Sunday, June 21, 2026 Intelligence for Portfolio Executives closing the AI Wage Gap. ———————————————————————————

Anthropic quietly published an analysis this week of 400,000 Claude Code sessions — real work, real users, across the full range of people who now sit in front of an AI agent and ask it to build something. The point of the study was not which model is best or how many lines of code got merged. It was a question about who succeeds.

The finding is the part you should not skim past. Lawyers and managers completed complex agentic tasks at nearly the same success rate as professional engineers.

Read that again, because it quietly demolishes an assumption almost every organization has been operating on for two years.

The assumption was that the people who would win in an AI-saturated workplace are the people who can code. That technical fluency — the ability to write the script, wire the API, debug the stack — was the durable edge. So companies hired engineers, sent managers to Python bootcamps, and treated "can you build" as a synonym for "can you write software." The whole reskilling industry was built on that premise.

The 400,000 sessions say the premise is wrong, or at least expiring fast. When the agent writes the code, writing code stops being the moat. AlphaSignal, which covered the study this week, put it bluntly: "AI is absorbing specialist knowledge and redistributing it. Your moat is shrinking."

———

I want to be precise about whose moat is shrinking, because that one word is doing a lot of work.

The moat that's shrinking is the syntax moat — the part of expertise that was really just fluency in a tool. The lawyer who couldn't build software last year can now build it, because the agent handles the syntax and she handles everything that matters around it. What didn't shrink — what got more valuable — is the thing she brought that the engineer didn't: she knows which problem is worth solving, what the output is supposed to look like, and the precise moment a confident-sounding answer is actually wrong.

That's not coding skill. That's domain expertise. And it just became the scarce input.

Here's the second data point from this week, and it's not a coincidence that it landed in the same seven days. OpenAI trained GPT-5.5 Instant using 600 doctors as a direct training signal — not as end users, as the people shaping the model's medical reasoning.

[side note: I was in a project to train the same model for strategic HR acumen, so I saw how this is done from the inside - it’s quite impressive :]

The result was a 71% reduction in health-related hallucinations. Sit with the structure of that. The frontier lab did not solve medical accuracy by making the model bigger. They solved it by routing 600 licensed professionals' judgment into the training loop.

Two stories, one shape. In the first, domain experts out-perform coders because the coding got commoditized. In the second, a frontier lab pays domain experts because their judgment is the missing ingredient the raw model can't generate. Both are telling you the same thing from opposite ends: expertise is now the appreciating asset, and the technical skill that used to gate it is depreciating in real time.

———

This is the AI Wage Gap thesis stated almost too cleanly to be comfortable.

For two years I've argued that the premium doesn't go to the person who uses the most tools. It goes to the person who can direct AI with deep professional knowledge — who knows what good looks like in their function and can tell, instantly, when the machine is bluffing. The 400,000-session study is the empirical version of that argument. The people winning are not the most technical people in the room. They're the people with the deepest domain knowledge who also picked up the agent.

The danger in this — and there is one — is that "your domain expertise is the moat" sounds like permission to relax. It is the opposite. The expertise is only a moat if it's attached to an agent. The lawyer in the Anthropic study didn't win because she's a lawyer. She won because she's a lawyer who sat down with Claude Code and shipped. The lawyer who has the same expertise and refuses to touch the tool didn't appear in the success column at all. She appeared in the gap.

That's the whole game now. Expertise without leverage is a stranded asset. Leverage without expertise produces confident garbage at scale. The value sits exactly at the intersection, and the intersection is narrow, and almost nobody in your organization is standing on it yet.

Watch how this rewires a hiring decision. A year ago, if you had budget for one person on a build-heavy team, you hired the strongest engineer you could afford and hoped they'd pick up the domain over time. The 400,000 sessions invert that bet. The person who learns your domain is on a multi-year clock; the person who already owns your domain and learns the agent is on a multi-week one. Anthropic's own framing — domain expertise now beats coding skill — is a procurement instruction disguised as a research finding. The scarce hire is no longer the one who can build. It's the one who knows what's worth building and can tell when the build is wrong, and that person was probably already sitting in your finance or legal or operations function, doing work you'd never have labeled "technical."

———

So what does the Portfolio Executive do with this week's number?

You stop competing on the axis that's collapsing and start compounding on the one that's appreciating. Concretely: the worst use of your next ninety days is learning to code a little better, because the agent is lapping you on syntax faster than you can close the distance. The best use is taking the one judgment in your function that only deep expertise can make well — the comp decision, the diligence call, the policy edge case, the clinical read — and building a workflow where the agent does the assembly and you supply the judgment. Then document, in writing, every place your expertise changed the output. That document is your proof case. It's also, not incidentally, exactly the artifact the labs are now paying domain experts to produce.

And there's a positioning move underneath the build. When the people around you are still anxious about whether AI is coming for their expertise, you reframe the whole question for them: AI isn't coming for your expertise, it's coming for the gap between your expertise and your willingness to wield it. The person who says that out loud in the executive meeting — and then opens a laptop and shows the workflow — is the person who gets handed the AI mandate. The IBM data still holds: most Chief AI Officers are promoted from inside, not hired. They're promoted because, when leadership asked "who actually understands how to point this thing at our real problems," there was exactly one person in the room who had already done it in their own function.

The 400,000 sessions just told you which side of the line the value is on. It's not the side you'd have guessed two years ago. It's not the coders. It's the experts who picked up the tool — and there are still far fewer of them than there should be.

The frontier labs figured this out and started paying for it. The only question is whether you build your proof case before your sector's version of that study gets written with your name in the wrong column.

If you need the room to build it: the July cohort is where it gets built. More below.

———————————————————————————

This week's signal maps directly onto three chapters of Closing the AI Wage Gap:

Chapter 3 — The Multiplier is the chapter that anticipated exactly this. Its argument is that the operators pulling ahead are not the most technical — they're the ones who moved AI from assistant to primary producer while keeping their own judgment in the loop. The 400,000-session finding is the organizational-scale proof of the individual-scale pattern I documented across 2,300+ coaching engagements: expertise plus leverage beats raw technical skill, every time the work involves a decision that matters.

Chapter 9 — The Internal Champion is the 57% chapter — why the path to the CAIO title runs through internal proof cases built in your own function, not external credentials. This week sharpens the chapter's core claim: the champion who gets promoted is the domain expert who shipped, not the engineer who could. When coding stops being the differentiator, the differentiator becomes knowing what to point the agent at — which is precisely the thing a deep functional expert has and an outside hire doesn't.

Chapter 15 — The Recursive Advantage deals with compounding, and it's the uncomfortable one this week. If domain expertise is the appreciating asset and leverage is the multiplier, then the gap between the expert-who-shipped and the expert-who-waited doesn't grow linearly — it compounds. Six months of directing an agent against real problems isn't six months ahead. It's a different curve.

Manuscript in Tier-1 agent querying — Levine, Halpern, Sagalyn pending. If you know an editor or agent at the intersection of work, AI, and organizational economics, reply to this email.

———————————————————————————

The best six from a week that had sixty. Full daily version coming — see below.

GLM-5.2 ships open-weight at roughly a sixth of GPT-5.5's price. Z.ai released GLM-5.2 under an MIT license with a 1M-token context window, two reasoning modes, and an 81.0 on Terminal-Bench 2.1 — up from 62.0 — at about $1.40 per million tokens. Frontier-class capability you can run inside your own walls just got cheap. For any org that's been hiding behind "we can't touch real AI without sending our data to a vendor," that excuse has a shelf life measured in weeks now. Z.ai

Cursor Origin clocked 22.6 commits per second in a single repo. Cursor acquired Graphite and launched Origin, a git-hosting platform built for agents rather than humans — because GitHub's assumptions break when the committer is a swarm, not a person. The deeper signal: every piece of developer infrastructure built for human pace is being rebuilt for agent pace. If your tech stack's assumptions about "how fast work arrives" predate agents, they have an expiration date. cursor.com/changelog

John Jumper — Nobel laureate, AlphaFold — left DeepMind for Anthropic. The single clearest individual talent signal of the year. Where the best scientists move tells you where the scientific-AI moat is forming, and this week it pointed in one direction. For executives in healthcare, pharma, and research, the question isn't whether scientific AI matters — it's whether your AI literacy is keeping pace with where the frontier talent is voting with their feet.

Midjourney shipped a full-body medical scanner. A text-to-image company now makes a 60-second ultrasound scanner with 500,000 sensors, built with Butterfly Network, first "spa" opening in SF in 2027. This is not an AI feature bolted onto a product — it's a text-to-image company becoming a medical-hardware company overnight. Every industry now has an Anthropic-shaped hole opening in it. The portfolio move is to name the one opening under your sector before someone else builds in it. midjourney.com/medical

Agents got a wallet and a fake ID in the same week. Nous Research shipped a Hermes + Stripe integration that lets agents buy things and pay APIs, while Cloudflare launched temporary accounts that hand an agent a throwaway identity and a live deployment for 60 minutes. Agents have stopped just generating — they're transacting. If your procurement, finance, and compliance functions don't have a posture on "our AI can now spend money," build one before the first invoice shows up. blog.cloudflare.com

The US government is weighing equity stakes in frontier AI labs. Treasury and Commerce are reportedly exploring direct stakes in AI companies, Bernie Sanders floated a public-ownership bill, and the whole question got aired at the G7 AI summit alongside Amodei and Hassabis. The ownership and access layer of AI is becoming political fast. Enterprise buyers signing multi-year AI commitments now have a regulatory and access-risk variable that didn't exist a quarter ago — model it.

Six items from a week that had sixty. Full daily version published at The Leverage Brief.

———————————————————————————

Each week: one real tool I built, the stack behind it, what it took, what changed. This is a free preview — the full breakdown series, covering every tool in the Portfolio Leverage Co. stack with copy-paste architecture and adaptation guides, is coming as a paid add-on.

The problem showed up as a spreadsheet and a tired development director. The nonprofit was running a major-gifts campaign with a 110-donor cultivation pipeline — real relationships, real giving histories, real timing that mattered — and managing the whole thing out of a grid of cells and a development director's memory. Every donor needed a personal touch at the right moment: the right ask, referencing the right prior gift, in a voice that sounded like the organization and not a mail merge. At 110 relationships, that's not a workflow. That's a full-time job nobody had budget for.

That's not a fundraising problem. That's an architecture problem.

So I built ChaiRaise. Here's what it does. It holds each donor's full history — giving record, cultivation stage, last contact, personal context — and for any donor, or any segment, it drafts the next cultivation touch through the Claude API: the email, the ask amount calibrated to their capacity and history, the specific reference to what they last gave and why. The development director opens it in the morning to a stack of drafts that already sound right, edits what needs a human hand, and sends. The pipeline that used to live in one person's head now runs as a system.

The architecture is the same spine as everything in the stack: a lightweight layer to segment donors by stage and capacity, the Claude API doing the actual drafting and voice-matching, structured donor records enforcing that no relationship falls through a crack, and a clean review surface so a human approves every word before it goes out. No vector-database cathedral. No fine-tuning. It runs for less than a board dinner costs.

Three things made the difference, and none was the model. First, the voice calibration — getting the drafts to sound like this organization's relationship with this donor took more iteration than any code, because a major-gift ask that sounds generic is worse than no ask. Second, the capacity logic: teaching the system to scale the ask to giving history honestly, instead of either lowballing or overreaching, was pure domain judgment encoded into structure. Third, the human gate — ChaiRaise drafts, the development director decides. It does the codified 80%. The relationship stays human. That's the entire thesis of the book, applied to a 110-donor pipeline.

The result: a development director who was drowning now reviews instead of composes, the right donors hear from the organization at the right moment, and the relationships that fund the whole mission stopped depending on one overloaded person's memory.

Next week: back to the operator stack — the workflow I run to turn a week of raw research into this newsletter without losing the thread.

The full Build Breakdown series — every tool in the Portfolio Leverage Co. stack, the exact architecture behind each one, and how to adapt it for your org — is coming as a paid add-on to The Leverage Brief.

———————————————————————————

The research that feeds this newsletter — AlphaSignal, Rundown AI, TAAFT, AI Fire, The Code, and the operators I track daily — generates 50–60 items a week. The Leverage Brief carries six.

The Leverage Signal is the five-minute weekday read for the same audience: Portfolio Executives, CHROs, CTOs, CAIOs, CLOs, and fractional executives in AI-exposed roles; founders and operators running AI-leveraged businesses; executive coaches and L&D leaders whose clients are navigating AI transformation. Two or three highest-signal tool or agent releases from the prior 24 hours, each with a one-sentence "what this means for your work" framing. One open-source repo worth knowing. One deployment pattern from the operator community. One macro signal — model, funding, regulatory, talent — in your working context before 9 AM.

I want to know the readership is there before I build the pipeline.

→ SUBSCRIBE TO THE (M–F) LEVERAGE SIGNAL — Ten seconds to opt in.

———————————————————————————

Paid AI training and evaluation work for senior operators. Flexible, remote, NDA-bound.

The expert-data market is tracking toward $100B/year by 2027. Frontier labs pay domain experts for post-training evaluations, RLHF, and agent environment design. Senior operators and licensed professionals in HR, finance, legal, medicine, and engineering consistently land in the $50–$150/hr band. This week's news that OpenAI trained GPT-5.5 with 600 doctors is the same market in plain sight: your expertise is the training signal, and it's getting paid for.

Mercor — Premium rates ($75–$150+/hr for qualified domain experts). Strictest screening; best fit for CHROs, attorneys, physicians, and senior engineers. Valued at $10B after its October 2025 Series C. Note: impacted by a March 2026 supply-chain attack — review their post-breach disclosures before onboarding. → Apply via referral link

micro1 — Faster onboarding via the Zara AI interview; multiple attempts allowed. Crossed $100M ARR in December 2025. Expanding into robotics pre-training and agent simulation. $20–$150/hr depending on domain. → Apply via referral link

Meridial (by Invisible Technologies) — Expert contractor work across law, STEM, finance, linguistics, coding, and safety. No prior AI training experience required. Typically responds within 48 hours. Strong fit for specialized domain experts. → Apply at meridial.ai

Apply to all three. A $50–$150/hr income node built on expertise you already have.

———————————————————————————

The July cohort starts in about two weeks. The last seats are open now.

This week's signal is the whole argument for the room. When coding stops being the moat and domain expertise becomes the appreciating asset, the only thing that converts your expertise into leverage is a workflow you actually shipped — pointed at a real decision only you can make well.

Fifteen seats. Twelve weeks. Three things you leave with: a redesigned operating week built around compounding output rather than calendar entropy; a custom AI workflow or tool you actually ship in your real work — not a prototype, a working artifact with before-and-after data that proves your expertise changed the output; and a positioning narrative that names your value at the convergence of talent and technology, because that's where the CAIO titles are going.

You don't close your gap by coding a little better. You close it by attaching the agent to the judgment only you have — and proving it ships.

→ Apply: portlev.com/cohort

———————————————————————————

Find the one decision in your function that only your expertise can make well.

Not the busywork — the judgment call. The comp exception. The diligence verdict. The policy edge case. The read that a smart generalist would get wrong and you'd get right in ten seconds because you've seen it a hundred times.

Now build the workflow around it. Let the agent do the assembly — pull the context, draft the analysis, structure the output. You supply the judgment at the one point where it matters. Then write down, in plain language, every place your expertise changed what the machine produced.

That document is two things at once: proof that you can direct AI with knowledge the agent doesn't have, and the exact artifact frontier labs are now paying domain experts to create. The 400,000 sessions told you which side of the line the value is on. This is how you get on it.

Start it this week.

———————————————————————————

Reply to this email. I read every one.

Yuri Kruman / Founder, Portfolio Leverage Co. · 3x CHRO · AI Trainer: OpenAI, Meta, Microsoft / PortLev.com · LinkedIn · AIWageGap.com

———————————————————————————

"Four hundred thousand sessions, and the people who won weren't the coders — they were the experts who picked up the tool. The syntax moat drained overnight. What's left is judgment, and judgment plus an agent is the most valuable thing in the building."

— The Leverage Brief, June 2026

———————————————————————————

© 2026 Portfolio Leverage Company. You're receiving this because you subscribed to The Leverage Brief. Unsubscribe in one click.