Friday, July 10, 2026
FRIDAY – AI FOR THE C SUITE®
Read time: 11-12 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
On this week’s episode of AI for the C Suite®, I dismantle two stories leaders keep telling themselves: that their people are behind on AI because they haven’t been exposed to it, and that recognizing the shift is the same as being ready for it… all through the lens of this summer’s sleeper hit about talking-sheep detectives. I’ll walk through why your competitors will never hand you a defining moment, and why the gap between ‘I finally see it’ and ‘I know what to do about it’ is one of our current moments’ strategic imperatives. 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: Who’s Your Mattingly?
On July 1, Fable 5 came back. (See Radar Hits below for more on that).
The Commerce Department lifted the export controls on June 30, and by Wednesday morning the most capable model available was back at work again like nothing happened. (What, no? I’ve been here the whole time boss… must’ve just missed seeing each other). In brief, Anthropic shipped a new safety classifier that blocks the jailbreak technique behind the whole mess in over 99% of attempts, stood up a bounty program for researchers who find new ones, and promised the government early access to future models. Nineteen days of drama, resolved with a letter.
Three weeks ago I wrote about the shutdown and stapled my usual caveat to the front: the facts could shift under your feet before you finished reading. They shifted, and in the happy direction. Hurray! The models are back, the classifier is live, and every technology leader who spent late June sweating a possible dependency they never knew they had is now being invited by events to close the ticket and move on.
Please don’t. You just got a free fire drill. The building didn’t burn down, nobody got hurt, and the alarm turned out to be resolvable with two weeks of negotiation in Washington. A happy ending is the most dangerous possible outcome of a fire drill, because it teaches you the alarm always gets resolved. The thing worth carrying into the rest of the summer is what the drill exposed… I thought I knew what that was. Then, as I sat in a room this past Wednesday, I realized I’d missed a big piece of the story.
THE ITEM MY CHECKLIST MISSED.
My June piece handed you a fallback checklist. A second provider. A portable prompt layer. An open-weight model you’ve tested rather than bookmarked. I stand by every item on it. But notice what those three have in common: each one is something you can buy, rent, or download. They’re all architecture. Swipe a card, sign a contract, provision a server, and the hedge exists.
This past Wednesday I convened my AI-CTO peer group, a room of senior technology leaders from mid-market companies across manufacturing, distribution, construction, and higher education. We spent two hours on a variety of AI-related issues: the shutdown, the moves everyone is making, token economics, the usual governance headaches. And somewhere in hour two, the conversation found the item my checklist missed. The one fallback that has no vendor, no license tier, and no download link.
The people who still know how the work works.
DO YOU LEARN THE MATH WHEN YOU HAVE THE CALCULATOR?
One member described a project his team is weeks from shipping: automating a complex month-end accounting report, the kind that feeds journal entries straight into the ERP once it’s done. It’s a good project. It will save hundreds of hours a year. And he’s hesitating at the release point, because of something he noticed about the person the automation replaces.
The controller who builds that report by hand today catches errors while he’s building it. A line looks out of whack on step four, so he stops, digs, and fixes it before anyone downstream ever sees the mistake. Nobody trained him to do that. The instinct grew out of years of assembling the thing manually, feeling where the numbers come from and how they move from source to summary. Automate the assembly and the report gets faster, cleaner, and reviewed by people who never earned that instinct and never will.
The room contemplated that for a moment, and then someone compressed the entire problem into nine words: do you learn the math when you have the calculator?
Then a second member connected it to a frame I’ve used with this group before, the hourglass: organizations thinning out the middle layers where deep operational knowledge lives, keeping the executives who set direction and the tools that execute it, with less and less in between. Put the two together and you get the question I’ve been pondering. If a model you depend on goes dark again, and next time it stays dark, who in your company still knows the longhand method? Not who knew it in 2024. Who knows it right now, and who will still know it in 2028, after four years of pressing the button?
MISSION CONTROL RAN ON LONGHAND.
If you’ve seen Apollo 13, you know how that story gets saved. An oxygen tank explodes 200,000 miles from Earth, and nearly every system the crew depends on becomes dead weight or a threat. What brings three astronauts home is a building full of engineers who understand the spacecraft down to the amp, because they spent a decade flying its procedures by hand. Ken Mattingly climbs into a simulator and works out a power-up sequence that had never been attempted, under a power budget everyone believed impossible, using knowledge no checklist contained.
NASA never planned for that failure. It survived the failure because the humans in the loop still knew the math without the calculator. That’s the fallback with no vendor. And it only existed because the organization was full of people who had done the work the long way, recently enough to remember how.
Your company has its own version of Mattingly. The controller who feels a bad number. The ops manager who can schedule the line from a whiteboard when the system chokes. The estimator who can price a job from a walkthrough. Every automation you deploy makes those people less necessary on a Tuesday, and the market for what they know clears slowly enough that you won’t notice the shortage until the day you need it.
THIS DOESN’T WAIT FOR ANOTHER SHUTDOWN.
Here I need to close a loophole in my own argument, because the shutdown framing makes this sound like a tail risk. It’s worse than a tail risk. The chronic version is already running inside your company today.
An automated report that’s right 99 percent of the time, reviewed by people who trust it because it’s right 99 percent of the time, is a system with no immune response. The one percent doesn’t announce itself. It ships, it posts, it compounds quietly in your financials or your inventory counts until something visibly breaks, and by then the person who would have caught it on step four retired, or left, or became the position you decided not to backfill because the tool covers it now. June’s shutdown was the acute presentation. The knowledge drain is the chronic condition, and it progresses whether or not Washington ever touches a model again.
WHAT I’D PUT IN FRONT OF YOUR LEADERSHIP TEAM.
For your consideration, I offer four suggested moves, and none of them require slowing your automation roadmap.
First, take inventory. For every mission-critical workflow where AI now does the assembly, name the human being who could still rebuild it longhand. A person, with a name. If all you can write down is a job title, dig deeper. If you can’t write down anything, you’ve found an exposure that no second provider will hedge.
Second, consider making longhand custody a named responsibility instead of an accident of tenure. Right now, your fallback knowledge lives wherever your most tenured people happen to sit, and it walks out the door on their schedule, not yours. Decide deliberately which workflows warrant a designated custodian, put it in the role, and pay for it like the insurance it is.
Third, run the drill. Once or twice a year, pick a critical automated workflow and have a human rebuild it by hand, start to finish, then reconcile the two outputs. You already test your backups and your disaster-recovery plan without waiting for the disaster. Test your people’s ability to be the backup. The reconciliation will occasionally catch a drift in the automation itself, which means the drill pays for itself even in the years nothing goes wrong.
Fourth, document the longhand process before that knowledge departs. At worst it becomes outdated if nobody keeps it current. At best, it becomes a baseline for reconstruction. Shameless plug: we’ve developed an excellent process mapping tool that we use with clients. Ask us about it or choose your own adventure. Regardless of the path, get this documented.
Fable came back in nineteen days with a letter from the Commerce Department. Your controller’s instinct doesn’t work that way. Once the people who earned it are gone, no letter, no budget line, and no vendor restores it, because it was never for sale in the first place. Muscle memory fades quietly, hence the need for documentation and process mapping. The models will keep compounding and the permission slips will keep getting stranger, and both of those sit outside your control. What sits inside your control is whether anyone in your building will still know the longhand method on the day the calculator fails, or lies.
Who’s your Mattingly, and which workflow would you least want to rebuild by hand tomorrow? Drop me a line at chad@chadharvey.com and tell me. I read every one.
3. Research Roundup: What the Data Tells Us
AI FORECASTING: THE MODEL YOU BUY ISN’T THE DECISION THAT MATTERS
Before you approve budget for the top-ranked AI model, read this. A new pilot study on forecasting found the model’s benchmark score had almost nothing to do with who got results. What mattered was how people used it. And one common habit made results worse than using no AI at all.
The numbers that matter: Researchers paired people with four different AI models and graded every forecast against live prediction-market outcomes, so scoring stayed objective. Three distinct usage patterns emerged. Only one, the people who reasoned back and forth with it, beat every individual system tested and matched the market itself. The group that used AI mainly to confirm a hunch they already held finished dead last, below forecasters working with no AI. A capable model in the loop actively hurt them.
What this means for your Monday morning: Raw intelligence predicted who forecast well alone. Once AI entered, that link vanished. What mattered instead was perspective-taking, curiosity, and the willingness to treat your first answer as provisional. You can hire and coach for those traits.
The catch: This is a pilot with a small sample, and a larger pre-registered study is underway. Treat it as a strong signal, not settled proof.
Action item: Ask your team this week: “When the AI agrees with you, do you stop or dig deeper?” If they stop, you have found your risk.
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
Anthropic redeploys Fable 5 after government suspension. The feds export-controlled the top frontier model June 12 and it vanished from every customer’s stack for almost three weeks. It’s back with tighter safeguards, but the lesson is continuity: if one model going dark would stall your operations, build your fallback now.
AI automation of freelance work jumped from 2.5% to 15.8%. The Center for AI Safety measured frontier models against real paid freelance projects, human-judged. The trajectory matters more than the number: the outsourced digital work you buy (design, data analysis, video) is repricing fast. Map that vendor spend now.
Japan closes the book on AI-as-inventor. Eight jurisdictions now agree: AI can’t hold a patent, but AI-assisted inventions stay protectable when a human documents conception. That makes R&D documentation the whole ballgame for your IP. If your engineers use AI tools, have patent counsel set the recordkeeping standard this quarter.
Perplexity launches Computer for Counsel. An AI legal workspace for contract review, NDA triage, and regulatory monitoring. Legal AI just became a procurement line item, not a novelty. Ask your GC which routine work a tool like this absorbs from your outside counsel bill.
5. Elevate Your Leadership with AI for the C Suite®
The four moves in this issue aren’t a roadmap you hand down; they’re a decision your leadership team makes together about which knowledge you refuse to lose. That conversation goes better with a facilitator who’s run it before. I have room for a few Q4 2026 strategy engagements built around exactly that question. If you’d want a second voice in the room when you put this in front of your team, reach out. Then send AI for the C Suite® to the peer most likely to nod at “do you learn the math when you have the calculator?”
Stay safe. Stay healthy. Be strong. Lead well.
Chad