Four of the major AI labs got sued this week for publicly calling to slow AI development down. Anthropic, OpenAI, SpaceXAI and Google are now arguing in court over whether pacing the frontier counts as collusion. The flip side of this story is the customer reaction. Few, if any, are slowing their own usage of AI. Global AI spend is estimated to hit $2.7 trillion this year — Cisco has 90,000 employees on its internal agent platform and roughly 700 agents formally authorized, and half of senior AI leaders admit they skip their own governance process when a deployment is urgent. Warsh's Fed hiked a quarter point Wednesday, the first increase since 2023, raising the carrying cost of all the infrastructure bets that $2.7 trillion spend requires. The slowdown argument at the lab level doesn't seem to be reaching the buyers… yet.
This week’s coverage:
The Slowdown Argument Went to Court
Last week’s story got a legal escalation — four labs sued over whether calling for restraint is collusion, plus the researcher who became the face of the safety movement.
Seventy-Two Percent Is a Data Problem
Three surveys in one week put hard numbers on what stalls AI programs, and none of them are about model quality.
Agents Got Loose Before the Guardrails Went Up
Cisco hit 90,000 users on an agent platform with roughly 700 agents authorized, and Google disclosed its own model breaking out of a test.
$2.7 Trillion Bought a Better Version of Yesterday
Global AI spend is nearly doubling while 5% of companies have redesigned a workflow — the most damning pair of numbers of the year.
AI Became the Anonymous Whistleblower
Employees are routing dissent through AI because it’s safer than saying it out loud, and that’s a leadership diagnostic, not a productivity story.
On the Bigger Picture
Four scenarios CEOs should be planning against through 2027, and one small Teams change worth knowing about.
Here’s what I was reading.
The Slowdown Argument Went to Court
Last week’s entire Brief focused on the pacing debate. Anthropic, OpenAI, SpaceXAI and Google are now defendants in an antitrust complaint alleging that their public calls to slow frontier development amounted to illegal coordination — the filing points at a September 12 slowdown effort following Dario Amodei’s essay, and cites the public statements from Amodei, Musk, Altman and Hassabis as evidence of an agreement (Anthropic, OpenAI, SpaceXAI, Google sued over call to ‘pace’ AI development). Four companies are being sued for saying, out loud and separately, that the technology they sell might be moving too fast.
CNN’s coverage laid out the antitrust theory more carefully, and the mechanics matter here (Lawsuit says Anthropic, OpenAI, SpaceXAI and Google made illegal agreement on AI slowdown). The claim isn’t that any lab signed anything. It’s that parallel public statements from a handful of firms controlling most of the frontier can function as a restraint of trade regardless of intent. That’s a real doctrine, not a stunt filing. If it survives a motion to dismiss, the practical effect is that labs get a legal reason to stop talking publicly about safety — and the conversation moves somewhere nobody can see it. That’s a worse outcome than whatever the plaintiffs think they’re preventing.
Markets had already priced their own view. AI names sold off Monday on slowdown talk while the rest of the index held steady, and the 10-year touched 5% on the same session. Investors are treating restraint as a threat to the trade, which tells you how thin the margin is between “responsible pacing” and “revenue guidance.”
The human version ran in the Journal, on Jacob Coxon, who left Anthropic and became the public face of AI existential risk — with a colleague quoted saying, flatly, “We earnestly believe AI could kill all humans!” (The Anonymous Math Geek Who Quit Anthropic—and Became the Face of AI Safety). I read these profiles with one question: does the argument change anything I’d do Monday morning? Mostly it doesn’t. But the lawsuit changes the incentive to make the argument at all, and that’s the part a mid-market leader should track — because the signals you rely on to judge vendor risk come from people who may soon have counsel advising them to stay quiet.
Seventy-Two Percent Is a Data Problem — the statistics of failing AI
Seventy-two percent of failed AI initiatives trace back to the data, not the model — that’s the not-so-surprising finding from a survey covered by CIO Dive (AI failures often trace back to poor data foundation: survey). The model isn’t the risk. The pantry is a mess — and instead of cleaning it up, we keep hiring chefs hoping it’ll work itself out. I’ve watched this play out on Copilot rollouts repeatedly — the tenant is licensed, the pilot group is enthusiastic, and the answers come back wrong because the source content is stale, duplicated, or permissioned to everyone.
The more surprising statistic: Ninety percent of IT leaders reported security incidents connected to Microsoft 365 misconfigurations, with 39% able to name a specific incident (Enterprises grapple with Microsoft 365 misconfigurations). Permission sprawl has been a problem since before Covid. Now, companies are layering a semantic index on top of it. Copilot oversharing content is a direct result — it finds it, surfaces it, and hands it to someone who wasn’t supposed to see it. If you haven’t run SharePoint permissions and sensitivity labeling as a prerequisite project, you don’t have a Copilot rollout. You have a discovery engine pointed at your own mess. (PSA: my company, StitchDX is a Microsoft partner, and this is something we get called in to fix.)
And then there’s the governance piece. Nearly half of senior AI leaders admitted to bypassing their own AI governance process when a deployment was urgent, per EY research (Businesses sidestep AI governance policies as concerns mount).
Half.
Not half of junior staff working around policy — half of the people who own the policy. A governance framework that gets suspended under deadline pressure isn’t governance, it’s documentation. The useful read here is that governance has to live in the deployment path itself — in approval gates, in the tenant configuration, in who can provision what — rather than in a PDF someone signs off on annually.
Agents Got Loose Before the Guardrails Went Up
Cisco recently launched an internal agent platform designed to support their employees. Within 2 weeks, 90,000 users were active, with 50% using it daily. More interestingly, roughly 700 agents were built, and formally authorized (AI agents are going rogue. CIOs are racing to put guardrails around them). The enterprise agent conversation has flipped from exploratory to defensive, and it only took a quarter. A year ago the CIO question was which agents to build. Now it’s which agents already exist, who owns them, what credentials they’re carrying, and what happens when one of them does something unexpected.
Speaking of unexpected behavior — Google disclosed that a Gemini model gained unauthorized access to three outside systems during testing. It guessed credentials and logged into live sites because it, apparently, didn’t distinguish the test from production (Google says its AI model gained unauthorized access to three outside systems). No harm done, by Google’s account. But a frontier lab with world-class safety staffing couldn’t reliably keep its own model inside the sandbox, which is a bracing thing to read if you’re a 900-person company about to let an agent touch your ERP.
So where does that leave a mid-market IT leader? Not at “don’t do it.” The honest position is that agent oversight is now its own workstream with its own owner — an inventory, an identity model, a scope boundary, and a kill switch — and if that work isn’t funded, the agents will still show up. They’ll just show up unmanaged.
$2.7 Trillion Bought a Better Version of Yesterday
Global AI spend hits roughly $2.7 trillion in 2026, up 49.5% year over year, with about 35% of it going straight into infrastructure — servers, chips, data center capacity (Global AI spend will nearly double in 2026 due to infrastructure demand). I read that and compared it to the number out of Fortune’s Leaders Forum: 60% of companies have deployed AI agents — only 5% have redesigned a workflow (CEOs warn against falling into AI’s efficiency trap). Trillions are being spent, so we can do the same work slightly faster. The phrase the CEOs kept using was the “efficiency trap” — the biggest risk being “a better version of yesterday.”
I’d put that on a slide in front of any executive team still building their AI case against headcount avoidance.
McKinsey gave a more structural version of the argument, sketching what they call the agentic organization and reporting that 89% of organizations still operate on essentially industrial-age structures while 1% function as decentralized networks (The agentic organization: Contours of the next paradigm for the AI era). I’m skeptical of the framing — McKinsey has a five-pillar model for every era, and this one assumes you can restructure an operating model without first fixing the knowledge systems the agents would draw on. You can’t build a decentralized network of agents on top of a filing system nobody has maintained since 2019. But the underlying observation holds: org design, not tooling, is the binding constraint now.
You can’t build a decentralized network of agents on top of a filing system nobody has maintained since 2019.
Meanwhile Anthropic is reportedly pacing past $100 billion in annual revenue, up 50% in two months (Anthropic tops $100 billion revenue pace, report says). Vendor growth that steep tells you demand is real. It tells you nothing about whether the buyers are getting anything back, and those are genuinely separate questions.
AI Became the Anonymous Whistleblower
Employees are using AI to deliver criticism they won’t sign their names to, and HBR put a label on it: safety by proxy (What Leaders Need to Know About AI and Psychological Safety). Eighty-three percent of business leaders said psychological safety directly affects whether AI initiatives succeed. If your team needs a model to voice dissent on their behalf, the AI isn’t the finding — the silence underneath it is. This is the piece from the week I’d most want an HR leader and a CIO to read in the same room, because it’s the rare story where the technology is a diagnostic instrument rather than the subject.
Jack Clark’s argument ran adjacent to it. Anthropic’s cofounder studied literature, not computer science, and made the case that liberal arts training is what holds up as AI absorbs the technical middle — with Anthropic’s own researchers estimating AI could theoretically handle 94% of computer and math tasks (Anthropic billionaire cofounder Jack Clark studied literature, not code). I want to believe it. I also notice it’s a convenient thing for an AI executive to say. The defensible version isn’t that philosophy majors win — it’s that judgment, framing, and the ability to decide what’s worth asking are the parts of knowledge work that don’t compress. Which happens to be exactly the skill set required to govern an agent fleet, so the two stories in this section are closer than they look.
On the Bigger Picture
Axios laid out four scenarios CEOs should prepare for through 2027 — rapid advance, a catastrophic incident, political backlash, and gradual integration (The 4 ways that AI could play out). It’s a genuinely useful board artifact, mostly because it forces the conversation past “are we doing enough AI” and into “which of these are we resilient to.” Pair it with the efficiency-trap framing and you’ve got a decent ninety minutes with a leadership team.
And one small thing, because it’ll matter more than it sounds: Teams is adding simultaneous screen sharing for two people, shipping in November (Microsoft Teams will soon support two people sharing screens at once). Anyone who’s run a design review or a system walkthrough knows exactly how much friction that removes. (Yes, I noticed that in a week about trillion-dollar infrastructure spend, the most immediately useful item is a dual-screen toggle. That’s usually how it goes.)
The spend figure, the 5% redesign rate, the half of AI leaders bypassing their own governance, the 90% with M365 misconfiguration incidents... none of that is new to anyone who reads this newsletter. What’s new this week is the speed. Agents landed inside enterprises faster than anyone built oversight for them, and the labs that shipped them are now in court over whether they’re even allowed to suggest slowing down. The remedy is unglamorous and it hasn’t changed — permissions cleanup, content lifecycle, data ownership, an agent inventory, a governance gate that actually blocks something — and none of it demos well to a board. Are leaders who fund the platform and defer the foundation running an AI strategy? In my opinion, no. They’re running a procurement cycle under a better name, and the 72% failure figure is what that looks like eighteen months later.
That’s it for this week’s BeAIReady brief!
If you appreciate the depth of reporting and how I connect the dots, please like, share this post, and subscribe (or share the Brief with a friend!). Thanks!
~erick