributable to the Applicant Screening Tool is approved subject to the single condition that a monitoring record for the Tool, measuring adverse impact by protected class against a baseline set by [date], be produced to the Committee with its most recent exception by [date], failing which further disbursement is suspended pending an explicit decision of the Board. RESOLVED, that 2027 spend attributable to the Procurement Assistant, notwithstanding its Asserted band score, is not approved under the Board's absent-owner rule until a named owner has accepted accountability in writing, at which point the request returns to the Board with the remaining conditions for an Asserted system; and that the Audit Committee will separately confirm the system's treatment under the Company's internal control over financial reporting. RESOLVED, that the decision to decommission any Tier 1 system is reserved to the Board; that the Chief Risk Officer is first owner of each Tier 1 system's stop route and trigger, the Chief Operating Officer of its fallback and test, and the Chief Technology Officer of its mechanism; and that the authority to exercise the pause is delegated to the system's accepted owner with a duty to inform the Committee Chair within 24 hours of any exercise. Director [name] recorded a concern that quote-to-order conversion may reward underpricing and asked that the denominator in condition (b) for the Engine address margin as well as error; the concern is recorded. The Committee Chair will confirm satisfaction of each condition and the item returns to the Board at its [month] meeting as a standing exception until all conditions are satisfied.

The two minutes describe meetings about the same request, and both are honest. The second contains seven things the first cannot: named systems with tiers; the evidence inspected, with each artifact's state; decisions, with conditions and dates; accountable owners, by name; stop authority, stated and allocated; a concern, recorded; and a follow-up with an owner and a date. Counsel should review any minute language against the company's own practice and jurisdiction; the language above illustrates a position, not legal advice.

February and March: the arrival moments

In the hypothetical, two things happen in the first quarter of 2027. In February a shareholder submits a proposal, modeled on the one Alphabet received in 2026 (C-0014), asking Halvard's board to amend the audit committee charter to provide formal oversight of AI-related risk and arguing that the charter is unclear on whether the board formally oversees material AI risk. In March the pricing engine, after a vendor model update nobody flagged to the committee, begins quoting a customer segment below cost.

On Path A the board's answer to the February proposal is a description of its structures: the Responsible AI Policy, the Governance Council, the charter language, the quarterly updates. It is the answer Alphabet gave, and Alphabet's shareholders accepted it by a wide margin; the vote is evidence neither that the answer was sufficient nor that the proponent was right (C-0014). What the Path A board cannot say is what it decided about any named system, because it decided nothing about any named system. In March the board learns of the pricing incident from the Chief Executive, who learned of it from a regional sales director, who learned of it from a customer. The first question in the room is who can pause the engine. The answer, found under pressure, is that the CTO can, the Vice President of Commercial Operations cannot, nobody has defined a trigger or a fallback, and the vendor contract is silent. The engine is paused on the fourth day. The board's record of the system, when later requested, is the November minute, which does not mention it.

On Path B the board's answer to the February proposal is different in kind. It can say that it adopted tiering criteria and an evidentiary standard by written consent in October 2026, that it inspected the five evidences for six Tier 1 systems in November, that it approved three, conditioned two, and withheld one under its absent-owner rule, that it reserved decommissioning to itself, assigned the stop route to the CRO, COO, and CTO by part, and delegated exercise of the pause to accepted owners with a 24-hour reporting duty, and that the conditions were satisfied on stated dates or the spend was suspended. Whether the proponent is persuaded is beside the point; the board has answered with a record rather than a description. In March the pricing engine's mispricing trips the wrong-outcome denominator set under condition (b), which was mispriced quotes by segment against a baseline established in December. The monitoring log, produced under condition (c), records the exception at 08:15. The Vice President of Commercial Operations, holding pause authority under her written acceptance, pauses the engine at 09:40 on the trigger the CRO defined, without asking anyone; the COO's fallback, manual quoting from the prior price book, is in place by 10:30; and the committee chair is informed by 10:00 under the 24-hour rule. The vendor model update that caused the drift is recorded as a "what changed" the checkup's first question should have surfaced, and the notification gap is itself logged as an exception. The board's record of the system is the November minute, the December baseline, the March exception, and the pause, each a document a later reader can check.

What the case shows and what it does not

The case does not show that Path B prevents incidents; the engine mispriced on both paths, because the vendor updated the model on both paths. It does not show that Path A breached any duty; the inspected Delaware authorities establish no AI-specific oversight duty and no court has decided such a claim (C-0021), and the Path A board acted in good faith on what it was shown. What it shows is narrower and, within its narrowness, decisive for a board. The thirty days before the approval determine what the board holds four months later. On Path A the board holds a minute of a briefing and discovers, at the moment of the request, that it has no record of a decision. On Path B the board holds a record of what it decided, by whom, on what evidence, and with what conditions, and the incident is handled by the person the board named, on the trigger the CRO defined, with the fallback the COO owned, and reported on the timeline the board set. The budget was the same. The signature was the same. What sat beneath it was not.

Thirty Days to the Next Meeting

A board can go from a briefing to a decision record in one meeting cycle if the committee chair asks for four things on day one and refuses to recommend the request without them.

The second path in the worked case is available to any board before its own November meeting on one condition: the committee chair must ask for the four things on day one and be willing to say, in October, that the committee will not recommend the request without them. Everything else in the thirty days is management's work and is feasible in the time, because none of it requires building anything; it requires writing down what exists and admitting what does not. Each step below carries an owner, a deliverable, a measure, and an escalation. (Position: the path is a Touch Stone Publishers framework and not a requirement of any law or regulator.) The board's part is concentrated in the first three days and the last five.

Days 1 to 3: the board writes

• Step 1. The committee chair sets the ask. Owner: the chair of the committee that holds AI oversight, with the board chair's concurrence. Deliverable: a one-page memorandum to the chief executive, dated, stating that the committee will recommend the 2027 AI request only on receipt of a tiered inventory, the five evidences for each Tier 1 system in whatever state they exist, a named owner per system, and a proposed stop rule per Tier 1 system (route and trigger from the Chief Risk Officer, fallback and test from the Chief Operating Officer, mechanism from the Chief Information Officer or Chief Technology Officer), by day 25. Measure: the memorandum is sent by day 1 and acknowledged in writing by the chief executive by day 3. Escalation: a chief executive who declines the ask, or proposes to substitute the existing program update, says so to the full board at the November meeting, in the minute, before the vote.
• Step 2. The committee writes the tiering criteria. Owner: the committee, drafting; the general counsel and corporate secretary, recording. Deliverable: two sentences in consequence language, stating what places a system in Tier 1 (its effect, when wrong or unavailable, on a person, a customer, a regulated obligation, the authorization of expenditures, or the safeguarding of assets) and stating that every system regardless of tier carries an inventory entry and an owner. Measure: the criteria are adopted by the full board by written consent, dated, by day 3, and the consent is in the minute book. Escalation: a dispute among directors about the criteria is resolved by the full board before the inventory begins, because criteria written after the inventory are written to fit it.
• Step 3. The board adopts the evidentiary standard. Owner: the full board, on the committee's recommendation. Deliverable: a resolution adopting the five evidences with the definitions of absent, asserted, and inspected, and stating that a report on a Tier 1 system arriving without them will not be accepted as a basis for approval. Measure: adopted by written consent by day 3, with the committee having tested each definition against one live system before recommending it. Escalation: a standard the committee has not tested against a live system is not put to the board; the test takes one afternoon.

Days 4 to 15: management produces

• Step 4. The chief executive names the inventory owner and the system owners. Owner: the chief executive. Deliverable: a memorandum naming one executive accountable for producing the inventory by day 15, and a provisional single owner for each system already known to the chief executive to make or shape a consequential decision. (Position: the Touch Stone Publishers Single Owner Protocol, one named human accountable to the board for each material system; and beneath it the Touch Stone Publishers Accountability Contract Model, under which accountability is a conversation held before results are demanded, covering what is to be done, what authority is granted, what success looks like, and by when.) Measure: the memorandum is dated by day 5 and every provisional owner has been told, in writing, the four terms of the conversation. Escalation: a system for which no executive will accept provisional ownership is reported to the committee chair by name on day 6, because a system nobody will own is the first exception the checkup has found.
• Step 5. The inventory is built and tiered. Owner: the inventory owner named in step 4, with every business unit head answerable for completeness in their unit. Deliverable: a register of every system in production or pilot that makes or shapes a decision about a person, a customer, or a regulated obligation, with what it does, whom it affects, which decisions it touches, its provisional owner, its tier under the board's criteria, and the date the entry was confirmed. Measure: the register is delivered by day 15 with a date and a confirming signature on every row, and the number of Tier 1 systems is stated. Escalation: a business unit that has not confirmed its rows by day 12 is named to the chief executive that day and to the committee chair on day 15, and is treated as having Tier 1 systems until it shows otherwise.
• Step 6. Owners assemble the five evidences for each Tier 1 system. Owner: each provisional single owner. Deliverable: for each Tier 1 system, the inventory entry; the owner's written acceptance or a statement of what prevents it (usually the absence of pause authority); the success measure defined at launch, its baseline if any, and the current figure, together with an honest statement of whether wrong outcomes are measured at all; the monitoring record and the last exception it produced, or a statement that none exists; and the pause or decommission route in its three parts, the trigger from the Chief Risk Officer as first owner, the fallback and any test record from the Chief Operating Officer, and the mechanism from the Chief Information Officer or Chief Technology Officer, together with who may exercise it and whether it has ever been exercised. Measure: every Tier 1 system has a file with five labeled sections by day 20, each section marked by the owner as absent, asserted, or inspectable. Escalation: an owner who cannot produce the last exception the monitoring generated reports that fact to the committee as an exception in its own right, rather than describing the monitoring process.

Days 16 to 25: the committee inspects

• Step 7. The committee inspects and scores. Owner: the committee, in a working session with each Tier 1 owner in the room for their own system and no one else's. Deliverable: for each Tier 1 system, a score of 0 to 2 per evidence (absent, asserted, inspected) for a total of 10, with the artifact inspected named beside each score, and a portfolio view by system, not in aggregate. Measure: every Tier 1 system is scored by day 25 and the score sheet is an attachment to the committee's report to the board. Escalation: a system whose owner presents a summary in place of the artifact is scored asserted for that evidence and the substitution recorded; scores are not negotiated.
• Step 8. The committee drafts the approval conditions and the stop allocation. Owner: the committee chair, with the general counsel. Deliverable: for each Tier 1 system, a draft resolution following its band: approved at Governed, approved subject to a single named condition at Evidenced, approved conditionally with each condition, its owner, its date, and the consequence at Asserted, not approved at Briefed, and not approved under the absent-owner rule wherever the inventory entry or accepted owner is missing; together with a draft allocation naming the Chief Risk Officer as first owner of each route and trigger, the Chief Operating Officer of each fallback and test, and the Chief Information Officer or Chief Technology Officer of each mechanism, delegating exercise of the pause to accepted owners, reserving decommissioning to the board, and setting a reporting rule for any exercise. Measure: the drafts are circulated to the full board with the pre-read by day 27, three days before the meeting. Escalation: a Tier 1 system for which management proposes terms more lenient than its band forces is flagged in the pre-read as a disagreement between committee and management, so that the full board decides it in the meeting and the disagreement is minuted.

Days 26 to 30: the board decides

• Step 9. The board decides and minutes each system. Owner: the full board; the corporate secretary records. Deliverable: a minute in the decision form, naming each Tier 1 system, its score, the evidence inspected, the decision, the conditions and dates, the owner, the stop authority, any concern recorded, and the follow-up. Measure: the minute contains every Tier 1 system by name, and no Tier 1 system is approved on terms more lenient than its band and the absent-owner rule force. Escalation: any director who believes a condition is insufficient says so in the meeting and the concern is recorded by name; a recorded concern is a decision the board made.
• Step 10. The committee chair owns the follow-through. Owner: the committee chair. Deliverable: a tracking sheet of every condition, its owner, its date, and its status, reviewed at every committee meeting, with the item standing on the board agenda as an exception until every condition is satisfied. Measure: at day 90 after the approval, the share of approved spend attached to Tier 1 systems that have satisfied every condition, reported by system. Escalation: spend attached to any system that missed a condition is suspended under the resolution's own terms and re-presented to the full board for an explicit decision to continue, condition further, or stop; the suspension is not waived by the committee.

The board's part of that path, steps 1 to 3 and 9 to 10, fits in one written consent and one meeting. The middle is management's, and a management that cannot produce an inventory and five files per Tier 1 system in three weeks has told the board something it needed to know before it signed. Reliance under Section 141(e) is on records and on reports from officers within their competence (C-0020); the thirty days make sure that what reaches the board is exactly that.

Ten director questions

The following ten questions are a tool for a single director; any one can be asked at the next meeting without preparation, and the form of the answer shows which record the board is producing. They are framed around the board's own decisions rather than management's production, because management's answers usually come more readily than the board's, and that asymmetry is the finding of this paper in miniature.

1. When we approved last year's AI spend, what did the resolution say we were relying on, and can we produce it? A satisfactory answer is a minute naming a document. An unsatisfactory answer is the deck.
2. In our own words, adopted by this board on a date we can cite, what makes an AI system material to us? A satisfactory answer is a resolution with criteria in consequence language. An unsatisfactory answer is a reference to management's classification or to the policy.
3. Have we ever said, in advance and in writing, what evidence we require before approving spend on a material system, and have we ever refused a report for lacking it? A satisfactory answer is a standard and at least one refusal. An unsatisfactory answer is that management has always been forthcoming.
4. For our highest-consequence system, which of us has personally seen the inventory entry, the owner's acceptance, the denominator, the last exception, and the tested stop route? A satisfactory answer is a director's name beside each artifact. An unsatisfactory answer is that the committee is comfortable.
5. Who may pause that system today without asking anyone, did this board decide that, and where is the decision written? A satisfactory answer is a name, a resolution, and a date. An unsatisfactory answer is that management would of course inform us.
6. What decisions about AI have we reserved to ourselves, and what have we delegated, and would the corporate secretary be able to find that allocation in the minute book? A satisfactory answer is a list the secretary can retrieve. An unsatisfactory answer is that it is understood.
7. Which of our approved 2026 AI spend is attached to systems whose success we defined before we spent it, and which is attached to systems where we will never be able to know? A satisfactory answer is a share, by system. An unsatisfactory answer is a description of the benefits case.
8. If we received a shareholder proposal tomorrow asking whether this board formally oversees material AI risk, would our answer describe our committees or cite our decisions? A satisfactory answer is a sentence that names a system and a date. An unsatisfactory answer is a paragraph that names a charter.
9. When a regulator or supervisor has told us that a class of system is outside a regime we rely on and left to our own governance, did we decide anything in response, or were we briefed? A satisfactory answer is a minute recording a decision. An unsatisfactory answer is that the risk function is handling it.
10. If one of us were asked in five years what this board decided about the systems we funded in 2026, which document would we hand over, and does it contain the word "decided"? A satisfactory answer is a decision minute with a system name in it. An unsatisfactory answer is the last quarterly update.

Minute language for the four reserved decisions

The following resolutions record the four reserved decisions in a form a corporate secretary can adapt. They are a Touch Stone Publishers position offered as illustration; counsel should review any minute language against the company's own practice, charter, and jurisdiction, and nothing here is legal advice.

Tiering criteria. RESOLVED, that for purposes of the Board's oversight of the Company's use of artificial intelligence, a system is designated Tier 1 if, when wrong or unavailable, it changes a price a customer pays, commits or disburses the Company's funds, affects a decision concerning a person's employment, credit, health, or access to a service, or bears on a regulated obligation of the Company; that every system that makes or shapes a decision about a person, a customer, or a regulated obligation, regardless of tier, shall carry a current inventory entry and a named accountable owner; that management shall assign tiers against these criteria and the [Committee] shall review the assignments; and that any disputed assignment shall be determined by the Board and recorded.

Evidentiary standard. RESOLVED, that the Board adopts, as the standard of evidence on which it will rely in approving expenditure on any Tier 1 system, the following five artifacts: a current inventory entry; the written acceptance of accountability by a named owner; an outcome measure defined against a stated denominator and baseline, including a measure of wrong outcomes; a monitoring record together with its most recent exception; and a route to pause or decommission the system, exercised or tested, with the record of the test; that each artifact shall be reported to the Board as absent, asserted, or inspected; and that a request for expenditure on a Tier 1 system that is not accompanied by these artifacts in the stated form shall not be presented to the Board for approval.

Stop authority. RESOLVED, that the decision to decommission any Tier 1 system is reserved to the Board; that for each Tier 1 system the Chief Risk Officer is first owner of the pause route and its trigger, the Chief Operating Officer is owner of the fallback and of the test, and the Chief Information Officer [or Chief Technology Officer] is owner of the mechanism; that the authority to pause a Tier 1 system on its defined trigger is delegated to that system's accepted owner, to be exercised without prior approval, with a duty to inform the Chair of the [Committee] within [24] hours of any exercise; that each Tier 1 system's pause route shall be tested at least [annually] and the test recorded; and that any Tier 1 system lacking a tested pause route shall be reported to the Board as an exception at its next meeting.

Approval conditions. RESOLVED, that expenditure for [year] attributable to each Tier 1 system is approved, approved conditionally, or not approved as set out in Schedule [A] hereto, which names for each system the evidence inspected, the decision, each condition with its accountable owner and date, and the consequence of failure to satisfy it; that failure to satisfy any condition by its date suspends further disbursement attributable to that system pending an explicit decision of the Board; and that the Chair of the [Committee] shall report to the Board at its meeting following the [ninetieth] day after this approval the status of every condition, by system.

Four resolutions, one schedule, and a standing exception on the agenda are the whole of the board's apparatus. A board that has them and nothing else has done its own work and left management's to management.

Close

The board that requires evidence before it signs has raised the expectation of every executive beneath it, and the record it leaves is the only measure of the board that will outlast the people in the room.

The highest function of a leader is not to manage performance but to raise expectations, and the difference between the two is the difference between a board that receives what a quarter allows and a board that says, in advance and on the record, what it will accept. (Position: the Touch Stone Publishers Expectation Elevation Model.) A board that approves on a briefing has managed performance: it has compared this year's deck with last year's and found it adequate. A board that approves on evidence has raised the expectation of every executive who reports to it, because the chief executive must now produce to the board what the owners produce to the chief executive, and an owner who was once asked how the rollout was going is now asked for a denominator and the last exception. That cascade begins with four resolutions the board can pass before its November meeting. The model's failure mode is expectation without contract, a standard raised and left unassigned; the accountability conversation between the chief executive and each owner makes the higher expectation real, and the board's inspection of the owner's written acceptance is how it learns the conversation took place.

The twelve-month horizon runs from the Q4 2026 approval, at which the four reserved decisions are made and the conditions minuted, through the ninety-day point at which the conditions are tested, to the 2027 proxy season, in which the board describes what it decided rather than how it is organized. The ten-year horizon is the record those minutes become. The systems will be replaced, the regimes revised, the directors changed. What remains is whether, in this year, this board wrote down which systems it regarded as consequential, what evidence it required, what it reserved the right to stop, and what it decided about each named system on what evidence.

The Legacy Test applies to governance records as much as to people. (Position: the Touch Stone Publishers Legacy Test, which holds that the measure of a board or a governance model is not whether it worked while its architects were present but whether what they produced can be carried forward.) The measure of this board will not be whether its AI systems performed while these directors sat; it will be whether the record it kept lets the next board know what was decided, by whom, and on what evidence. The approval before the budget locks is the moment that record is written, and the checkup, developed in full in The Ninety-Day AI Governance Checkup Playbook from Touch Stone Publishers, is how the board decides what it will say. The board can approve on evidence or on a briefing. Both are signatures. Only one is a decision.

<!-- TSP:END -->