Hey ,

Welcome to the 17th edition of the 3-2-1.

Three posts showed up in my feed within 72 hours last week. They came from McKinsey, the CEO of Zapier, and the CEO of Box. All three were making the same argument.

Your people are the product.

Let’s get into it.

3 Things for Work (in L&D)

1. AutomationBench (Wade Foster, Zapier)

• Zapier built a benchmark that drops AI models into real business environments (Sales, Marketing, Ops, Support, Finance, HR) and tests whether the work actually got done.
• Scoring is based on whether the right records are updated and the right messages are sent.
• Only GPT 5.5 cracked 10%
• 🏋 Effort ≈ 5 min to read, 15 min to explore

2. The Skill Change Index (McKinsey Global Institute)

• McKinsey’s new index maps which skills will be most and least exposed to automation.
• Negotiation, problem solving, and leadership matter more, not less.
• AI changes how skills are used, not whether they’re needed.
• For L&D leaders, this is the institutional anchor you’ve been waiting for. The capabilities you’ve been fighting to prioritize are the ones that just got classified as durable.
• 🏋 Effort ≈ 10 min to skim

3. The Agent Operator Job Spec (Aaron Levie, Box)

• Levie describes a role emerging in almost every company that uses AI seriously: dedicated people who map workflows with agents, wire internal systems, manage context, create evals, and figure out where the human stays in the loop.
• Not individual employees doing it on the side. A real job. “The future of software engineering that you’ll see grow in non-tech companies.” The number being floated is 500K to 1M of these jobs in five years.
• Nobody has defined how you develop them.
• 🏋 Effort ≈ 3 min read

2 Things for Life

1. AI computes the known world. The unknown one is still our job.

“Logic, powerful as it is, can only compute what is, not what could happen. It can only understand the future as a version of the past, so it can only survive inside board games, mathematical simulations, medieval theologies, planned communities, rigged economies, and other artificially bounded environments with never-changing rules. It cannot react to change, adapt to emergent threats, exploit fresh opportunities, cope with uncertainty, handle instability, innovate, or grow.”

— Angus Fletcher, Storythinking

2. Mortality is not the enemy of a good life. It is the condition that makes one possible, and separates us from AI.

“It is by consciously confronting the certainty of death, and what follows from the certainty of death, that we finally become truly present for our lives.”

— Oliver Burkeman, Four Thousand Weeks

1 Idea from Me

The 10% Ceiling

I want to sit with that Zapier number for a moment. Only one AI model cracked 10% on real business tasks. Not 10% of the easy ones. 10% of the full mix: filing the right CRM update, sending the right follow-up, not breaking a downstream process.

That’s the best models in the world, as of May 2026.

I don't think that number stays around 10% forever. But I also don't think it moves as fast as the hype cycle suggests. And while the industry debates when AI becomes “good enough,” three things are true right now, today, for every L&D leader reading this.

The first is that the 90% failure rate has to land somewhere. Every task AI gets wrong either gets caught by a human or spreads through the business. Which means the most valuable people in your organization are not the ones who can use AI fastest. They are the ones who can tell when the AI is wrong. That capability has a name. It is judgment.

The second is that McKinsey just validated what we've been saying for two years. Negotiation, problem solving, leadership. The messy, ambiguous, judgment-heavy work. These skills are not going away. They are getting more valuable because the stuff below them (retrieval, drafting, routing) is getting cheaper.

The third is that Levie just described a new job category with no development path. Agent operators. The people who sit between the business and the AI, translating workflows, curating context, designing evals, deciding where the human stays in the loop. Your organization is about to hire them, or promote them internally. And when HR comes to L&D and says “what’s our development track for this role?” most L&D functions will have to make one up in a month.

I want to name what most L&D strategies miss about this.

The agent operator job is not a technical role dressed up in a new hat. Yes, there is technical craft involved. But the durable capability underneath the role is the same capability we've been building for years:

• Framing a problem clearly enough that someone else can act on it.
• Breaking work into steps.
• Noticing when the output does not match what you actually wanted.
• Giving feedback that changes the next attempt.

These are the skills that make someone an effective manager of people. They are the same skills that make someone an effective operator of AI. One capability model, not two. One budget line, not two. And the L&D leaders who see this first will be the ones their CEOs listen to when the agent operator hiring wave hits.

So I would ask one question of your current strategy.

If no AI model can complete 10% of your organization’s real work, and McKinsey is telling you the human skills matter more, and Box is telling you there is a job category emerging with 1 million seats to fill, what is your L&D function actually building for?

Hit reply and tell me. I read every one.

That’s it for this week. Enjoy your Sunday.

I’ll be back in two weeks ✌️

Andrew

P.S. 👉 If you’re trying to figure out what an agent operator development path looks like for your organization, reply to this email. I do a handful of free 30-minute strategy calls each month for L&D leaders working through this exact question.