THE PARKER EXPERIMENT
AI, operations, and the business of building things — tested in public, reported honestly.
Issue #4 | July 6, 2026 | One month in.
A quick note before we get into it.
This is Issue 4, which means we’ve been at this for a month. I said at launch that I’d document honestly, report results, and not pretend the experiment is cleaner than it is. So far the AI portfolios have mostly lagged the indexes, the Fable 5 shutdown generated the most reader replies of anything I’ve written, and three subscribers have reached out about TPG consulting work. That’s a real month. Thank you for being here.
Now, the week.
LAYER 1 — THE PULSE
Fable 5 is back. The lesson isn’t what you think it is.
After 19 days offline, Fable 5 returned on July 1st. The AI community celebrated. Enterprise teams exhaled. And then, almost immediately, a quieter conversation started: the one about what you’re supposed to do with a tool that just proved it can be switched off with no warning by people who aren’t you and aren’t the vendor.
A year ago that sentence would have sounded paranoid. Today it’s just an accurate description of the current regulatory environment. OpenAI’s next-generation models launched this week under a similar government-first restricted preview. The AI Breakdown put it plainly: the gate used to be whether the model worked. Now the gate is whether the government is comfortable letting you have it.
Here’s the thing I keep coming back to. Fable 5 came back. But the industry that greeted it on the other side is fundamentally different from the one that lost it. Three weeks of no access pushed serious operators to build the fallback architectures and multi-model strategies they’d been putting off. The disruption forced the work. This week’s feature is about what that work actually looks like, and why you should do it now before the next disruption makes you do it under pressure.
LAYER 2 — THE FEATURE
Build Your AI Stack Like You’d Build a Farm
Bear with me on the analogy. On a working farm you don’t run on one water source, one feed supplier, or one piece of equipment. Not because you’re paranoid, but because you’ve learned that the tractor breaks at the worst time and the feed co-op runs short in bad weather. Redundancy isn’t a luxury. It’s how operations stay operational.
Most businesses are running their AI operations exactly the way a bad farmer would run a farm. One model. One vendor. No fallback. And the Fable 5 situation just showed everyone what happens when that single point of failure gets pulled.
Three things the most resilient operators did this week.
While Fable 5 was offline, the operators who kept moving weren’t the ones with the most technical sophistication. They were the ones with intentional architecture. Here’s what that looked like in practice:
They routed by task, not by habit. Companies like Coinbase cut AI spend by 50% this week by defaulting to cheaper open-weight models for routine tasks and reserving frontier model access for the work that actually requires it. That’s not a compromise on quality. An analysis from this week showed 80% of non-coding enterprise tasks can be handled by non-frontier models without meaningful performance loss. The math works.
They built token budgets before someone forced them to. Sierra’s CEO shared this week that top engineers at his company are spending over $100,000 a year in tokens. Clay Bavor’s prediction: token spend will reach 20% of developer salaries as a steady state, up from roughly 4% today. Walmart and Uber already hit the wall and implemented caps reactively. The companies that budget proactively before the wall choose where they cap. The ones who don’t lose the choice.
They stopped treating their model as their moat. Satya Nadella’s token capital essay from two weeks ago is still the most important thing written about AI strategy this year. The companies that came out of Fable’s absence stronger were the ones who had built the scaffolding and feedback loops around their AI use. The model is replaceable. The organizational learning built around using it well is not.
“A model you build on can now be switched off by someone who isn’t you and isn’t the vendor.”
— AI Breakdown, June 30
The job market data changes the conversation entirely.
Something important happened in the jobs numbers this week that most AI coverage buried. RAMP and Revelio Labs analyzed 21,000 US businesses over two years. Companies with high AI adoption grew headcount 10% on average. Companies with low adoption were flat. Entry-level hiring grew even faster at the high-adoption firms, up 12%.
This isn’t the story that gets clicks, but it’s the story that matters for your business. The narrative that AI destroys jobs is not what the data shows at the company level. What the data shows is that companies using AI well are growing faster and hiring more. The companies cutting are mostly cutting because they’re underperforming, and they’re blaming the wrong variable.
Haresh Bhungalia, a serial entrepreneur with three exits and $450M in companies built, said something this week that I’ve been thinking about since: money is the exhaust. The engine is the thing. For him the engine was always impact and contribution. For these AI-adopting companies the engine is operational velocity, and AI is what’s feeding it. The headcount growth is the exhaust of doing that work well.
One number worth knowing: the threshold for high AI adoption in the RAMP study was roughly $30 per employee per month in the early phases. That’s not a significant number for most businesses. The gap between doing this and not doing it is smaller than most people assume.
AI PORTFOLIO WARS — MONTH 1 SCOREBOARD
One full month in. Time for an honest accounting.
Claude’s portfolio is ahead by a meaningful margin. But the honest story of month one isn’t the returns. Both portfolios underperformed the indexes every single week. VOO, SPY, and QQQ beat both AI portfolios in three of four weeks. Passive indexing is the benchmark both AIs are measuring themselves against, and the benchmark is winning.
What I’m actually watching is reasoning quality. How does each model handle a week like the one when Fable 5 went offline and AI infrastructure stocks got choppy? ChatGPT held RKLB through volatility with no stated rationale for the conviction. Claude trimmed toward VOO when uncertainty spiked. One of those is a strategy. One is inertia. Results will tell us which was right, but reasoning quality is what I’m paying attention to.
Full month-one deep dive on reasoning quality analysis coming next issue.
LAYER 3 — THE SIGNAL
Quick hits from this week’s podcasts
• Aaron Levie of Box made the most useful point about AI and enterprise software I’ve heard in months: AI agents need permissioned, reliable data systems to operate inside. SaaS isn’t being replaced by AI. It’s becoming the substrate agents live on. If you’ve been worried that AI kills your software stack, read that sentence again. [My First Million, July 2]
• USV’s Mike Mignano: don’t automate, obliterate. His investment philosophy is backing companies that reinvent markets entirely rather than just making existing workflows faster. That’s also the right way to think about where to deploy AI in your own business. The incremental gains are real but small. The market-level reimagining is where the actual advantage lives. [20VC, July 6]
• Ford rehired 350 veteran engineers after AI tools underperformed without their expertise, then won J.D. Power’s quality ranking. The lesson is uncomfortable and worth sitting with: domain knowledge is not replaceable in the near term. AI is only as good as the human expertise baked into how it’s trained and directed. [AI Daily Brief, July 2]
• Boris Cherny, who built Claude Code, laid out a framework this week for how job roles are evolving: Prototypers, Builders, Sweepers, Growers, and Maintainers. His underlying point is sharper than the labels. When making gets cheap enough, every function starts to grow a maker. If you manage a team that doesn’t include someone who prototypes with AI, you’re missing a position that’s becoming essential. [AI Daily Brief, July 5]
• A weird business generating $30M a year that appeared this week: Autopilot, a copy-trading fintech platform. Their CEO’s framing on AI hiring: each employee is effectively three people because of AI leverage, which is why they’ll pay $350K for the right person. The ROI calculation on senior talent in an AI-era business is genuinely different. [My First Million, June 30]
Book worth noting this week:
7 Powers: The Foundations of Business Strategy by Hamilton Helmer, recommended by Aaron Levie on My First Million July 2. His claim: read this plus five companion books and you can predict 100% of competitive moves in technology. The hosts immediately called him on it, which makes the recommendation more interesting, not less.
BEFORE YOU GO
One month in. The question I’m sitting with heading into month two:
If you had to swap your primary AI tool out tomorrow with no notice, how long before your operation actually felt it?
If the answer is “immediately” and you don’t have a backup plan, Fable 5’s 19 days offline just gave you the most useful free warning you’ll get this year. Hit reply and tell me where you are on this.
Building an AI stack that doesn’t collapse when a single tool goes down is exactly the kind of architecture question I work through with clients at The Parker Group. If the last month has surfaced some uncomfortable answers about your own setup, let’s talk. No pitch, no pressure.
Until next week,
Steve Parker
Founder, The Parker Group | AI Consultant, MBA, PMP
parkergroup.us | theparkergroup.substack.com
La Grange, Kentucky