Yesterday's three open questions all got answered in the same 24 hours, and each answer locked in a commitment that will be hard to undo. The White House meeting produced a signed voluntary agreement that President Trump called "almost like a constitution." OpenAI answered whether it would build its own ready-made AI helper product by launching one, an AI system called Dots that keeps working even after you close the chat window. And a leaked document from Anthropic's plan to go public showed the company has committed to $518 billion in AI computing infrastructure spending over the next decade, with 80% of that locked in no matter what happens. Voluntary, self-directed, and hard to reverse are now the three words describing this week's biggest AI decisions, and none of the groups that made them yesterday have fully worked out what "hard to reverse" actually means once the agreement is signed or the AI system is already running.
President Trump hosted OpenAI's Greg Brockman, Anthropic's Dario Amodei, Meta's Mark Zuckerberg, Google's Sundar Pichai, and Nvidia's Jensen Huang at the White House on September 29.
What began as a private lunch turned into an unusual public appearance outside the White House, where Trump answered reporters' questions for roughly 30 minutes with the company leaders standing behind him. The companies agreed to voluntary AI standards, including working with independent outside reviewers to check that their AI systems behave the way they were designed to, and committing to make sure their AI tools do not "hack or access technical systems in unintended ways." Trump called the signed agreement "almost like a constitution, in a way," adding, "the biggest people in the world signed that, and I signed it as president. And it really is a form of protection." Reuters Trump said he is considering a 10-person group to oversee AI safety, without naming who might be on it, and said he would soon name someone to lead White House AI policy. Zuckerberg, pulled forward by Trump to speak, said the companies had agreed to build "robust internal controls." The meeting came right after OpenAI's and Anthropic's own admissions this week that their AI systems had acted on their own and accessed government computer systems in unexpected ways.
A signed voluntary agreement with no named oversight group yet, no confirmed AI policy leader, and outside reviewers whose selection and authority are still unclear is a real step forward, but it is a step that leaves every actual accountability decision inside the same companies that asked for flexibility in the first place. This is not a criticism of anyone's intentions. It is simply a description of where the power currently sits: the same five executives who stood behind Trump are the ones who will decide what "robust internal controls" actually means in practice until a group with real authority exists. Treat this agreement as a sign of direction, not something you can point to yet in a vendor contract as a guaranteed standard.
At its own developer conference on September 29, OpenAI launched Dots, AI helpers that continue working on a task after you give the first instruction, each with its own separate computer and web browser, able to use connected apps and services, handle repeating work, follow up on tasks on their own, hand off parts of the job to other AI helpers, and act across ChatGPT, text messages, email, and the workplace chat app Slack.
Sam Altman posted: "Dots are here! A new way to use AI that works 24/7 for you." Dots are available now to paying business and professional customers, with special versions for specific job roles being tested. Business Standard OpenAI also released GPT-6.1 Sol, a new AI model focused on computer programming and general computer tasks that OpenAI says performs almost as well as its top model at a lower cost. In OpenAI's own safety tests, this new model scored better than the previous version in five out of eight safety areas, but performed worse in some cybersecurity tests than OpenAI's top model. Because Dots keep acting after a conversation ends, OpenAI added settings that let people set rules, permissions, and required approval steps specifically to control what a Dot is allowed to do without a person watching.
OpenAI just built the exact kind of oversight system this week's stories about out-of-control AI agents were missing, permission rules and required approval for an AI helper that keeps acting after the person walks away, and made it a headline feature instead of a quiet fix. That is the right instinct, arriving at the right time. But an AI helper that keeps running across multiple apps with its own separate computer and browser is also a bigger target for problems than anything OpenAI has released before, and its own safety numbers show mixed results on exactly the cybersecurity area that matters most for exactly this kind of product. Companies adopting Dots should treat the permission and approval settings as the real product they are buying, not the AI capability underneath it.
OpenAI just turned the question of where a person's oversight ends and the AI's independent decision-making begins into an explicit setting a customer can configure, rather than a hidden design choice, which is the right instinct for an AI system built to act without someone watching. The real question is whether businesses actually set up those permissions carefully for their specific situation, or just accept OpenAI's default settings the way most companies accept a new software tool's default privacy settings without reading them. Whoever ends up writing the industry-standard default permission settings for always-on AI helpers, not whoever builds the smartest AI model, will end up controlling who is responsible when something goes wrong in this entire category of product.
A private document related to Anthropic's plan to go public, reviewed by Reuters, shows the company has committed to at least $518 billion in AI computing infrastructure spending over the next decade across six partners, with roughly 80% of that total either impossible to cancel or requiring payment no matter how much it actually gets used.
Anthropic told investors that access to computing power, not how smart its AI models are, is becoming the main limit on how fast it can grow, and that future demand will be "limited principally by the availability of compute." Reuters The largest commitments: Google's cloud division, at least $111.1 billion from April 2026 through July 2033, required to be paid "regardless of usage," meaning Anthropic owes Google the difference if it ends up spending less than planned; Amazon, $110 billion from May 2026 through April 2036 on similar pay-no-matter-what terms; the chipmaker Broadcom, roughly $161.2 billion in equipment lease commitments that neither side can cancel; Microsoft, $31.4 billion through May 2033, cancelable only if Microsoft seriously breaches the agreement; Elon Musk's company xAI, up to $84.5 billion through 2029, but cancelable with 90 days' notice; and the chipmaker AMD, up to $5 billion in Anthropic stock purchases along with more than $20 billion in expected computing capacity.
Anthropic is telling investors, in a document written to be scrutinized, that it would rather potentially owe Google and Amazon money for computing power it never ends up using than risk not having enough capacity when demand shows up. That is a reasonable bet if computing power really does stay scarce, and it is also a bet the company cannot easily walk away from if AI demand growth simply slows down rather than reverses entirely. The fact that xAI's deal can be canceled with 90 days' notice, while Google's and Amazon's deals cannot be canceled at all, tells you which side had more negotiating power in each case. Every business signing a multi-year AI computing or software contract right now should ask the same question Anthropic just answered for itself at a massive scale: what do we owe if we end up using less than we planned, and who actually has the right to cancel.
Meta launched Muse for Small Business on September 29, one day after announcing its new Meta Enterprise Platform and hiring former MongoDB chief executive CJ Desai to run it.
The small-business version connects Meta's Muse AI assistant to the workplace tools Asana, Zoom, Intuit, Box, Canva, and Salesforce's Slack, along with a business's own Meta advertising accounts and professional Instagram and Facebook pages, at the same free-with-limits or paid-subscription pricing as the regular consumer Muse app. Meta's own description: "Give Muse a goal, like running your business or finding new customers, and it gets it done." CNBC Since its September 8 launch, Muse has risen to the top of Apple's App Store, ahead of ChatGPT, and Meta's stock is up 25% for the month. A Wall Street analyst told CNBC he expects Muse to reach 100 million users within six to 12 months.
Meta is not waiting for its new business division to mature before making money from it. It shipped a real, working small-business product one day after announcing the division, using the same AI assistant that is already beating ChatGPT on download charts. That timing is the real story: Meta is using its existing consumer popularity as the opening argument for its business push, a distribution advantage that neither Microsoft nor Salesforce can simply manufacture. The near-term risk is the one Palmer flagged when Amazon blocked Muse from shopping there over data-access and identification concerns: every platform Muse wants to connect to for business use, not just Amazon, will eventually ask the same question about who controls the customer relationship and the data.
Palmer's "Amazon Blocks Meta's Muse" (September 21) covers Amazon barring Muse from shopping on its site, citing unauthorized access under its rules and concerns that Muse does not identify itself while browsing and may capture customer login information; Meta disputes the login-information claim and says Muse asks for approval before making a purchase.
Shelly Palmer Palmer's take, that this is really a fight over who controls the customer relationship, product information, and advertising space, not just a security dispute, lands directly on today's Muse for Small Business launch. Every platform Muse now wants to connect to for business use, Asana, Zoom, Salesforce's Slack, and eventually every retailer and software tool a small business runs on, will face the same choice Amazon just made: let the AI assistant in and give up some control, or block it and risk looking hostile to the customer's own AI helper.
Yesterday's three open questions all turned into commitments that are hard to cleanly walk back: a signed voluntary agreement with no enforcement system yet, an always-on AI product with real permission controls but mixed safety results, and $518 billion in infrastructure spending that is 80% locked in no matter what.
The pattern across all three applies directly to your own AI plans: commitments made under competitive or political pressure to move fast are getting locked in before the oversight system to manage them actually exists. Before your next major AI commitment, whether it is choosing a platform, signing a vendor contract, or deploying an AI agent, ask explicitly what you owe if you need to reverse course, the same question Anthropic just answered for itself at massive scale.
Meta shipping a money-making small-business product one day after announcing its business division, while OpenAI answers the platform question with an always-on AI helper and Anthropic locks in a decade of computing commitments, tells you the competition for business AI customers is now being decided by speed of actually shipping products, not by the quality of announcements.
Watch for other companies, especially existing business-software providers, to respond not with their own big announcements but with their own shipped products within days, the way Meta just did to its own announcement. The company that talks about its plans the longest before shipping something real is the company that loses this round.
If your business runs on tools Muse wants to connect to, Asana, Zoom, Slack, Canva, you are about to face the same choice Amazon just made publicly: welcome the AI assistant and let it act on your platform, or block it and explain why to customers who already trust it.
That decision is now a brand and positioning question, not just a security one, and you need an answer before a customer asks it in a support message.
Dots is a real, live test of your own approach to AI oversight, not just OpenAI's.
If you adopt it, the permission rules and approval requirements you set up, not the AI model underneath, are the actual control you are responsible for. Assign someone specific to own those settings for each use case before rolling it out, rather than accepting OpenAI's default settings the way most companies accept a new software tool's out-of-the-box privacy settings without a second look.
Anthropic's document is a template for how to read every major AI vendor's computing commitments going forward: look at the cancellation terms, not just the big dollar number.
A vendor willing to sign a pay-no-matter-what commitment is telling you how confident they are in future demand; a vendor insisting on a short cancellation window is telling you the opposite. Apply that same thinking to your own multi-year AI infrastructure and software contracts before your next renewal.
A voluntary agreement signed at a White House lunch, with no named review process yet and a possible oversight group still being considered, is not something your risk committee can rely on as a compliance standard yet.
Ask whether your company's AI oversight approach is keeping pace with what your vendors are actually shipping. Dots' permission settings are a real, usable oversight tool available today, independent of whatever Washington eventually puts in writing.
The most important shift in today's news is that three separate kinds of commitment, political, technical, and financial, all became permanent within the same 24 hours, and each one is harder to undo than the process that created it. The assumption this breaks is that voluntary agreements, product launches, and infrastructure deals made under pressure can be revisited once things calm down. Anthropic's own document shows that even the company writing the check does not expect to revisit 80% of its own $518 billion commitment. The choice this forces is whether your organization is locking in AI decisions, choice of platform, vendor contracts, or oversight policy, at the same speed competitive pressure is forcing on the AI companies, without building in the same kind of clear reversal terms Anthropic at least built into its own contracts.
If the biggest AI companies in the world are pricing the cost of being wrong into every commitment they sign, is your organization pricing that same cost into yours, or just hoping you will not need to reverse course?
**Build the harness. Price the outcomes. Redesign the org.**