Anthropic’s October 26 IPO filing gives investors a reason to ask a question usually left until a crisis: who can make an AI company miss its revenue target?
Imagine a CEO facing a launch delay after customers have been promised delivery and computing capacity has been reserved. The safety team wants more testing. Sales wants the release. The CFO sees a hole opening in the quarter.
Everyone endorsed responsible AI when it was a mission statement. Now someone has to sign off on a missed revenue target. Who can insist on the delay and can that decision survive the CEO’s objection?
That is where investors should test AI safety commitments: whether the company has allocated enough authority, money, and accountability to honor them under commercial pressure.
Protecting the business can mean missing the quarter. If the board won’t make that trade-off, the safety team is left arguing with people whose targets depend on the launch. Investors should ask who can stop it, what they can require before it resumes, and what happens if management goes ahead anyway.
Anthropic’s proposed charter gives its Long-Term Benefit Trust the right to elect a board majority, with additional approval requirements for one seat. That is substantial institutional power.
But the filed trust agreement sets a demanding threshold for outside shareholders seeking to enforce its terms: an investor or group must hold at least 10% of the voting power of “Non-Trust Voting Shares” continuously for at least a year and maintain that threshold throughout litigation. That calculation excludes trustee-owned shares and the founders’ Class F share; voting trustees, the company, and the Class F holder have separate enforcement rights.
These restrictions concern enforcement of the trust, not every shareholder remedy. An investor who does not qualify depends on other authorized parties to enforce it. Institutional protection is not the same as individual recourse.
Anthropic’s filing describes a release process. Before each model release, its Frontier Red Team analyzes capabilities that could assist cyberattackers. Development and deployment decisions are made case by case; assessments may lead the company to delay releases, restrict access, or withhold a model from general use. A board safety and risk committee helps oversee model safety and security.
The company’s separate Responsible Scaling Policy identifies decision-makers: under its Risk Report process, the CEO and Responsible Scaling Officer make the ultimate determination on the adequacy of risk assessments and resulting development or deployment plans. When proceeding relies heavily on the argument that Anthropic’s systems add relatively less risk given the risks unavoidably posed by other AI systems, both the board and Long-Term Benefit Trust must explicitly approve the Risk Report. These are substantive safeguards. Investors should nevertheless ask how disagreements among management, the board, and the trust would be resolved, who ensures that required approvals are obtained before proceeding, and whether those arrangements can withstand the financial pressure of a delay.
For boards across the industry, the practical test has three parts: authority to pause, money to sustain the delay, and consequences for bypassing the rules.
First, make pause and override authority explicit. Specify who can impose a pause, what evidence triggers it, what must change before restarting, and who can authorize an override. Any override should identify the decision-maker and record the reasons before release. The CEO should not be able to proceed first and explain later.
Second, fund the pause. Ask the CFO to model a 90-day delay before approving the launch. Which computing bills continue? What customer obligations remain? What happens to cash, guidance, and compensation? Ninety days is a stress test, not a universal standard. The point is to discover whether exercising the safeguard would trigger a financial crisis. A board that budgets for acceleration but leaves no capacity for restraint has already tilted the decision.
Third, decide what happens when management bypasses a safeguard. Require prompt escalation to the responsible board committee, a corrective-action plan, and clear consequences for unauthorized overrides. Companies buying AI should negotiate notice when promised protections materially change, rights to remediation, and an exit where feasible. Investors should know who can enforce each promise and what they can demand if it is broken.
Before launch, rehearse a failed safety review: who stops release, who funds the delay, and who authorizes a restart?
These arrangements will not guarantee safe AI. But they will expose the point at which a company’s promises exceed its willingness to pay for them.
The competitive opportunity is larger than avoiding a bad headline. Credible restraint can give enterprise customers a reason to trust a supplier with consequential work. That capability can support long-term relationships; it also costs money to maintain.
Investors should examine both sides before assigning it value. Boards should confront both before approving growth plans. A published policy is a starting point, not proof that a safeguard will survive commercial pressure.
A promise can survive on paper. The ability to honor it may not survive the budget.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.
The post Anthropic’s IPO filing raises a costly question: Who can say no? appeared first on Fortune.
