The AI oversight a board
can only prove if it recorded it first.
There is a scene that repeats in boardrooms everywhere. Management presents what it is doing with AI, directors ask how peers are handling the topic, someone suggests another update next half, and the meeting moves on. The topic was on the agenda. The record will say the board followed it. None of the instruments that began addressing AI oversight ask that. On 4 December 2025, the SEC Investor Advisory Committee approved a recommendation that the Commission require companies to define what they mean by “artificial intelligence”, disclose board oversight mechanisms, if any, and report separately on material deployments. The phrase “if any” does the heavy lifting: it concedes that many have none, and turns the absence into something disclosable. The recommendation is neither a rule nor Commission guidance. Two members voted against, two abstained, and the Commission has not acted on it as of this reading. What matters is the axis of the question, and the axis is assignment.
Recognition, assignment and evidence, in that order.
Inside the same document, the committee cites three surveys by named third parties. Read in sequence, they describe a ladder.
- 01
60% of S&P 500 companies treat AI as a material risk
Joint Deloitte and USC Marshall report, cited by the IAC.
Recognising the risk is the widest step on the ladder, and the cheapest to reach. It records that the topic arrived at the company.
- 02
About 40% have charged a board committee with AI oversight
EY Center for Board Matters, October 2025, cited by the IAC.
An organisation that recognises the risk has not necessarily named anyone. This is the step the instruments ask about.
- 03
15% disclose board oversight of AI
ISS research, cited by the IAC.
An organisation that has named someone cannot necessarily show it to whoever asks. It is the narrowest of the three steps.
The IAC also records that shareholder proposals asking boards to measure AI impact quadrupled in 2024 over the prior year. The rates above are the committee citing third parties, not SEC findings, and should be read that way. The documents cited are listed at the end.
One policy, two states.
The policy stays silent
- Generally no voting recommendation based on a company's AI oversight
- Generally no voting recommendation based on a company's AI disclosure
- Structure left open: named directors, the full board, a dedicated committee or an existing one
The policy asks who was in charge
- Reviews the company governance practices
- Identifies which directors or board committees were charged with overseeing AI risks
- Assesses the board response and the associated disclosures
- May recommend voting against the appropriate directors
There is a structural precedent in the same document. For environmental and social matters, the policy records that, absent explicit board oversight, it may recommend voting against audit committee members. That is a punitive default for non-assignment. For AI that default does not yet exist. The reading that AI sits one step behind ESG on the same ladder, in the same document and under the same issuer, is Alkemia's, not Glass Lewis's.
The policy leaves the structure open. What it presupposes is that an answer exists.
Alkemia reading
AI shifts the evidentiary terrain.
What the instruments share is the moment at which the demand arrives. Delaware corporate doctrine offers the clearest mechanism. The Caremark standard, settled there almost three decades ago, holds directors liable for oversight failure in two situations: failing to install a reasonable information system, or ignoring the signals it produced.
In the reading of Pierluigi Matera, published on the Oxford Business Law Blog in March 2026, AI leaves the Caremark legal standard intact and shifts the evidentiary terrain on which good-faith oversight is demonstrated. The reason is structural: AI systems have come to produce the very information the board oversees, such as compliance monitoring, anomaly detection and whistleblower triage. They accelerate reporting and also shape what gets reported. Oversight migrates from signals people can detect to algorithm-mediated black boxes, and the second Caremark prong is compromised at the root, because a board cannot consciously disregard an alert that never appeared.
The doctrine is settled. Matera grounds the application of Caremark to AI oversight in that doctrine, and the article cites no Delaware ruling on an AI oversight failure. That is the standing in which this reading should reach the board.
The operational point follows. The question of assignment lies dormant until the moment it stops being dormant, and that moment is precisely the one in which it can no longer be answered well. A board does not enter an assignment into the minutes after the harm and obtain any effect from it. The window for the record precedes the event that makes the record necessary, and it closes without warning.
A partial functional analogy.
The American discussion usually reaches Brazil as an imported novelty. Brazilian law contains only part of the oversight function that discussion describes.
Arts. 142, I and III, and 153 of Law 6,404/76, the Brazilian corporations statute, give the board the general direction of the business, supervision of the officers, the power to examine books and papers and to request information, and a general duty of care. That is where the oversight function lives. Where the failure involves unlawful acts by other administrators, art. 158, §1º adds a specific rule: an administrator is liable who is complicit, who is negligent in discovering the acts, in the statute's words negligenciar em descobri-los, or who, knowing of them, fails to act to prevent their practice. Dissent entered into the minutes exonerates, within the limits of the provision.
Brazilian doctrine reads these situations through vigilance and culpa in vigilando. Even so, the object of §1º is narrower than that of the American doctrine: it addresses unlawful acts by other administrators, whereas Caremark addresses oversight of the company information systems, controls and operations. The mental element differs too: §1º allows liability for negligence in discovering those unlawful acts, while Caremark requires bad faith, shown by a conscious failure to oversee. The functional analogy between the regimes is only partial.
CVM votes help set the bounds of that comparison. In the Forjas Taurus case, in 2020, the vote of the reporting commissioner, Gustavo Gonzalez, treated the board's monitoring duty in light of the general duty of care, as an obligation of means rather than of result, in an analysis that links the supervisory competence of art. 142, III, with art. 153. In the T4F case, in 2025, the reporting commissioner, João Accioly, wrote that the Caremark standards are, in his words, por aqui em certa medida recepcionados, received in Brazil to a certain extent. Commissioner Marina Copola joined the acquittal with reservations about the reasoning, without addressing that reception expressly. These passages, on their own, do not show a settled CVM position equating the tests of the two regimes. Neither case concerns AI.
§3º adds the consequence that matters to this discussion. In publicly held companies, the joint liability of §2º, covering breach of legal duties necessary to the company's normal functioning, is restricted, save for §4º, to administrators who have received by provision of the bylaws the specific assignment to fulfil those duties. Assignment in the bylaws is legally operative in that specific regime. A committee charter, an internal policy, a responsibility matrix or minutes may carry evidentiary and governance weight, and are not automatically equivalent to the bylaw provision that §3º requires.
This does not mean that all liability is determined by the bylaws, nor that the absence of a designation produces immunity. Liability for one's own conduct remains, as do the duty of care, the board's supervisory function, the rule of §1º and the joint liability of §4º for anyone who knows of the breach and does not report it to the shareholders' meeting. The effect of the absence of a bylaw assignment on §2º is not expressly resolved. The restrictive reading, supported by §3º, is that in publicly held companies the civil joint liability of §2º does not expand automatically without a specific bylaw assignment. That produces no immunity. Individual liability under the caput, the article's main provision, and under §1º remains, as does the situation of §4º and, on a separate plane, the duty of care of art. 153, always depending on the facts, the imputation and the causal link. In sanctioning proceedings, CVM votes have diverged on using art. 144 to assign administrative liability in matters the bylaws do not allocate. The divergence was explicit in PAS CVM 19957.004388/2023-41, decided in 2025: the reporting commissioner invoked the provision to reach the entire executive management, the diretoria, in a vaguely described area; the voto-vista, an opinion delivered after a request to review the record and diverging on the reasoning, required a functional link and the elements of the administrative offence, and the CVM Collegiate unanimously acquitted the technology and operations officers. That controversy is administrative and does not resolve the civil joint liability of §§2º and 3º.
The two T4F votes, both for acquittal, converge on the insufficiency of a merely formal provision for controls. For João Accioly, merely formal provisions for internal controls would not suffice to fulfil the duty to supervise adequately if, in practice, the controls were not used, or if the board members ignored the alerts they generated. For Marina Copola, the mere existence of internal policies, detached from a critical analysis of their effectiveness, is not enough to determine whether the administrators are properly fulfilling the duty of care. The transposition to AI is Alkemia's reading: the record of the assignment documents the function received, and the assessment of how it was exercised also considers whether the controls are adequate and actually work, and the response to the alerts they produce.
The mechanism has a Brazilian case that needs no hypothetical. In PAS CVM 24/2006, an officer accused of failing to establish effective internal controls over an area that reported to him was unanimously acquitted by the CVM Collegiate, in a 2013 judgment in which the reporting commissioner read the supervisory dimension of the duty of care as subject to an assessment of the reasonableness and adequacy of controls. On the mandatory appeal that follows an acquittal, the recurso de ofício, the Conselho de Recursos do Sistema Financeiro Nacional, the last administrative instance, converted that acquittal into a fine in February 2018, on a majority decided by the President's casting vote, with the appeal's reporting councillor on the losing side. More than eleven years passed between the 2006 inspection that gave rise to the case and the CRSFN tiebreak.
The approximation to the American doctrine is a comparative reading by Alkemia. The Brazilian decisions and votes examined in this analysis do not show a settled institutional position equating the tests of the two regimes. Its application to a company or to a specific incident depends on the bylaw structure, the functions actually exercised and the applicable sector rules.
What the board decides before it needs to.
- A
Assignment exists only if it sits in a dated instrument
A named body without a record fails the same test as no named body at all.
What answers the question is a dated document, and its class matters. The bylaws are the constitutive instrument, and it is to the bylaws that Brazilian law attaches joint-liability consequences in a publicly held company, a companhia aberta. Board charters, committee terms of reference and minutes are organisational instruments: they carry weight as evidence of who did what, and they do not produce that effect. A presentation stating that the board follows AI is neither of the two. The distinction reads as formalism until the day someone asks for the document. The document shows who received the oversight function. In Alkemia's reading, showing that the function was exercised also takes evidence that the controls worked and that alerts got a response. Two votes in the T4F case at the CVM maintain that a merely formal provision for controls is not enough: the reporting commissioner's, to fulfil the duty to supervise; Commissioner Marina Copola's, to assess whether the duty of care is fulfilled.
It is worth measuring the distance between the two things today. If a company board were asked right now who answers for AI oversight, the answer either exists in some record or exists only in the memory of whoever was in the room.
- B
The choice of oversight structure belongs to the company
Named directors, a dedicated committee, an existing committee or the full board are all defensible.
None of the instruments prescribes a design. Glass Lewis lists four accepted arrangements and leaves the decision with the company. The IAC recommendation contemplates disclosing oversight mechanisms, if any. In Alkemia's reading, the board should define and record the arrangement suited to the company's risks.
That freedom of design spares the board a discussion that usually consumes more time than it deserves. The choice of arrangement is secondary to the fact that one exists. A board that spends two quarters deciding whether to create a technology committee or attach the topic to the audit committee spends those two quarters with no recorded assignment.
- C
Disclosing is a decision separate from assigning
The 60 / 40 / 15 ladder shows the two coming apart in practice.
About 40% have assigned oversight to a committee and 15% disclose that oversight. The two figures come from separate surveys, with their own samples and methodologies, and they measure different cohorts. The comparison holds only at an order-of-magnitude level. At that level it supports a modest and useful reading: there is a meaningful set of companies that decided internally and have not yet recorded it publicly. For those, the cost of disclosing is close to zero and the benefit appears exactly in the situation where the question is asked by someone holding a vote.
The Glass Lewis policy states that companies developing or using AI should disclose the board role in oversight, including how they keep directors up to date on the topic. That last part tends to be forgotten: assignment presupposes a body capable of exercising it, and that capability is also a matter of record.
One meeting today, or no answer afterwards.
The board that answers this today spends one meeting. The one that answers after the incident finds that the answer is no longer available. For the board that needs to record the assignment and the CEO who answers for it, the review starts in three records: the instrument that names the responsible body, the class of that instrument, and the separate decision to disclose.
Discuss your decisioncontact@alkemia.io
Ten documents, read at source.
- SEC Investor Advisory Committee, recommendation approved 4 December 2025
- Glass Lewis, 2026 Benchmark Policy Guidelines, United States
- Matera, Algorithmic Oversight, Oxford Business Law Blog, March 2026
- Brazilian Law 6,404/76, consolidated text, arts. 142, 144, 153 and 158
- CVM, PAS 19957.004388/2023-41, judgment concluded 5 August 2025
- CVM, PAS 24/2006 (Telemig Celular), judgment of 18 February 2013, vote of reporting commissioner Otavio Yazbek
- CRSFN, minutes of the 410th Judgment Session, 20 and 21 February 2018
- CVM, PAS RJ2014/13977 (Forjas Taurus), vote of reporting commissioner Gustavo Gonzalez, 30 January 2020
- CVM, PAS 19957.003158/2024-46 (T4F), vote of reporting commissioner João Accioly, 11 November 2025
- CVM, PAS 19957.003158/2024-46 (T4F), separate vote of Commissioner Marina Copola, 11 November 2025
Board oversight of AI, as a matter of record.
- Does the SEC require companies to disclose board oversight of AI?
- No. What exists is a recommendation of the Investor Advisory Committee, approved on 4 December 2025 with two votes against and two abstentions, that the Commission require that disclosure. It is neither a rule nor Commission guidance, and the Commission had not acted on it as of this reading.
- Does Glass Lewis recommend voting against directors over AI oversight?
- Absent material incidents, the 2026 policy states it will generally not make voting recommendations based on a company’s AI oversight or disclosure. Where material harm is attributable to insufficient oversight, the policy identifies which directors or committees were charged with overseeing AI risks and may recommend voting against the appropriate directors.
- Which AI oversight structure should the board adopt?
- None of the instruments prescribes a design. Glass Lewis accepts named directors, a dedicated committee, an existing committee or the full board, and the choice belongs to the company. The IAC recommendation contemplates disclosing oversight mechanisms, if any. In Alkemia's reading, the board should define and record the arrangement suited to the company's risks.
- Does the Caremark doctrine apply in Brazil?
- Not as a rule. The approximation to arts. 142, 153 and 158 of Law 6,404/76 is a partial functional analogy. In the T4F case, decided by the CVM in 2025, the reporting commissioner wrote that the Caremark standards are received in Brazil to a certain extent, and Commissioner Marina Copola joined the acquittal with reservations about the reasoning, without addressing that reception expressly. On their own, these passages do not show a settled CVM position equating the tests of the two regimes. The comparative reading is Alkemia's, and its application to a concrete case depends on separate legal assessment.
- Is recording the assignment in the bylaws mandatory?
- No. What art. 158, §3º does is make the bylaw assignment legally operative in restricting the joint liability of §2º in publicly held companies. Board charters, committee terms of reference and minutes carry weight as evidence of who did what, without automatically producing that effect, and the effect of a total absence of assignment is not expressly resolved.
- Does recording the assignment prove that the board exercised oversight?
- Not on its own. The record shows who was in charge. In the T4F case, in 2025, reporting commissioner João Accioly stated in his vote at the CVM that merely formal provisions for internal controls would not suffice to fulfil the duty to supervise adequately if the controls were not used or if board members ignored the alerts they generated. Commissioner Marina Copola stated in her vote that the mere existence of internal policies, without critical analysis of their effectiveness, is not enough to assess whether the duty of care is being fulfilled. Both voted to acquit. The votes do not concern AI. In Alkemia's reading, assessing AI oversight also considers whether the controls are adequate and actually work, and the response to alerts.
- What happened in the CRSFN case cited in this analysis?
- In PAS CVM 24/2006, an officer accused of failing to establish effective internal controls was unanimously acquitted by the CVM Collegiate in 2013. On the mandatory appeal, the CRSFN converted that acquittal into a fine in February 2018, on a majority decided by the President's casting vote. More than eleven years passed between the 2006 inspection that gave rise to the case and the CRSFN judgment in 2018.
- Does recording the assignment in the minutes after an incident work?
- A board does not enter an assignment into the minutes after the harm and obtain any effect from it. The window for the record precedes the event that makes the record necessary, and it closes without warning. That is why an assignment recorded today costs one meeting.
This analysis offers governance criteria. It is not a legal opinion and does not determine the liability of any administrator or company.
The text was checked against the recommendation approved by the SEC Investor Advisory Committee on 4 December 2025, the AI oversight section of the Glass Lewis 2026 Benchmark Policy Guidelines, the analysis published on the Oxford Business Law Blog in March 2026, the consolidated text of Brazilian Law 6,404/76 and administrative decisions of the CVM and the CRSFN on the allocation of competence and supervision among administrators, as read on 23 August 2026, and against the CVM's 2013 decision in PAS 24/2006 and the votes in the Forjas Taurus (2020) and T4F (2025) cases, read on 15 October 2026. The committee recommendation is neither a rule nor Commission guidance. The rates cited are the committee citing named third parties. The comparative reading between the Brazilian and American regimes is Alkemia's, and its application to a concrete case depends on separate legal assessment.