Friday, June 19, 2026
FRIDAY – AI FOR THE C SUITE®
Read time: 10-11 min · Read online
Hi, it’s Chad. Every Friday, I serve as your AI guide to help you navigate a rapidly evolving landscape, discern signals from noise and transform cutting-edge insights into practical leadership wisdom. Here’s what you need to know:
1. Sound Waves: Podcast Highlights
This week on AI for the C Suite®, I skipped the usual one-on-one and convened the “AI Super Friends.” They’re five practitioners who build, deploy, and govern this technology for a living. No slides, no script: just a candid riff on whether our existing playbooks stretch to cover agentic AI, why “well-governed agents” beat the autonomous hype, and what the 95% AI-failure rate really tells us. If you want to get your bearings on agentic AI in one episode, make it this one. Hit one of the below links to check it out:
Apple · Spotify · iHeart · Amazon · YouTube
Subscribe for free today on your listening platform of choice to ensure you never miss a beat.
2. Algorithmic Musings: When the Frontier Requires a Permission Slip
For two years I treated one question about AI as settled. Can I get access to the best model? Sure. Swipe a card, accept the terms, you’re in. The interesting questions all lived a level up: what can it do, how fast is it improving, what should you build on it. Access was plumbing. Capability was the show.
Then last Friday the plumbing became the show.
On June 12, the U.S. government issued an export control directive that resulted in Anthropic cutting off access to its two most advanced models, Fable 5 and Mythos 5, for any foreign national, including the company’s own foreign-born employees. Rather than try to gate those users one at a time, Anthropic disabled both models for everyone. Its other models, including the one many of you use day to day, kept running.
A caveat before we go further, because this is the kind of forward-looking piece where the facts can shift under your feet. Anthropic says it believes the order is a misunderstanding and is working to restore access. As I write this, an Anthropic executive says access could be back within days. Cutting the other way, the one outside expert known to have read the underlying technical paper, Katie Moussouris, argues the flaw can’t be cleanly fixed without weakening the model… so a fast resolution may be harder than the optimists suggest.
By the time you read this, the specific dispute may be half-resolved. Treat what follows as a snapshot of a fast-moving story, with one structural lesson underneath it that will outlast the news cycle.
And, by the way, this entire situation is a remarkable sequence. A government reached for an export-control lever, and a company’s most capable products went dark by Friday night. To see why it matters to you, and why it matters far more than one company’s bad week, you need two definitions. While they sound like inside-baseball jargon, they’re actually the whole ballgame.
What “open weight” means
An open-weight model is one whose trained weights, the billions of learned parameters that make the thing work, are published for anyone to download. You can run it on your own hardware. You can fine-tune it on your own data. You can deploy it without paying a toll per query to a vendor. Llama, DeepSeek, Qwen, Kimi, and OpenAI’s gpt-oss all live here.
One clarification, because the terms get muddied. Open weight is not the same as open source. With most of these models you get the weights but not the full training data or the complete recipe used to build them. For our purposes the distinction that counts is simpler. Once you’ve downloaded the weights, the model lives on your machine. It’s been copied to thousands of other machines too. Nobody can reach in and switch it off. Not the vendor. Not a government. It’s loose.
What we mean by “frontier”
A frontier model is one of the small handful sitting at the leading edge of what’s currently possible: the top performers on reasoning, coding, and the agentic tasks businesses are racing to automate. For most of the last few years the frontier has been closed. Claude, GPT, and Gemini set the pace, and you reached them only through a vendor’s door.
That’s been changing. Open-weight challengers have closed much of the gap. DeepSeek’s latest, Qwen, Kimi, and Meta’s Llama 5 now trade real blows with the closed leaders on coding and reasoning, even if the very hardest problems still tilt toward the closed frontier. So the frontier is no longer one kind of thing. Part of it you rent through a door someone else controls. Part of it you can download and own outright. Hold onto that split, because Friday’s news drove a wedge straight through it.
The asymmetry the directive exposed
When the government wanted Fable 5 and Mythos 5 out of foreign hands, it could make that happen. Closed, hosted models run through a chokepoint. Every request passes through identity, billing, region, and authorization, most of it riding on a cloud provider like AWS. Pull the right thread and the model goes dark for everyone, which is exactly what happened. Anthropic couldn’t comply selectively, so it shut the whole thing off.
Now ask the question the order can’t answer. What would the same directive do to an open-weight model? Nothing. There’s no door to lock, because the model was handed out, copied, and forked long ago. Anthropic made a quieter version of this point in its own defense, noting that the capability the government feared could be coaxed out of other public models too, including OpenAI’s GPT-5.5, which faces no such restriction.
If you grew up on Jurassic Park, you already know the shape of this. The park could control the dinosaurs it kept behind the fences. It had no answer for the ones already breeding out in the jungle. Closed models are the ones behind the fence. Open weights are loose in the trees. A government can recall what it can reach, and it cannot reach what’s already in the wild. That asymmetry is the most important thing the directive revealed, and it’s getting lost under the louder takes about munitions and nationalization.
What this changes for you
Step out of the geopolitics for a moment, because the lesson for a middle-market operator sits much closer to home.
Maybe you standardized on a frontier closed model because it was the best tool you could buy. (Totally reasonable, btw). You wired it into a customer-facing product, or an internal workflow your team now leans on every day. Also reasonable. What Friday added to your risk register is a category you probably never underwrote: a regulatory shutoff. An outage passes. A price hike you can negotiate. Yet a government directive can revoke your access overnight, possibly for a reason you’re never even given. Case in point: the letter to Anthropic didn’t even spell out the concern.
Long-time readers will recognize the family resemblance. This is the geopolitical cousin of The SaaS Overhang: the slow accumulation of dependence on tools you rely on but don’t control. The overhang I wrote about back then was commercial, a stack of subscriptions quietly compounding. This version adds a state actor of dubious expertise with a hand on the switch.
It also reframes the bet so many of you are making on what comes next. Capability is going to keep compounding. That part of the AI story is about as safe a wager as exists right now. But access to the top of that curve has quietly decoupled from capability and turned into a question of permission. When you standardize on a single frontier closed model, you’re stacking two bets, not one. The first, that the model keeps getting better, looks solid. The second, that you’ll still be allowed to use it, just got a lot less certain.
What to do with it
I won’t tell you to rip out your closed models and self-host everything. That would be its own kind of foolish. The closed frontier is still the strongest option for plenty of work, and running open weights well carries real costs: the infrastructure, the talent to manage it, and the fact that the hardest tasks still favor the closed leaders.
The move is quieter and more disciplined. Know which bet you’re making on each workload. For the ones where continuity is non-negotiable, build a fallback before you need it: a second provider, a portable prompt layer, or an open-weight option you’ve tested rather than bookmarked. Treat your model choice as an architecture decision with a political-risk line item, not just a capability-and-price comparison. Regular readers know the frame I’d hang this on: Cut, Coast, or Redeploy. Some dependencies you hold, some you hedge, and some you redeploy around before the choice gets made for you.
Which brings me back to where I started. For two years I filed access under “settled and boring” and spent my attention on capability, because capability was the part that felt alive. Friday corrected me. The boring part turned out to be load-bearing. Turns out the fence was the thing worth watching all along.
My takeaway? The exponential curve everyone’s betting on isn’t going anywhere. Yet. But it now travels with a permission slip which can be revoked by people who don’t have to explain themselves or their reasoning. The leaders who come out of this ahead will be the ones who already know what they’ll do on the morning their most powerful model won’t open for them. If you’re trying to figure out where your own dependencies sit, and what a sane fallback looks like, write me at chad@chadharvey.com. I’d like to hear which bets you’re stacking.
3. Research Roundup: What the Data Tells Us
AI CODING AGENTS: THE MODEL IS THE SMALL PART YOU’RE OVERPAYING TO COMPARE
Before you greenlight an AI coding tool, a new independent study should change the questions you ask. Researchers took apart a production coding agent and found the part everyone shops for, the AI itself, is a sliver of what makes it work or fail.
The numbers that matter: The model’s decision logic accounts for under two percent of the codebase. The rest is operational plumbing: permission gates, context management, error recovery. People approve the large majority of permission prompts, and for heavy users, automatic approvals climbed from one in five to more than two in five over time.
What this means for your Monday morning: Benchmark scores tell you almost nothing about whether a tool is safe to deploy. As models converge on raw skill, the real differentiator is the harness: how permissions are enforced, how the system recovers, how sessions get audited. That is where reliability is won or lost.
The catch: This is careful work, but it reads a publicly extracted source snapshot, not the live product. The authors are candid: reverse-engineered code cannot confirm what runs in production. Treat it as illustrative, not a current spec sheet.
Action item: Ask any agent vendor what happens after the model decides: how it gates risky actions, recovers from errors, and logs every step for audit. If the answer is all benchmarks and no harness, keep looking.
Read our full analysis of this and all other analyzed research papers at AI for the C Suite®.
4. Radar Hits: What’s Worth Your Attention
The Case Against Building Your Own Agent Platform. If your team is pitching a homegrown AI agent platform, the market already voted. Enterprise build-versus-buy flipped in one year, from 47% building internally in 2024 to 24% by late 2025, and Gartner expects 40% of agentic projects dead by 2027. The platform layer is now the expensive path. Build the agents your business needs; buy the memory, orchestration, and governance underneath.
Generative Engine Optimization in 2026. Adobe just paid $1.9 billion for Semrush, putting a price tag on getting cited inside AI answers instead of buried in blue links. With ChatGPT past 900 million weekly users, the question a customer asks an AI is the new front door to your business. Ask marketing this week: are we tracking whether ChatGPT and Google’s AI Overviews mention us, or only our competitors?
Bots now outnumber humans online, and AI traffic is growing 6.5x faster than human traffic. Cloudflare clocked the crossover in June; most web requests are now machines, mostly AI agents shopping and researching on someone’s behalf. Those agents increasingly decide whether a customer ever sees you. Stop filing bot traffic under security alone. Decide which AI systems you want reaching your site and which to block, because that call now shapes who finds you.
5. Elevate Your Leadership with AI for the C Suite®
Here’s a 30-minute exercise for your next leadership meeting: list every workload running on a single frontier model, then ask what you’d do the morning it won’t open. If the room goes quiet, that’s the work. I run that pressure-test with middle-market teams, and have availability next quarter to do it with yours. Ping me at chad@chadharvey.com. If this issue earned its spot in your inbox, forward AI for the C Suite® to a peer who’s stacking the same bet.
Stay safe. Stay healthy. Be strong. Lead well.
Chad