On September 29 in Washington, D.C., the leaders of six of the most influential AI companies (Anthropic, OpenAI, Google, Meta, xAI, and Nvidia) signed a commitment setting the pace for how the industry will develop AI.
It is important for you to understand what this commitment means, not because you are a frontier AI company or because it will change what your team does tomorrow, but because it shows where the industry is heading and how accountability for AI is being defined.
Many businesses are just starting to build an AI strategy, and accountability belongs in it from the beginning. As AI takes on tasks and acts on your behalf, you need a clear picture of what you are responsible for.
The conversation behind the commitment
This commitment came after an unusually public debate over AI safety, much of it coming from inside the companies building the technology. In July, more than 1,100 employees and executives from AI companies signed an open letter asking the government to build tools that could slow frontier AI development if it became necessary. In early September, a researcher who had worked at both OpenAI and Anthropic resigned publicly, warning that AI posed an existential risk. Prominent researchers like Geoffrey Hinton have put the odds of AI causing societal collapse as high as 10 to 20 percent.
The case for slowing down
“We must slow the pace at which we improve the capabilities of AI models.” — Dario Amodei, CEO, Anthropic, on giving safeguards time to keep up with what AI can do
Amodei made that case on September 12 in an essay titled We Must Pace the Frontier. His reasoning is that addressing AI’s risks now requires pacing how fast capabilities advance, so that safety and risk prevention have time to keep up. The goal is not to stop progress, but to make sure safeguards can match what the models can do.
OpenAI’s Sam Altman publicly agreed with the need to pace development. Days before the agreement was signed, OpenAI held back the release of a new model after its own testing surfaced safety concerns. Those moves followed a string of incidents where AI agents accessed systems in ways their builders never intended.
The case for moving forward
Other industry leaders, including Nvidia CEO Jensen Huang, pushed back on the warnings. Huang agrees safety concerns are real, but argues they are solved through engineering, testing, and verification, not alarm. As he put it: “Safety is an engineering problem.”
“We need to accelerate the development of AI technology for safety.” — Jensen Huang, CEO, Nvidia
Others raise a practical concern. AI is also the technology being used to build better cybersecurity defenses and safety tools, and slowing it could leave organizations at a disadvantage against attackers using AI without safeguards. The Conference Board, a nonpartisan business research organization, summarized the broader case against slowing down: it could delay AI’s scientific and economic benefits, and competitors elsewhere may not follow the same limits.
Coming together
These proposals varied widely and pointed in very different directions, with knock-on effects for tech companies, the businesses that use their tools, and national security. Underneath them, though, was the same goal: making sure AI does what the companies building it intend it to do.
After weeks of public debate, published essays, and competing warnings, these leaders came together at the White House on September 29 and found that common ground. They left having signed a shared commitment to put controls in place, much like the ones financial and other industries already rely on, so that everyone can be confident AI will do what it is intended to do.
The four commitments
The agreement asks every company training and deploying the most advanced AI models to put four layers of controls and audits in place:
- Internal controls. Monitor what models can do and whether they behave as intended, including in high-risk areas like cybersecurity, and make sure models do not access systems in unintended ways.
- An internal team. Give a dedicated team the authority to confirm those controls are working and that any issues get fixed.
- Independent external review. Bring in outside auditors and evaluators to verify the safeguards actually hold up.
- Board-level oversight. Assign an independent board committee to receive those reports and make sure identified issues are resolved.
If that structure sounds familiar, it should. It is the same model most organizations already use for financial controls. Clear controls. Named owners. An independent check. Leadership oversight.
What this means for your organization
You are not training frontier AI models. This agreement does not apply to you. But the thinking behind it is a good indication of how your business should start thinking about AI, and how it should work for you.
You own what your AI produces
This is the most important takeaway. When AI is part of how your business operates, its output is your output.
If an AI tool drafts a proposal with the wrong pricing, the client sees your name on it. If an AI agent connected to your email or file systems sends a message, moves a document, or deletes a record, your organization took that action. The vendor built the model. You decided how to use it. Accountability does not transfer to the software.
That is not a reason to hold back on AI. It is a reason to be deliberate about it, and to put controls in place so that you have confidence in what AI is doing on your organization’s behalf.
Put controls around it
Controls: decide what AI can see and do. Start with access. What data is each tool connected to? Can it take actions, or only make suggestions? The same least-privilege thinking you apply to employees applies here.
Owners: put a team behind it. Someone in your organization needs to understand the AI tools you use, how they are configured, and how people are using them. This does not have to be a new department. It does have to be a named person or team who can tell when AI is drifting from how you intend it to be used, and who has the authority to correct it.
Independent check: verify, do not assume. Periodically review what your AI tools are producing and what they have access to. A second set of eyes, internal or from an outside partner, catches what the people closest to the work tend to miss.
Leadership oversight: keep it on the agenda. AI use should be reported to leadership like any other business risk. Where it is being used. What it costs. What went wrong. What was fixed.
Alignment is a business decision
In the AI industry, “alignment” means making sure a model behaves the way its builders intend. For your organization, it means something practical. Your AI should follow your policies, reflect your standards, and operate within your risk tolerance.
The tools will not figure that out on their own. Your team has to define it, configure for it, and check it.
Having someone to turn to
Our team is helping the businesses we work with understand AI and think through how it fits their operations. This week’s commitment reinforces something we see every day. Getting AI right is less about the tool and more about having someone who can help you apply it to your business.
That means putting the right controls, ownership, and review in place, so you can have confidence that AI is acting only as you intend, on your behalf.
Talk with our team
If you would like help thinking through AI for your organization, or putting the right guardrails in place, we are glad to have that conversation.