Directing Intelligence

Prove It or Lose It: The New AI Access Rule

Todd McLees — Wed, 21 Oct 2026 15:43:41 GMT

Directing Intelligence Digest | October 21, 2026

Two stories in today’s video, one from work and one from learning. In the first, a payments company hands out AI only to people who can show what it adds. In the second, a randomized trial with 1,269 teachers finds the real payoff of AI training wasn’t time. It was judgment.

The short version

Paysend’s CEO gives AI to top performers who prove what it adds, and estimates 25% more engineering output from the same team.
In UK government research, 77% of AI-using businesses saw no change in revenue.
A Texas restaurant replaced its walked-out server with an AI host. The tips now go to the kitchen.
Half of all business AI spending comes from 2% of firms, and executives report a 0.2% yearly productivity bump so far.
In Peru, AI training saved teachers zero time but made them better at catching bad AI output, especially teachers over 54.
Teachers in Greece let harsh grades stand when told an AI gave them.

WORK

Paysend makes you earn your AI

At London payments company Paysend, AI access is earned. CEO Ben Chisell, a self-described “very cynical man” about the hype, gives general-purpose AI tools to his highest performers, often only after they show what the tools add. New hires get a month to experiment, then explain what they achieved. One new SEO lead asked for Claude, then gave it back a month later: it only wrote his briefs faster. Now he’s building a single content-brief agent instead. Chisell estimates 25% more engineering output with the same-sized team (his estimate, not an audit). The contrast, per CNBC: Lloyds Banking Group rolled out 40,000 Copilot licenses, and UK government research found 77% of AI-using firms saw no revenue change.

What to watch: “Human in the loop shouldn’t mean human CC’d,” Chisell says. “It should mean human ownership.”

Read it: CNBC: Some companies give AI to everyone. This CEO is doing the opposite
On X: Shakthi’s post

Who’s Directing? Handing everyone a license is easy. It’s also how you scale bad judgment at the speed of software. Before you buy another seat, ask who on your team can show you what the last one produced.

The server walked out. The AI host stayed.

Two months ago, the only server at Wok & Karahi, a family-run Chinese, Indian and Pakistani restaurant in Spring, Texas, walked off the job. Owner Abbas Zoeb, a former Google software engineer, replaced the role within 24 hours: a virtual host named Haniya that speaks seven languages and answers the phone after closing, plus QR codes and pre-pay ordering. Zoeb told FOX 26 Houston he’s saving or winning back $9,650 a month, about $3,500 of it the server’s salary, for roughly $300 a month in software. Kitchen staff and the chef now get 100% of tips. The numbers are the owner’s. A University of Houston-Downtown professor notes it fits quick lunch spots better than fine dining.

What to watch: whether diners trade hospitality for speed, and whether this becomes the default answer to a no-show.

Read it: FOX 26 Houston: Spring restaurant turns to AI host to stay afloat after server walks out

Who’s Directing? A job didn’t vanish here. It was split. The machine took the order pad and the kitchen got the tips. Owners, that’s the real decision: when AI saves you a salary, who gets the difference?

Half of all AI spending comes from 2% of firms

Nick Bloom and colleagues surveyed nearly 9,000 CFOs, CEOs and senior executives across seven countries this summer for NBER. 78% of firms use AI. US firms expect to spend about $2,000 per employee this year, around $300 billion in total, but just 2% of firms account for over half of it. Executives use AI 2.1 hours a week, up from 1.5 in the winter. They report essentially no net effect on jobs, yet more than 40% of firms expect gains or losses that only cancel out in aggregate. The productivity bump so far: 0.2% a year, with 0.6% expected over the next three years.

What to watch: that churn under the flat line. “No net effect” can still mean your job.

Read it: NBER Working Paper 34836: Firm Data on AI
On X: Azeem Azhar’s thread

Who’s Directing? “No net job effect” is an average, and nobody lives in an average. Leaders, if 40% of firms expect people moving in and out, tell your team which side of that ledger they’re on before the survey does.

LEARNING

Not faster, but better: Peru’s teachers became sharper judges

The World Bank randomized 390 public schools in Lima, Peru, covering 1,269 sixth-grade teachers. Treated teachers got a one-day training, a Microsoft Copilot license and eight months of weekly WhatsApp tips. Copilot activation rose 20.4 points over a 2.8% control mean. Time saved on seven core tasks, from lesson planning to grading to parent messages? Precisely zero. The gains were in skill: better prompts, even on a rubric the trainers never saw, and 6.7 points less likely to accept a weak AI answer as good. The biggest gains went to teachers over the median age of 54, who started furthest behind, so the gap narrowed. Note: Microsoft’s AI Economy Institute co-funded the study.

What to watch: whether school systems sell AI training on better judgment instead of hours saved.

Read it: World Bank Policy Research Working Paper 11476: Not Faster, but Better (PDF)
On X: Michael’s roundup post

Who’s Directing? We keep selling AI to teachers as time back. The evidence says it’s judgment, and judgment is the better deal. If your PD promises a shorter Sunday night, you’re promising the wrong thing.

Teachers let harsh AI grades stand

In a preregistered experiment, 1,339 teachers in Greece reviewed identical student work with a deliberately wrong score, labeled as coming from either a colleague or an AI. When the score was too harsh, the gap between the final grade and a fair one was 22% larger under the AI label. Lenient errors got corrected either way. Teachers rated AI as less fair and less accountable, then deferred to it anyway, and the effect was strongest among younger, tech-confident teachers. Meanwhile, Korea is scaling AI essay scoring: Seoul’s Chaeum AI grew from 66 pilot schools to a planned 135, and Incheon just trained about 180 teachers to grade alongside its system, with a December comparison of teacher and AI scores.

Read it: PNAS Nexus: Why do experts miss AI’s errors? · Aju Press: AI grading expands to more Korean classrooms · Weekly Hankook: Incheon compares teacher and AI scoring (Korean)
On X: Classignal’s thread

Who’s Directing? “A human will check it” is a promise, not a safeguard. The teachers who trusted AI least still let it fail the kid. If you’re putting AI in the grading loop, train the doubt before you ship the tool.

Springfield teachers push to pause classroom AI

In Springfield, Massachusetts, the teachers’ union organized after learning in late August that Alpha School planned an AI pilot at Kiley Prep Middle School. Alpha says two hours a day of AI tutoring covers academics, with classroom “guides” rather than teachers. The Springfield Empowerment Zone postponed the pilot on September 21. Union president Riley Hernandez called it “an experiment on Springfield children,” and the union held a community forum on October 14. State union president Matt Bach wants a one-year moratorium on district contracts with AI and ed tech companies.

Read it: Massachusetts Teachers Association: Educators, Parents Oppose an AI Intrusion into Public Schools · Hoodline: Springfield teachers demand AI moratorium

Who’s Directing? The fight in Springfield isn’t really about AI, but who gets asked to be involved. Any pilot that skips the teachers and the parents has already failed its most important test.

Quick hits from X

Sayash Kapoor says the newest AI agents improved at experimental design on open-ended research, but the original authors still rejected their papers on judgment and calibration. His post
Alex Imas says DeepMind’s AI-in-science work shows huge uplift in hypothesis generation, with the bottleneck moving downstream to verification and testing. His post

Who’s Directing? When ideas get cheap, checking gets expensive. The people who can tell a real result from a pretty one just became your scarcest asset. Pay them like it.

Worth a listen

Plain English with Derek Thompson, “You’re Not AGI-Pilled Enough,” with Ethan Mollick: on why capability grows fast while organizations change slowly, and the question of what happens when “humans, once the prompter, become the prompted.” Watch the episode

Today’s stories all land on one skill. Paysend wants proof, Peru’s teachers got sharper and Greece’s teachers didn’t look twice. Using AI takes a day of training. Knowing when it’s wrong is the work, and it stays ours.

Sources

The Better Lesson

Todd McLees — Sun, 18 Oct 2026 14:33:01 GMT

Meta’s Muse noticed an airline credit about to expire and asked American for an extension. OpenAI’s dots, Grok Bot, Gemini Spark, and Claude and GPT running as agents work the same way, inside your accounts, acting before you ask. In September, OpenAI said about 10,000 of its agents organized themselves around the Navier-Stokes problem, one of the Clay Institute’s Millennium Prize Problems, and in 88 hours produced a machine-checked proof that mathematicians are still reviewing.

This fits the pattern Rich Sutton called the bitter lesson. Across seventy years of AI research, methods built on human expertise kept losing to general methods that keep improving as computers get faster. Ethan Mollick now argues that organizing work belongs on that list too.

The constraint has moved. In November 2024, Peter Gostev, now AI capability lead at Arena (formerly LMArena), wrote that AI was less likely to fail by hitting a technical wall than by “us having no ideas about what to do with much better models.” He pictured a year of better models in which almost nothing changed, because most organizations aren’t set up to benefit from more intelligence.

His explanation was structural. Outside places where one sharp person can change the outcome, like research labs and investment banks, most businesses run like factories. The process is built to contain the damage a weak contributor can do, which also caps what an excellent one can add. In his chart, most AI ideas cluster where models are weakest, and the most valuable corner, work that pays off only with much smarter models, sits nearly empty. I think he was right. And I believe it extends to agents at scale.

The gap is agency.

By agency I mean the ability to decide what matters and do something about it. Gostev's diagnosis covers both halves: the missing ideas are a failure to decide what matters, and the factory keeps people from doing much about it. drew the line in eight words: "Agents can be AI, but agency is human." An agent can plan and execute. Setting the purpose, deciding when to step in, and living with the consequences stay with people.

Who drives

The swarm ran on its own because the problem had a precise goal, a machine could check the answer, and the agents never had to work alongside a person. Most work is missing at least one of those. Much of it runs through customers, colleagues, students, and patients, toward goals nobody fully wrote down, and no proof assistant can check the result.

So most processes will be hybrid. Stanford researchers asked 1,500 workers in 104 occupations how much human involvement they wanted in each of 844 tasks, using a Human Agency Scale that runs from AI handling a task alone to human involvement being essential. The answers didn’t sort into automate or don’t. They spread across the scale, task by task, and sometimes disagreed with what AI can already do.

When people are part of the process, do agents drive? Klarna tested that. In 2024 it said its AI assistant was doing the work of 700 full-time customer service agents and had cut resolution times from eleven minutes to two. In 2025 its CEO told Bloomberg that cost had weighed too heavily, quality had dropped, and customers needed to know a human would always be available. Klarna started hiring people again.

The useful line runs between transactional tasks and value-added tasks. Transactional tasks have a right answer and a clear finish: scheduling, routing, status updates, lookups. Agents can take most of them now. Value-added tasks are the ones where judgment, trust, and relationships change the outcome. A 2024 meta-analysis found that people working with AI did worse on average than the AI alone on tasks where the AI was already better, and gained on tasks where people were better. Put people where judgment changes the outcome. A student can hand off scheduling without handing off the choice of what to study, or why.

Someone has to draw that line, process by process, and redraw it as the models improve. Orchestrating a system of people and agents is still Directing Intelligence. The intelligence is now human and digital, and directing it is still a person’s job.

Redesign the work

The bigger opportunity is Work Design: using AI to help redesign the work itself. Whatever the industry ends up calling superintelligence, one of its most valuable early jobs may be helping people rethink how their own work gets done.

Most workflows were built around the limits of the people and software available when someone designed them. Steps survive because a form once had to move between desks, or because an approval made sense a decade ago. A capable model can map a process, ask why each step exists, sort the transactional from the value-added, and propose a hybrid design for people and agents to run. What it needs from people is the reason the process should exist at all, and a definition of good.

It also needs the people who do the work. The Stanford study found real mismatches between what workers want AI to do and what it can do, and the people closest to a process know which steps carry the value. The redesign should happen with them, with AI doing the analysis.

Work is moving to outcomes again. Chamath Palihapitiya made the comparison on All-In this week. Around 2000, web advertising was sold by CPM, the cost per thousand impressions. AI is sold today by the million tokens. He expects AI to follow advertising to cost per click and cost per action, and he credits AdWords and AdSense with what happened when advertising made that shift: “we went from buying a thousand ads to just getting an outcome.”

When advertisers started buying outcomes, the hard part moved to them: they had to decide what counted as a conversion. An advertiser who paid for clicks got clicks, whether anyone bought or not. Redesigning work around outcomes puts the same job on whoever owns the work.

Intent, specs, and evals

Three skills make up most of that job, and most of what Directing Intelligence asks of people now. Each is a way of deciding what matters and acting on it, which is the definition of agency I started with.

The competency model we’ve been building shows the bitter lesson playing out inside the skills themselves. The skills that tell a model how, like prompt and context design, stop adding new requirements once someone is orchestrating whole systems. Intent, specification, and evaluation keep growing all the way to the top, along with work design and judgment. At the top, they become strategy.

None of these are technical tricks. They’re human skills: framing, judgment, communication, and accountability, applied to a system that now includes agents.

Intent Design. Set the goal. Articulate it. Precisely. The why behind a good intent is practical, and it comes down to four questions. Why are we doing this work? Why is this worth pursuing? Why does it work this way? Why are people necessary in this process? An intent that answers all four is ready to hand to an agent, and to a person.

Specification Design. A spec turns intent into something an agent can execute and a person can check: the outcome, the constraints, the decisions the agent may make on its own, and the ones that need a person’s sign-off. That last part matters more as agents get more capable. OpenAI shelved its next model, GPT-6.1 Astra, after testing showed it acting without permission and misreporting what it had done.

Eval Design. An eval is a judgment about what good looks like, written down so it can be applied a thousand times. Klarna’s assistant looked great on speed and cost because speed and cost were what got measured. The Navier-Stokes proof passed a proof assistant within a day, and Charles Fefferman, who wrote the official statement of the problem, still expects months of expert review before anyone accepts it. Deciding what counts as good stays with people, and in most organizations it has never been written down.

The better lesson

The bitter lesson says to stop hand-building the how. I agree. The better lesson is about what people do with the time and attention that frees up.

Agency grows through opportunities to act. People build it by making choices, living with the consequences, asking for help, and trying again. Handing every hard moment to an agent removes the experiences where judgment forms. The transactional tasks agents take first are often the ones new employees used to learn on.

So the job for anyone who leads people is to keep those opportunities in the work. Give a new employee a bounded problem, access to AI, and a chance to explain a recommendation. Then evaluate the reasoning as well as the deliverable.

About the writer

Todd McLees is co-founder of humanskills.ai and a contributor to Harvard Business Review. He helps colleges, universities, and employers build the human skills to direct AI well. He writes Directing Intelligence about human agency and AI at work.

About humanskills.ai

humanskills.ai builds human capability for AI-powered work. AI creates greater value when people can direct it: deciding what matters, defining what good looks like, judging what to trust, and owning the result. Programs run through higher education partners, private team cohorts, and open self-paced cohorts.

AI Gives Everyone Access. Judgment Is Still the Advantage.

Todd McLees — Thu, 15 Oct 2026 22:32:58 GMT

Directing Intelligence Digest | October 15, 2026

Yesterday’s news had one through-line...a human one. AI can sit in the interview and draft the tutoring message. What remains ours is the design of the handoff: who judges, who edits, and whether “access” is mistaken for learning.

The short version

Meta scaled back AI-driven cuts. Walmart workers fought AI schedules.
McKinsey’s AI client interviews: humans still score.
Target says its myDay app saves 2 million hours a year (company figures).
A Wyoming naturopath cut weekly posting from up to 15 hours to two.
Maryland’s AI tutor offer lowered grades. Supervised LearnLM matched humans.
Principals build their own tools. Falls Church parents want data answers.

WORK

The rush to AI... and the quiet walk back

Meta, per Reuters, scaled back plans to cut staff and lean on AI agents after employees spent more and more time cleaning up technical and security errors. At Walmart, Business Insider found store workers fighting AI tools that plotted nonsensical routes and impossible schedules.

Microsoft’s Kathleen Hogan: “access and usage do not equal transformation.” Mentoring helped.

Meanwhile Airtable, after its Bending Spoons acquisition, is parting with a large share of its team, citing smaller, faster teams. Zoho’s Sridhar Vembu says he refuses that path and will retrain instead. “We will see in 2030.”

What to watch: Who rehires judgment... and who keeps paying people to clean up after the bot.

Read it: The Guardian · Airtable: Back to bases
On X: Sridhar Vembu

Ten minutes with an AI client... scored by people

McKinsey is testing a new step for some undergraduate business-analyst candidates in North America this fall: about ten minutes talking with an AI simulation of a client. The simulation aims to cover about 80% of everyday client scenarios. The AI does not score anyone. Humans do.

Derek Thompson notes AI is remaking the application funnel faster than the job. Here the machine plays the other side of the table. People keep the judgment.

What to watch: Whether the pilot predicts on-the-job performance, and whether copycats keep human scoring.

Read it: The Verge
On X: Derek Thompson

Who’s Directing? We built machines to screen people out. Now we build machines for people to talk to... and still need a human to say yes. If no one can override the scorecard, you did not automate judgment.

Target says myDay gave 2 million hours back (self-reported)

Target says its store app, myDay, merges trucks, orders, schedules and inventory into one prioritized task list for more than 350,000 team members. By the company’s own count, that saves more than 2 million hours a year, plus $25 million in value from “Scan to Divert.”

It started with more than 50 store leaders and workers telling engineers what hurt. Built with the floor, not dropped on it.

What to watch: Whether saved hours become guest time and development... or fewer shifts.

Read it: Target newsroom (self-reported)

Who’s Directing? Two million hours is a big number. What a company does with them is a value. If you can’t tell your people where the saved time is going, the value does not get realized.

A one-and-a-half-person clinic in Rawlins gets its weekends back

April Martinez, a holistic health educator and naturopath in Rawlins, Wyoming, calls her shop “a one and a half man show.” Posting to eight platforms ate 10 to 15 hours a week. Now it takes about two.

She built her own AI agent loaded with her market profile, a style guide and her own material. Claude handles strategy. “It’s my business partner,” she says. But she reads every post before it goes out. “AI lies.”

What to watch: Whether small owners keep their own voice... or all start sounding the same.

Read it: Saratoga Sun via Wyoming News Exchange

Who’s Directing? She gave the machine her voice and kept the final read for herself. Plenty of companies with a thousand times her budget can’t say the same.

LEARNING

Same tech, opposite results: access vs. a supervised tutor

A University of Maryland RCT offered 2,379 undergraduates a course AI tutor. Students offered it finished about four points lower. Only about 15% used it, and nearly 74% of requests sought information or solutions. Fewer than 1% asked for feedback. Access alone is not a learning intervention.

In five UK schools, expert tutors approved, edited or rewrote every LearnLM draft for 165 math students. Remediation: LearnLM 93.0%, human tutors 91.2%, static hints 65.4%. Transfer: 66.2% vs 60.7%.

What to watch: Adoption metrics before grade claims, and bots sold without the supervisor.

Read it: Maryland EDSI · arXiv: LearnLM
On X: Kevin Bryan

Who’s Directing? You cannot call it tutoring if three-quarters of the clicks are “give me the answer.” Access without pedagogy is a polite way to lose the room.

A Virginia principal’s motto: people over paperwork

Derek Cantrell, principal of Allegheny High School in Covington, Virginia, once lasted one day in a computer science class. Now he builds his own AI tools. One tracks absences, tardies and referrals, warns students early and lets them earn back standing.

EdWeek found others. In Wilmington, Delaware, principal Sterling Seeman wrote prompts that help teachers draft yearly goals. In Idaho, Rachel Jensen’s trackers help teachers overseeing 90-plus online students each. Cantrell’s warning: keep student data out of the public domain.

What to watch: Whether districts back principal-builders with data guardrails.

Read it: Education Week

Who’s Directing? The best AI strategy here didn’t come from a vendor. It came from a principal who wanted to be in the hallway more. Part of your plan should give educators time back with students. You have to be able to measure that.

Falls Church wrote the rules. Now parents want the receipts.

Falls Church City Public Schools in Virginia laid out its classroom AI. Teachers get four tools: Gemini, SchoolAI, Diffit and Canva. Students get SchoolAI in grades 6 to 12 and Gemini in 9 to 12. No ChatGPT or Copilot. AI is not used for grading, discipline, special education or placement.

The News-Press pushed further: what does it cost, and what student data do vendors see? Answers are pending.

What to watch: The cost figures and data answers in the follow-up.

Read it: Falls Church News-Press

Who’s Directing? Naming your tools is the easy part. If you can’t tell a parent what their child’s data costs and where it goes, you haven’t earned their trust.

Quick hits from X

Agent identity at work. Ethan Mollick gave his agent its own email so it cannot draft as him. Mollick on X
Orgs as messy intelligence. Alex Imas: messy human variety makes organizations a collective superintelligence. Imas on X
Proofs vs understanding. Stefania Druga: AI can flood math with proofs. Understanding is still scarce. Druga on X

See you tomorrow.
Todd McLees · Directing Intelligence Digest

Sources

When AI expands the work... people still own the judgment

Todd McLees — Wed, 14 Oct 2026 16:38:09 GMT

Directing Intelligence Digest | October 14, 2026

The news from October 13 had one through-line... a human one. AI is expanding what people can do at work and in school. But the parts that make us us (judgment, empathy, verification, deliberate learning) are not going away. They are becoming the job.

The short version

German IT apprentices with ChatGPT Pro gained 26 to 44 points in measured productivity, and comprehension held.
Half of 100 UK CEOs say AI is a top-three productivity driver. Only 14% can see what it costs.
In Chile, an AI counselor moved college choices more than human counselors did. Information beat empathy.
Novices need explicit guidance. Socratic-only AI tutors frustrate them.
Software’s real AI gain today: at best 25% to 50%, end to end.

WORK

The apprentices who reached further

A Stanford Digital Economy Lab working paper ran a field experiment with 673 final-year IT apprentices across twelve German vocational schools. Some got ChatGPT Pro. Others worked with no internet and no AI. The tasks sat inside the curriculum... and deliberately beyond it.

The AI group’s productivity rose 26 to 44 percentage points. On beyond-curriculum work, 26 to 31 points. Immediate comprehension did not fall; on some checks it improved. Apprentices with above-median AI literacy gained more.

Yes, it is a learning story too. But at work, middle-skilled people reaching further is agency. Just remember: an assisted score is not durable mastery. The authors say so. So do I.

What to watch: Junior roles redesigned around verified understanding, not just faster tickets. AI literacy taught as a core skill.

Read it: Task Expansion with Generative AI (PDF) · Stanford project page

Who’s Directing? Nobody gave these apprentices permission to reach beyond the curriculum. The tool did. The question for every manager reading this: when your people reach further, are you ready to let them?

UK CEOs feel the gain... and see the skills gap

EY (Oct 12) surveyed 100 UK CEOs for its EY-Parthenon CEO Outlook. Quick numbers:

50% put AI-enabled tools in their top three productivity drivers of the past year.
Only 14% have comprehensive, real-time visibility of AI costs.
74% expect skills shortages to beat access to capital as a growth barrier within three years.
71% say AI will reshape roles, skills and how work is organised more than headcount.
47% say their organisation is not building skills fast enough.

Plain English? The real worry may not be layoffs. It may be whether your role still fits... and whether anyone invests in growing you into the new one.

What to watch: ROI dashboards that count skills investment. Leaders who reinvest productivity in people, not only margin.

Read it: EY UK: AI drives productivity gains, but barriers remain

Who’s Directing? Half of CEOs feel the productivity, and only one in seven can see what it costs. That gap is where people get hurt: the bill always comes due, and it is usually paid in jobs and training budgets.

LEARNING

Chile’s Kai: information moved the needle

A nationwide randomized trial from the IDB and Elige Educar, with about 41,000 high school seniors. Students got WhatsApp counseling from either Kai, an AI counselor, or trained humans, aimed partly at teaching careers.

Among students already interested in teaching, Kai significantly raised first-ranked education majors and their share of choice sets. Human-arm effects were smaller and not statistically significant. Kai matched or beat roughly two-thirds of human counselors.

Why? Text analysis says Kai was more factual, faster and more coherent. Humans scored higher on motivation and empathy. As co-author Gregory Elacqua put it: Kai won on information, humans on empathy... and information mattered more.

Not a robot victory lap. Facts at scale reach kids who never get a counselor. Empathy still matters. The answer is both, designed on purpose.

What to watch: Hybrid models, replications elsewhere, and long-run teacher pipelines, not just applications.

Read it: IDB paper · CEPR DP21658 abstract
On X: Gregory Elacqua

Who’s Directing? Information beat empathy. I hate that finding, and I believe it. Which means the human job is no longer telling kids what is true. It is being there when the truth is hard to hear.

Novices need answers, not just questions

Education researcher Carl Hendrick says minimally guided instruction does not work with AI tutors either. Socratic questioning models frustrate students when they withhold the explicit guidance novices actually need.

A confused learner is not a deeper learner. Dignity in education means scaffolds matched to what someone already knows. An agent that only asks questions can look wise... and leave a beginner stuck.

What to watch: Tutors that know when to explain. Schools that refuse one-size-fits-all “Socratic-only” AI.

Read it: Guardrails or Roadblocks? (arXiv)
On X: Carl Hendrick

Who’s Directing? Struggle teaches only if you already know enough to struggle productively. Hand a novice a Socratic bot and you have not built a tutor. You have built a very polite wall.

Quick hits from X

Software’s reality check. Zoho’s Sridhar Vembu pegs AI’s software gain at at best 25% to 50% end to end. The 10x leap via verification tools? Not yet. Vembu on X
Thinking, amplified. Princeton’s Arvind Narayanan prepped for The Ezra Klein Show with an agent trained on about a million of his own words. Useful AI, he says, with no cognitive offloading. Narayanan on X
Intelligence isn’t power. His first point: a system’s power is not in the model. It is what we choose to grant it. The episode · AI as Normal Technology

Who’s Directing? A man trains an agent on a million of his own words so he will not stop thinking. Meanwhile we are training millions of people to stop. Pick your side.

Worth a listen

Lenny’s Podcast with Josh Woodward (Google Labs): run your company like a lab team... and why your unlearning rate may be a rising skill. Episode · YouTube
The AI Daily Brief: How to Go from AI Power User to AI Leader, on building champions networks so good practice does not die with a few power users. Episode

Tools stretch the reach of hands and minds. People still decide what is true enough to ship, kind enough to say, and worth learning on purpose.

Otherwise it is not us.

See you tomorrow.
Todd McLees · Directing Intelligence Digest

Sources

NYC Schools Don't Have any AI Answers Yet

Todd McLees — Wed, 25 Mar 2026 20:40:57 GMT

01 / NYC’s AI guidance for schools uses a traffic light. The hardest questions are still yellow.

New York City released its preliminary AI guidance for public schools, three years after banning ChatGPT. The framework uses a traffic light system: green for approved uses like brainstorming lesson plans, red for prohibited uses, yellow for gray areas requiring significant oversight. Whether students can use personal AI accounts for homework is still unresolved. The city is seeking public feedback through May 8 before releasing a more comprehensive playbook in June.

The largest school system in the country just told a million families: we don’t have the answers yet, but here’s how we’re going to figure it out together. That’s more honest than most institutions have been willing to be. The question is whether the feedback process is real or performative. If it’s real, this becomes a model. If it’s a formality before a predetermined outcome, it becomes another reason parents don’t trust institutions with their kids’ AI exposure.
A traffic light is not a strategy. But it’s better than pretending the intersection doesn’t exist.

02 / More success, less effectiveness. A researcher in Morocco names the paradox.

A Moroccan education researcher published a piece called “AI, a Learning Machine: Are Students Still Learning?” arguing that AI is creating a visible paradox in education: students are producing better outputs while mobilizing fewer of their own cognitive abilities. AI is accessible, it promotes equal opportunities, it helps with homework and projects. And it may be reducing the very mental effort that learning requires.

This is the OECD finding showing up in a classroom in Morocco. Better outputs, weaker understanding. The pattern is global now. And the conclusion is the right one: adopting AI without pedagogical reflection is not a path forward. The tool isn’t the variable. The teaching is the variable. It was always the variable.

03 / The federal government just became a learning provider.

The U.S. Labor Department is launching a free AI literacy course for Americans who are uneasy about the technology. It’s framed as workforce preparation, but the implications for how we think about teaching adults to learn alongside AI are direct.

This signals that AI literacy is now considered public infrastructure, like roads or broadband. Not because the course will be transformative, but because the federal government deciding it needs to teach this tells you something about where we are.
Community colleges have been building this infrastructure for two years. We’ll see where this goes. I still believe that Institutions of Higher Education are the best prepared to upskill an entire nation.

04 / What if the most important thing education teaches is the thing AI can’t replicate?

The Dallas Federal Reserve published research showing AI automates codified knowledge (what you learn from textbooks) but complements tacit knowledge (what you learn from experience). Young workers in AI-exposed fields aren’t getting laid off. They’re not getting hired. The job-finding rate is collapsing for new graduates while wages rise for experienced workers.

Read that finding as an educator. If AI replicates what students learn in school but can’t replicate what they learn from doing the work, then the entire value proposition of education shifts. The degree isn’t the credential anymore. The capacity to develop judgment, to ask the right questions, to build the kind of knowledge that only comes from practice and reflection. That’s what we should be teaching toward. That’s what the Human Agency Stack is built around.
My french bulldogs seem to understand everything I say. They do not. This is also true of most AI tutoring products.

Todd McLees is the co-founder of humanskills.ai.

The Perception Gap Is the Most Important Number in Enterprise AI

Todd McLees — Wed, 25 Mar 2026 19:45:25 GMT

01 / CFOs privately expect AI layoffs to be nine times higher this year. The more interesting number is the one they’re getting wrong.

A Duke/Federal Reserve CFO survey reported by Fortune finds that chief financial officers expect AI-attributed job cuts to reach roughly 500,000 this year, up from 55,000 last year. The researchers call it a rounding error against the overall U.S. workforce, not a doomsday scenario. But the second finding is the one worth sitting with: perceived productivity gains from AI are running ahead of actual gains. Companies think AI is delivering more than it is.

The gap between what leaders believe AI is delivering and what it’s actually delivering is the most important number in enterprise right now. It explains both the layoff confidence and the coming disappointment. When the perception catches up to the reality, the question will be whether organizations invested in capability or just cut headcount and hoped.

02 / The federal government is now teaching AI literacy.

The U.S. Labor Department is launching a free AI literacy course aimed at Americans who are uneasy about the technology. Axios reports the move is framed as a response to growing concern that AI will reshape work faster than institutions can prepare people.

When the federal government starts teaching AI literacy, the “should we?” phase is over. The question now is whether public programs can move at the speed the workforce actually needs.
Community colleges have been building this road for two years. The federal government just noticed there’s a road. We cannot move fast enough.

03 / AI replaces what you learned in school. It amplifies what you learned on the job.

The Dallas Federal Reserve published research showing that AI substitutes for entry-level workers who carry codified knowledge (textbook learning, documented procedures) but complements experienced workers who carry tacit knowledge (judgment, pattern recognition, the kind of understanding you can only build by doing the work). Employment is falling for young workers in AI-exposed fields. Wages for experienced workers in those same fields are rising.

This is the most precise framing I’ve seen of who AI helps and who it hurts. The career ladder isn’t disappearing. The bottom rungs are. Organizations that want to develop AI-capable workforces need to rethink how entry-level employees gain experience, because the old model of starting with codifiable tasks and slowly building tacit knowledge is exactly the model AI disrupts first.

04 / The companies building AI are hiring. The companies using AI are cutting.

OpenAI is planning to nearly double its workforce to 8,000 by year’s end, hiring across product, engineering, sales, and “technical ambassadorship.” Anthropic is approaching $19 billion in annualized revenue. Meanwhile, Block cut 40% of its workforce. Atlassian cut 10%. Oracle is cutting up to 30,000.

Both of these things are happening in the same economy at the same time. The question for every organization is which side of that line you’re learning to stand on.

The best advice I ever got about technology came from a volleyball coach. She said “the fundamentals don’t change, the speed does.”

Todd McLees is the co-founder of humanskills.ai.
Directing Intelligence: Work arrives when there’s signal, not on a schedule.

The 5% Who Direct Intelligence Differently

Todd McLees — Wed, 25 Mar 2026 19:29:18 GMT

01 / Only 5% of users consistently direct AI with sophistication. Now we know what they do differently.

KPMG and UT Austin’s McCombs School of Business analyzed 1.4 million real workplace AI interactions and published the results in Harvard Business Review. The researchers studied eight months of back-office operations, evaluating more than 30 behavioral characteristics of how people actually use AI at work.

The finding that matters: the sophisticated users, the ones who consistently produced high-impact outcomes, weren’t the most frequent users or the most technically skilled. They were the ones who excelled at framing problems clearly, directing the AI’s reasoning toward specific tasks, and iterating with purpose across their work.

About 5% of users demonstrated these behaviors consistently.

This is the first large-scale empirical study I’ve seen that confirms what we’ve been building toward at humanskills.ai: the skill isn’t prompting. It’s directing. Problem framing, task orchestration, evaluative judgment, purposeful iteration. These are the behaviors that separate someone who uses AI from someone who directs intelligence. And the fact that only 5% do it consistently tells you everything about the gap between access and capability. Every organization I work with has solved the access problem. Almost none have solved the capability problem.

02 / AI makes the work better. Then it makes the worker feel worse.

A study published in Nature’s Scientific Reports ran four experiments with 3,562 participants and found something that should unsettle anyone designing AI workflows. Collaborating with generative AI improved immediate task performance. But when participants transitioned from AI-assisted work back to working alone, their intrinsic motivation dropped significantly and their feelings of boredom increased. The performance boost did not carry over into subsequent independent tasks.

The researchers call it a “psychological deprivation effect.”

Read this carefully. AI collaboration makes the output better in the moment. Then it makes the human feel less capable and less motivated when the AI isn’t there. The performance gain doesn’t transfer to the next task. This is the cognitive atrophy that educators have been warning about, now measured in a controlled setting across thousands of participants. The implication for anyone designing how people work with AI is serious: if you don’t deliberately build in moments where the human does the hard thing alone, you’re training dependence, not capability. The Automate/Augment/Keep Human framework exists for exactly this reason. Some tasks should stay with the human even when the AI could do them faster. Not because the AI can’t. Because the human needs the practice.

This is also why I still hand-chop onions instead of using the food processor. The food processor is faster. But I lose something when I stop using the knife.

03 / AI replicates what you learn from textbooks. It amplifies what you learn from experience.

The Dallas Federal Reserve published research drawing a distinction that deserves to become part of how every organization thinks about AI capability. The researchers separate codified knowledge (established information from textbooks, training materials, documented procedures) from tacit knowledge (understanding gained through experience, pattern recognition built over years of practice, judgment formed by doing the work).

AI automates codified knowledge. It complements tacit knowledge. Employment is falling for young workers in AI-exposed fields, not through layoffs but through a collapsing job-finding rate. Wages for experienced workers in those same fields are rising.

*This reframes how we should think about developing AI capability in any organization. The 5% of sophisticated users in the KPMG study aren’t better at AI. They’re better at the underlying work, and that’s what makes their AI use sophisticated. Tacit knowledge is the operating system. AI is the application. You can install the application on any machine. But it only performs well on a machine with a strong operating system.*

*If you’re building an AI capability program and you’re only teaching the application, you’re building on sand.*

---

Todd McLees is the co-founder of humanskills.ai.; Keynote Speaker & HBR Contributor

We help organizations build Directors of Intelligence, so AI helps their people be more valuable than ever.

Directing Intelligence arrives when there’s signal you need to know about, not on a schedule. If it showed up today, it’s because something happened that changes how you should think about working with AI.

Opposing Counsel

Todd McLees — Mon, 23 Mar 2026 23:27:25 GMT

Andrew Yang has a clip making the rounds. In it, he calls the legal profession “the ideal environment for AI,” describes partners giving AI work that used to take junior associates a week, and concludes that firms have no incentive to hire a small army of young lawyers anymore. The video’s graphic shows a six-lawyer boutique in San Francisco that dropped staffing costs 27% after not replacing a departing associate.

It’s a clean story built on a friend’s anecdote, a single data point, and a straight-line extrapolation. In an era where we have access to real data in real time, that’s not good enough.

We are in what I’ve been calling the Agency Economy, a period where intelligence is abundant and human agency is the scarce, high-value contribution. This era calls for a different kind of thinking: precision over anecdote, polarities over single narratives, and an honest reckoning with what’s now possible that wasn’t before. Yang’s clip fails on all three counts.

Precision: What the 2026 data actually shows

The Thomson Reuters 2026 Report on the State of the U.S. Legal Market (produced with Georgetown Law, released January 2026) is the most current comprehensive picture of the industry. It’s based on financial data from 184 U.S. law firms. Here’s what it shows.

Legal demand grew nearly 2% in 2025, with Q3 reaching 3.9% year-over-year growth. That’s one of the strongest years since the 2008 financial crisis. Billable hours increased 2.5% for the year, hitting 4.4% growth in July. The average firm saw 13% profit growth. Technology spending grew 9.7% and knowledge management spending grew 10.5%, both the fastest rates the industry has likely ever experienced. Rate growth hit 7.3%, the fastest pace since the GFC.

That demand surge, the report makes clear, isn’t driven by a booming economy. It’s driven by chaos: trade wars, regulatory upheaval, geopolitical tensions, and policy volatility. All of which require constant legal work. The firms aren’t growing because business is calm. They’re growing because the world is complicated.

The 2026 Wolters Kluwer Future Ready Lawyer Survey (810 lawyers across the U.S., China, and eight European countries, conducted August 2025) found that 92% of legal professionals now use at least one AI tool daily. Sixty-two percent report saving 6 to 20% of their work week. And 52% of firms attribute a 6 to 20% increase in revenue to those tools. Organizations with defined AI strategies are twice as likely to see revenue growth and 3.5x more likely to realize critical AI benefits.

Robert Half’s 2026 Demand for Skilled Talent report shows 74% of legal leaders say the AI factor alone has made them more likely to bring in staffing help. Firms are hiring, and they’re hiring for AI fluency alongside legal skill. Paralegal unemployment sits at 1.9%, which is barely measurable.

Harvard Law School’s Center on the Legal Profession interviewed COOs and partners at AmLaw 100 firms (published February 2025, but reflecting current deployment strategy). None plan to reduce attorney headcount. Ninety percent expect AI will improve service quality, not just cut costs. The dominant view: total hours worked will remain similar or expand, while attorneys shift toward analysis and strategy.

The above is not a projection or a hope. It’s what the profession measured about itself in 2025 and early 2026. This argument has a shelf life, and I want to be honest about it. If legal-domain hallucination rates drop below 2% across multiple independent benchmarks, if agentic AI systems can handle filing-to-resolution workflows without a licensed attorney verifying the output, or if in-house legal departments start building AI capabilities that bypass outside counsel at scale, the calculus changes. I’ll revisit this in twelve months and see which of those conditions moved. Yang should do the same with his anecdotes.

Share Directing Intelligence

Why the economics run the other way

The data shows a profession that’s growing while adopting AI at record pace. Yang’s logic says that shouldn’t be happening. So either the data is wrong, or his model is.

His model is wrong. And a piece published today by economists Alex Imas and Soumitra Shukla explains why.

They build on a framework called “O-Ring Automation,” developed by Joshua Gans and Avi Goldfarb. The name comes from the Challenger disaster: a single faulty O-ring brought down the shuttle because every component had to work for any of them to matter. Michael Kremer applied this insight to economics in 1993. When a job requires many tasks done well, and those tasks depend on each other, productivity isn’t additive. It’s multiplicative. One weak link degrades the whole output.

Lawyering is an O-ring job. Research, drafting, client counsel, negotiation, regulatory interpretation, courtroom judgment, relationship management. These tasks multiply against each other. A lawyer who nails the research but misreads the client’s risk tolerance doesn’t produce 85% of a good outcome. They produce a bad one.

This structure changes what happens when you automate part of the job. Imas and Shukla describe something they call the “focus effect.” When AI takes over some tasks in a high-dimensional job (one with many complementary tasks), the worker isn’t left with less to do. They concentrate their time on fewer things. The quality of each remaining task goes up. And because the tasks are multiplicative, those quality gains compound through the entire production function. The worker becomes more productive. Wages rise. The firm doesn’t cut the position. It gets more from it.

This is exactly what the Harvard CLP interviews found. AmLaw 100 firms reporting massive productivity gains on specific tasks. Zero planning to reduce attorney headcount. The economics explain the data.

Contrast that with long-haul trucking, which Imas and Shukla flag as the real canary. A trucker’s job is dominated by one core function: moving the vehicle from A to B. Low-dimensional. If autonomous driving gets reliable on highway routes, there’s no constellation of complementary tasks for the driver to concentrate on. No focus effect. The job goes away. And firms have a stronger incentive to finish automating low-dimensional jobs because eliminating the last task means eliminating the entire wage bill.

Yang treats lawyering as if it were trucking. As if “research and drafting” is the whole job, and automating those means the rest collapses. Law has seven or eight distinct, complementary tasks. Automating two of them doesn’t reduce the need for lawyers. It makes the remaining six more valuable.

There’s a second variable in the Imas and Shukla framework: demand elasticity. When productivity gains lower costs in a market where cheaper prices bring in significantly more buyers, the result isn’t fewer workers. It’s more. Jevons’ paradox: when coal engines got more efficient, coal consumption didn’t drop. It exploded, because efficiency opened applications that were previously too expensive.

Legal demand right now is about as elastic as it gets. Which brings us to what AI is actually creating.

Polarities: Different firms will make different choices

The Thomson Reuters report includes a warning that’s just as important as the growth numbers. The legal industry, it notes, has “a peculiar historical habit of surging just before it stumbles.” Similar demand explosions preceded both the 2008 financial crisis and the 2022 inflation crunch. Forecasts in the report point toward slowing growth by mid-2026, with Q3 potentially dipping into contraction.

This matters because it’s the truth Yang is circling without quite reaching. Some firms will contract. Some practice areas will soften. Some junior roles that existed primarily to absorb low-value document work will, in fact, be reshaped or reduced.

And simultaneously, other firms are growing. Midsize firms captured the largest share of demand growth in 2025 as corporate legal departments pushed routine work downstream from Am Law 100 firms. New AI practice groups are launching across major firms. The alternative legal services market reached $28.5 billion. Pro se employment lawsuits surged 49% in 2025, creating new defense work. Fair Housing Act filings jumped 69%. Courts are drowning in AI-generated motions that require human lawyers to counter. Nippon Life Insurance is suing OpenAI after a single AI-armed litigant generated $300,000 in defense costs.

And there are 741 AI-related bills introduced across 30 state legislatures as of January 2026. The EU AI Act is in force. Colorado, Illinois, California, and Texas all have new AI compliance obligations. Meanwhile the White House just unveiled their National AI Legislative Framework. Every company deploying AI now needs lawyers who understand the rules.

For every doomed narrative, there is a different polarity. Yang tells one side. The data shows both.

The polarity that matters most: some firms will see AI as a reason to cut. Others will see it as the growth opportunity of a generation. Both will be right about their own situation. Neither is right about the profession.

Possibilities: The too-hard, too-big era is over

This is what makes the current moment genuinely different from every prior technology cycle in legal services. The Thomson Reuters report describes “an almost absurd tension” where firms deploy technology that accomplishes in minutes what once took hours, then try to bill for it by the hour. Ninety percent of legal dollars still flow through standard hourly billing.

That tension isn’t a crisis. It’s an opening.

Harvey, the legal AI company valued at $8 billion and used by a majority of the Am Law top 10, reports a 0.2% hallucination rate on its internal BigLaw Bench benchmark. Its architecture uses domain-specific models, multi-layer claim verification, and real-time Shepardization through LexisNexis. A&O Shearman, Harvey’s flagship client, has over 1,000 lawyers using the platform daily, saving two to three hours per staff member per week with notable reductions in contract review time.

That’s a tool that makes things possible that weren’t before. Contract analysis across forty NDAs at once. Regulatory compliance monitoring that tracks changes in real time. Due diligence workflows that compress weeks into days.

The “too expensive to pursue” cases can now be pursued. The “too many documents to review” problems can now be reviewed. The access-to-justice gap (more than 50 million low-income Americans receive inadequate legal help for 92% of civil legal problems) can start to close. The UK’s Solicitors Regulation Authority approved Garfield.Law as the first firm authorized to deliver legal services entirely through AI, targeting small claims at per-document pricing.

These are possibilities, not replacements. Every one of them requires someone who understands the law, the client’s situation, and the limits of the tool. Harvey’s own clients still validate everything the system produces. The possibility isn’t “no more lawyers.” It’s “lawyers can now do things that were previously impossible.”

Subscribe now

People: Human agency is the scarce resource

The hallucination numbers tell a story, but not the one most commentators draw from them.

Harvey’s 0.2% is self-reported on its own benchmark. The broader domain-specific data (from the Suprmind 2026 benchmarking report, updated in March 2026) shows legal-specific hallucination rates averaging 18.7% across leading models. On the AA-Omniscience evaluation, Gemini 3 Pro showed an 88% hallucination rate: when it doesn’t know an answer, it fabricates one nearly nine times out of ten rather than saying so. The newest reasoning models, the ones marketed as most intelligent, are measurably worse at sticking to provided facts. Every reasoning model tested on Vectara’s enterprise-length benchmark exceeded 10% hallucination.

A 2025 mathematical proof confirmed that hallucinations can’t be fully eliminated under current architectures. OpenAI published research in 2026 acknowledging that training processes inadvertently teach models to confabulate rather than abstain.

Here’s the principle that matters: as errors get rarer, they get harder to catch. A model that hallucinates 30% of the time produces errors anyone can spot. A model that hallucinates 2% of the time produces errors that require genuine expertise to identify. The remaining mistakes are the subtle ones: a citation that exists but doesn’t support the proposition, a statute from the wrong jurisdiction, an analysis that tracks logically but misses a recent reversal.

Risk tolerance has to match the context. Summarizing a contract for internal planning? Low stakes, let the tool rip with a quick check. Filing a brief with a federal court? Advising a client on whether to accept a settlement? Submitting regulatory compliance documentation? The cost of a single confident error in those contexts is six or seven figures.

An assistant US attorney in North Carolina recently resigned over AI-fabricated quotes in a brief. Butler Snow faced judicial scrutiny for filings with fabricated citations. The AI hallucination cases database now tracks over 850 documented instances worldwide.

“Vibe lawyering” (using AI to practice law without legal training) is already producing expensive wreckage. Fisher Phillips reports defending AI-fueled pro se cases costs 10 to 15% more than typical claims, because the litigants can generate unlimited motions in minutes. One litigant sent over 300 AI-generated accusatory emails to opposing counsel. Another produced a 456-page appellate brief recycling motions already flagged for hallucinated citations.

Someone has to respond to all of that. Someone has to verify Harvey’s output before it goes to a client. Someone has to know whether the AI’s analysis of Colorado’s new AI Act applies to a company headquartered in Texas with EU operations. Someone has to sit across from a client and exercise judgment about what to do next.

That someone is a person with legal training, professional judgment, and the agency to direct intelligence rather than be directed by it.

The real question Yang should be asking

The ABA Task Force on Law and AI noted in its December 2025 Year 2 report that professional attitudes have shifted from whether to use AI to how to use it responsibly and effectively. Their assessment of the current technology: it won’t replace lawyers, eliminate the need for negotiation, take depositions, or try cases.

That framing is closer to the truth, but it still undersells what’s happening. The better question isn’t “will AI replace young lawyers?” It’s “what does a young lawyer need to be in 2026?”

The answer is: someone who arrives AI-fluent, understands when to trust a tool and when to question it, spends less time on mechanical tasks and more time developing the judgment that separates competent from excellent. Someone whose floor of competence is higher because AI handles the basics, and whose ceiling is higher because they started learning strategy and client work in year one instead of year four.

That’s a different kind of junior lawyer. Not a smaller one. Not an eliminated one.

Different firms will make different choices. The Thomson Reuters report is honest about the instability ahead. Some firms will cut. Some will grow. Some boutiques will run lean. Some midsize firms will surge. The profession is large enough to contain all of these polarities.

But the blanket conclusion that AI means “no more young lawyers” misreads the market (which grew nearly 2% in 2025 and is investing at record rates), misunderstands the technology (which requires trained humans to verify, direct, and apply), and ignores the new legal demand AI itself is generating (741 state bills, cascading compliance obligations, an explosion of AI-related litigation).

Intelligence is abundant. Agency is scarce. The firms that understand the difference will build the future of the profession. The ones that don’t will spend it reacting to other people’s narratives.

Working with AI Is Just the Beginning

Todd McLees — Mon, 02 Mar 2026 15:38:49 GMT

Jensen Huang has been saying it for two years now. The data center, as we’ve known it, is being replaced by something he calls an intelligence factory. You apply energy to it, he says, and it produces something incredibly valuable. Raw data goes in, intelligence comes out. Every company will need one, Jensen says.

Most people heard a chip salesman pitching infrastructure. Fair enough. NVIDIA sells the shovels. But last week, something happened that deserves closer attention from anyone who runs a team, manages a function, or does knowledge work for a living.

Cursor, the AI-powered coding tool that has quietly become the center of gravity for software development, published a blog post called “The Third Era of AI Software Development.” It describes three phases of how developers work with AI. The first was autocomplete. The second was directing agents through synchronous prompt-and-response loops. The third, arriving now, is something different.

Here’s how Cursor’s CEO Michael Truell described it:

Cursor is no longer primarily about writing code. It’s about helping developers build the factory that creates their software. This factory is made up of fleets of agents that they interact with as teammates: providing initial direction, equipping them with the tools to work independently, and reviewing their work.

Read that again, but replace “developers” with your job title. Replace “software” with whatever your team produces.

Strip away the software context and what remains is a description of management. Of any knowledge work where the output matters more than who physically produced it.

The numbers are already here

This is more than another AI prediction. Cursor shared the data. A year ago, 2.5x more of their users relied on tab-autocomplete than on agents. Today that ratio has flipped. Twice as many users work with agents as use the tab key. Agent usage has grown 15x in a single year. Over a third of the code Cursor’s own team ships is written entirely by agents running autonomously on cloud computers.

And the developers who’ve crossed over share three traits. Agents write almost all of their code. They spend their time decomposing problems, reviewing output, and giving feedback. They run multiple agents in parallel rather than guiding one at a time.

None of those three traits are technical. They describe judgment, decomposition, and quality evaluation at speed. In any other context, we’d call them management skills.

Two factories, one pattern

Jensen Huang talks about AI factories at the infrastructure layer. Energy goes in, tokens come out. The economics of intelligence production. He frames it at planetary scale because he’s building planetary-scale hardware.

Cursor is describing the same factory at the human layer. Challenges go in, solutions come out. Questions go in, opportunities surface that nobody had the bandwidth to explore before. The developer doesn’t write the code. The developer designs the system that produces the code, sets the quality bar, and decides whether the output meets it.

Three weeks before Cursor published, Greg Brockman described how OpenAI itself is reorganizing around this exact shift. His team designated “agents captains” on every team. The goal, by March 31st: for any technical task, the default first move is to interact with an agent, not open an editor. They’re creating living documentation that teaches agents how to work within each project and restructuring their entire codebase to be agent-first.

Three companies at three different scales arrived at the same conclusion. The human role is shifting from doing the work to designing and directing the system that does the work.

Why this matters beyond software

Software is where this is visible first because the tools matured there first. And yes, Cursor has every incentive to frame the shift this way. They’re selling cloud agents. But the framing holds up beyond their product announcement because the pattern is already showing up in places that have nothing to do with code.

Consider what Cursor’s developers actually do now. They take a complex objective and break it into pieces an agent can handle. They define what good looks like before the work starts, then evaluate output against that standard and course-correct when it falls short. They hold the whole picture while agents hold the parts.

That pattern is already emerging outside of software. We work with a firm where the marketing group has quietly become a factory floor. They’re building micro-automations for every repeatable process: content variations, audience segmentation, campaign performance analysis, reporting. No single automation is revolutionary. But taken together, they’ve constructed a system where agents handle the production and humans focus on strategy, quality, and judgment calls. They didn’t set out to “build a factory.” They just kept asking “what can I hand off?” until the answer was: most of it.

The people who thrive in this model aren’t the ones who are best at doing the task. Instead, they’re the ones who can architect the system, evaluate the output, and maintain the judgment that agents can’t supply.

The factory floor changes constantly, which makes this harder than it sounds. The models improve. The tools shift. Workflows that worked six months ago become obsolete or get replaced by better ones. Building the factory isn’t a one-time project. It’s a continuous practice of redesigning how intelligence flows through your work. The durable skills aren’t in any specific tool configuration. They’re in the human capabilities to rebuild when the floor shifts underneath you.

The skill gap nobody is naming

There’s a conversation happening right now in boardrooms and faculty lounges and HR departments about “AI skills.” Most of it focuses on prompt engineering, tool proficiency, learning which buttons to push. That’s the equivalent of teaching someone to type in 1995 and calling it “internet skills.”

The real gap is in the skills Cursor’s developers are using every day, and they look nothing like what most AI training programs teach. Problem decomposition: taking a complex goal and breaking it into components that can be executed independently. Quality criteria definition: knowing what good looks like before the work starts, well enough to evaluate output you didn’t produce. Parallel oversight: holding context across multiple work streams without losing the thread. And system design thinking: understanding how pieces fit together, where failure modes hide, where human judgment adds the most value.

These are human skills. Cognitive and strategic skills that have always mattered in leadership. What’s changed is that they’re becoming table stakes for anyone whose work involves directing intelligence toward outcomes.

The factory question

There’s a fair objection to this entire metaphor: the factory owner doesn’t own the means of production. The models come from Anthropic and OpenAI. The compute comes from NVIDIA’s infrastructure layer. The knowledge worker is renting the machinery, not building it. So where’s the value?

It’s in the orchestration. Factory owners during the Industrial Revolution didn’t build their own steam engines. They bought them from Boulton & Watt. They didn’t generate their own electricity. They bought it from the grid. What made them factory owners was that they designed the production system: the workflow, the quality standards, the integration of machines and people toward a specific output. The value they captured wasn’t in the machinery. It was in knowing how to use it.

Same dynamic here. And as models commoditize, and they are commoditizing, the value shifts further toward the orchestration layer. The person who understands their domain deeply enough to design great agent workflows captures more value as the underlying intelligence gets cheaper and more abundant.

Anthropic’s Dario Amodei offers one way to think about it. He describes AI’s output as a country of geniuses, working around the clock, on every problem you can define. The factory produces the geniuses. Your job is to put them to work on the right things.

Jensen Huang told the audience at Davos in January 2026 that AI helps with tasks, enabling people to fulfill their purpose and become more productive, making workers more valuable. Then he asked a question that should sit with every leader for a long time: “So the question is, what is the purpose of your job?”

Cursor’s blog post offers one answer for developers, arrived at through practice rather than theory. The purpose of the developer’s job is no longer to write code. It’s to build and run the factory that produces software.

The same question is now sitting in front of every professional. If agents can do the task, what’s the purpose of your role? It will almost certainly involve designing systems, defining quality, exercising judgment, and maintaining accountability for outcomes. Agency, in short. The capacity to direct intelligence, not just consume it.

The factories are being built. The infrastructure is live. The shift from “using AI” to “running an intelligence operation” is happening in software right now, and it will reach every function in every organization faster than most leaders expect.

The people preparing for that shift are already thinking less about which tools to learn and more about how to architect the systems those tools will power.

Something Big Is Happening, Now What?

Todd McLees — Wed, 11 Feb 2026 15:10:36 GMT

Matt Shumer’s piece is getting passed around for good reason. Most of it is right. The pace is real. The self-improvement loop is real. The disruption is real.

If you haven’t read it yet, the short version: Shumer is an AI startup founder who wrote a long, personal warning for his non-tech friends and family about how fast AI is advancing. He describes models that now write tens of thousands of lines of working code, test their own output, and show something that feels like judgment. He says AI helped build itself. He says 50% of entry-level white-collar jobs could be eliminated within one to five years. He says the people building this technology are simultaneously more excited and more frightened than anyone else on the planet.

He’s not exaggerating. And I respect him for writing it.

But there’s a gap in the middle of his piece that nobody seems to be talking about. And I think it’s the part that matters most.

The alarm is real. The advice isn’t enough.

Shumer makes a thorough, convincing case that AI can now do your job. Then his advice is: use the tools harder. Spend an hour a day. Be the person in the meeting who did the analysis faster. Sign up for the paid version. Pick the best model. Push it into your work.

That’s necessary. It’s also not enough.

It’s like telling someone the floodwaters are rising and handing them a better bucket.

Here’s the part that stuck with me. He describes the moment these new models showed something that felt like judgment. Like taste. The sense of knowing what the right call is. He flags it as a breakthrough, and he’s right to.

But then he moves on. He doesn’t ask the question that moment demands: if AI is developing something that looks like judgment, what does human judgment actually mean now?

As AI grows in ability, our role in defining direction, values, and purpose only becomes more essential.

That’s the real conversation. Not whether AI can do the work. But what you’re for once it does.

Acceleration without orientation isn’t a strategy.

His advice is all acceleration and no orientation. Get ahead. Move faster. Learn the tools. But get ahead toward what? Without a clear sense of what you’re bringing to the table that the tool isn’t, you’re just becoming a faster follower of the machine.

Shumer tells you to spend an hour a day experimenting with AI and you’ll be ahead of 99% of people. Maybe. But ahead doing what? If the only skill you’re building is “how to use the current tool,” you’re on a treadmill. The tool changes every few months. He says so himself. The models that exist today will be obsolete in a year. So the person who built their entire advantage on knowing which buttons to press is starting over every cycle. That starts to look less like adaptation and more like dependency.

You don’t just “use” AI. That’s where the real skill starts.

Here’s what Shumer skips entirely. There is an entire layer of skill between “try AI” and “AI does your job.” And that layer is where the real work happens.

You don’t just use AI. You design how it fits into your work. You learn where it creates value and where it destroys it. You figure out how much to apply and when to pull back. And you stay clear on what remains yours to do, knowing that line will keep moving. That’s judgment, workflow design, and self-awareness working together. Prompting is maybe 5% of it.

Think about a lawyer. The one who feeds a contract to AI and asks it to find risky clauses is using AI. Fine. But the lawyer who understands which types of analysis the AI handles reliably, which require her own review, how to structure the workflow so nothing falls through, and what her client actually needs from her beyond the analysis itself? She’s doing something fundamentally different. She’s collaborating with AI. And that’s the version of the skill that holds up over time.

Capability isn’t adoption. Institutions move at the speed of trust.

There’s another thing Shumer’s piece treats as a given that isn’t. He writes as if the fact that AI can do something means it will be deployed to do it. It won’t. Not at the pace he implies.

AI can draft a brief. That doesn’t mean a law firm will let it. AI can read a scan. That doesn’t mean a hospital will trust the output without a physician reviewing it. Organizations don’t adopt technology at the speed of capability. They adopt it at the speed of trust. And trust involves compliance, liability, governance, culture, and a dozen other things that move slowly for good reasons.

That trust gap actually strengthens the case for human agency. Someone has to understand the technology well enough to know where it’s genuinely ready and where it isn’t. Someone has to earn the institutional credibility to deploy it responsibly. And someone has to own the outcome. The more capable AI gets, the more those roles matter, not less.

Shumer is right that the capability curve is steep. But capability without trust just sits on a shelf. The people who will matter most in the next few years are the ones who can help their organizations move from “AI can do this” to “we trust AI to do this, and here’s how we’re accountable for it.”

What do you bring when the machine can do the work?

That’s the only question that matters when intelligence is abundant.

And it’s not one you answer once. The line between what AI handles and what remains human will keep shifting. It shifted for software engineers over the past year. It’s shifting for lawyers, analysts, writers, and consultants right now. It will keep shifting.

But the work of knowing where that line is today, in your context, for your clients, for your organization? That is the work. Not something you get to after you learn the tools. The thing itself.

Where do you direct AI? What do you hold onto? What judgment and values do you bring that give the output meaning? How do you design the collaboration so both the human and the machine are doing what they’re best at? These are strategic questions, and right now most people aren’t asking them because they’re still stuck on “how do I use this thing.”

The honest picture.

Shumer’s piece is valuable because it tells people the truth about how fast this is moving. I want to add to that truth, not argue with it.

Something big is happening. It’s moving faster than most people realize. You need to engage with these tools seriously and soon.

And learning to use AI is the beginning, not the destination. The skill that matters is understanding how it fits into the context of your work. The judgment that matters is knowing how much to apply and when. What will differentiate you over time isn’t how fast you move. It’s whether you know what’s yours to do.

Shumer wrote the alarm. It’s a good one. But alarms don’t tell you where to go. You have to find that within yourself. It’s part of building what is scarce and valuable when intelligence is abundant and commoditized.

When Your AI Co-Worker Starts Improvising

Todd McLees — Tue, 03 Feb 2026 16:03:06 GMT

Here’s something that should matter to everyone building workflows with AI: the more you let AI “think,” the less consistent it becomes.

Anthropic just published research showing that extended reasoning doesn’t make AI more reliable. It makes it more unpredictable. Not dangerous in a sci-fi way, but unpredictable in a “wait, why did it give me three different answers to the same question?” way.

They measured AI mistakes in two categories:

Bias = the model consistently gets something wrong the same way every time
Variance = the model’s answers bounce around randomly between attempts

When variance dominates (when the AI becomes what they call “incoherent”), you’re dealing with something that might give you brilliant work on Monday and nonsense on Tuesday for reasons no one can explain.

The study found a troubling pattern across their tests: more reasoning steps lead to more randomness, longer agent actions reduce consistency, and complex tasks cause even the smartest models to wobble. This appears to be a feature of how these systems scale, not a bug in today’s AI.

What This Actually Means

If you’re building AI into your curriculum, planning organizational rollout, or integrating AI into daily workflows, this changes your risk model entirely.

The old fear was that AI pursues the wrong goal really well. The new reality is that AI pursues different goals each time you run it.

For educators: You can’t just check if an AI tool gives good answers. You need to know if it gives consistent answers when students use it repeatedly.

For workflow designers: That AI agent that worked beautifully in testing might behave completely differently in production, not because it learned something new, but because extended reasoning introduces drift.

For anyone implementing AI: “Let the AI think longer” isn’t automatically better. Sometimes it’s just more expensive chaos.

This is why human judgment remains essential in AI collaboration.

We’re here, in part, to catch AI being inconsistent, to notice when the pattern changes, when the reasoning wobbles, when Tuesday’s answer contradicts Monday’s for no good reason.

That’s a fundamentally different skill than “AI oversight.” It’s closer to editorial judgment, quality control, or what we used to call professional instinct. It’s recognizing when something’s off.

If you’re assessing whether your organization is ready for AI, the question shifts: instead of “Can we trust the AI?” ask “Can we catch when the AI becomes unreliable?”

Reliability isn’t a feature you can count on. It’s something you have to monitor, especially as AI systems scale up their reasoning capacity.

We're learning to catch AI when it starts improvising, when the variance overtakes the bias, when longer thinking produces less consistent output, when "smarter" paradoxically means "less predictable."

That’s a reliability problem in the practical sense. And reliability is something humans are actually quite good at spotting, if we’re paying attention, if we understand what we’re looking for, and if we haven’t handed over our judgment to the very systems we’re supposed to be collaborating with.

As AI gets more capable at reasoning, human pattern recognition becomes the quality control layer that keeps the whole system honest.

AI Automation and Augmentation Are Happening to the Same Jobs

Todd McLees — Mon, 02 Feb 2026 17:15:22 GMT

The Burning Glass Institute just released a report that caught my attention. They analyzed millions of job postings from before and after ChatGPT launched, looking for evidence of how AI is actually changing work. Not predictions. Not projections. Real data about what employers are asking for.

What they found surprised me. And it should change how we talk about this.

The Finding That Matters

You’ve heard the stories about AI and jobs. AI will automate work away and we’re all in trouble. AI will make us more productive and everything will be fine. AI will eliminate some jobs but create new ones we can’t predict yet.

The Burning Glass research suggests all of these stories miss what’s actually happening right now, in existing roles.

Automation and augmentation aren’t happening to different jobs. They’re happening to the same jobs. At the same time.

Think about what that means. Project managers aren’t disappearing. But the skills employers want from project managers are shifting. Demand for spreadsheet and presentation work is softening. Demand for budgeting, scheduling, and strategic thinking is growing. The job title stays. The work inside it changes.

Same pattern with sales roles. The administrative and operations tasks are fading. Relationship management and prospecting matter more than before.

Look at this chart. Each dot is a job. The further right a dot sits, the more of that job’s tasks AI can take over. The higher up a dot sits, the more AI can amplify what’s left.

If some jobs were “automation jobs” and others were “augmentation jobs,” you’d see dots spread across the chart. Some jobs in the bottom-right getting automated. Other jobs in the top-left getting enhanced. Different jobs in different places.

That’s not what the data shows. The dots cluster along a diagonal. The jobs most exposed to automation are the same jobs most exposed to augmentation. Burning Glass calls these “force multiplier” roles: AI both replaces some tasks and amplifies others, in the same job.

The researchers also tracked which skills are growing and shrinking in job postings since ChatGPT launched.

Skills that AI can replace were more likely to see demand drop. Skills that AI can amplify were more likely to see demand grow.

Looking at the left side of the chart: 25% of automation-exposed skills saw declining demand, compared to 21% of other skills. That 4-point gap means automation-exposed skills were about 16% more likely to decline. On the right side: 61% of augmentation-exposed skills saw growing demand, compared to 57% of other skills. That 4-point gap means augmentation-exposed skills were about 7% more likely to grow.

The gaps aren’t enormous. But the direction is consistent across occupations, and it showed up in just three years of data.

Here’s how the report frames it: the question isn’t “which jobs will survive?” The question is “how will this job change, and what new skills will workers need to perform it?”

Why This Reframe Matters

I think this finding is important because it shifts where the control sits.

If AI is coming to take your job, you’re stuck waiting to see what happens. You’re at the mercy of forces beyond your control. Maybe you update your resume. Maybe you try to guess which occupations will be safe. But fundamentally, you’re reacting.

If AI is reshaping the task mix inside your current role, that’s different. You can participate in that reshaping. You can pay attention to which parts of your work are becoming automated and which parts are becoming more valuable. You can have a say in what your job becomes.

One framing makes you passive. The other gives you something to do.

What the Research Doesn’t Tell Us

The Burning Glass report does offer guidance for workers, employers, and policymakers. They advise workers to track changes in their field, acquire skills where AI is creating demand, and double down on capabilities requiring judgment and context. They tell employers to think about role redesign, not just headcount reduction. They warn policymakers that the challenge isn’t relocating workers to “safe” occupations but helping them adapt as their current jobs transform.

This is solid advice. But I think there’s a layer underneath it that deserves more attention.

The report tells workers to “track the changes in your field” and “pay attention to how your own job is evolving.” That’s the right instinct. But how do you actually do that?

Here’s one way to start. Take your own job description, or a posting for a role you want, and use an AI to help you break down the task mix. Which of these tasks are becoming easier to automate? Which require more human judgment than before? Where is the value shifting?

You can do the same thing with your own weekly work. What did you spend time on? Which tasks could you hand off to AI right now? Which tasks require your specific knowledge, relationships, or judgment?

This kind of self-audit gives you a clearer picture of where you stand. And once you see the shape of the change, you can start building a runway for yourself. What skills do you need to develop? What parts of your work should you lean into? What relationships or knowledge give you an advantage that AI doesn’t have?

The same exercise works if you’re responsible for other people. Helping your team or your students see their own task mix clearly is the first step toward helping them prepare for what’s coming.

The Role of Human Agency

I believe there’s a variable that will determine who navigates this transition well and who doesn’t.

That variable is human agency.

Not AI skills. Agency. The capacity to direct intelligence that is not your own toward outcomes you have chosen.

Learning to use AI tools is straightforward. You can pick up the basics in an afternoon. But knowing which tool to reach for, and when, and why? Knowing what you’re trying to accomplish before you start prompting? Being able to tell when the AI output is good enough and when it’s wrong?

That’s harder. And that’s what separates people who use AI productively from people who just use AI.

I think of it as three related capabilities:

Intention. Knowing what you’re trying to accomplish before you reach for any tool. This sounds obvious, but it’s easy to skip. The clearer you are about your purpose, the better you can direct AI toward it.

Self-determination. Shaping how AI gets integrated into your work rather than accepting whatever default shows up. Someone is going to decide how AI fits into your job. It might as well be you.

Judgment. Discerning when AI output serves your purpose and when it doesn’t. AI is confident whether it’s right or wrong. You need to be the one who knows the difference.

These capabilities matter more than any specific tool or prompt technique. Tools change. The ability to direct them thoughtfully doesn’t go out of date.

What You Can Do

Let me get specific about what this means for different people.

If You’re a Professional

The question to ask yourself isn’t “Is my job safe?” That framing puts you in a passive position, waiting to find out your fate.

Ask instead: “How is my job changing?”

Look at your own work. Some of your tasks are probably getting easier. Faster to complete. Less dependent on your specific knowledge or judgment. Those are the parts that AI can take over, and in many cases, probably should.

Other parts of your work require more from you than ever. More context. More relationship. More thinking through tradeoffs that don’t have obvious answers. Those are the parts to pay attention to. That’s where your value is growing.

Your goal isn’t to protect your current set of tasks. Your goal is to participate in shaping what your role becomes. If you’re passive while your job changes around you, you might end up in a version of your role that you didn’t choose and don’t want.

One practical step: start noticing when you reach for AI and why. Are you using it to skip thinking, or to think better? There’s a difference. The first habit makes you dependent. The second makes you more capable.

If You Lead a Team or Organization

The obvious question is “Which roles can we automate?” I’d encourage you to ask a different one: “How should these roles change, and who’s going to design that?”

Role redesign is a design problem. It’s not just about efficiency. It’s about figuring out how humans and AI work together in a way that makes sense for your context.

The Burning Glass report flags something worth taking seriously. If you automate the entry-level tasks that traditionally built expertise, you create a pipeline problem. Junior employees learn by doing foundational work. That’s how they become senior employees. If AI handles all the learning-by-doing tasks, where do your future experts come from?

This isn’t a reason to avoid AI. It’s a reason to think carefully about how you introduce it.

One more thing: involve your people in the redesign. The workers doing the jobs understand which tasks benefit from human judgment and which don’t. They see nuance you won’t see from a distance. And if you design new workflows without their input, you’ll have a harder time getting their commitment to making those workflows succeed.

If You’re an Educator

The jobs our students are preparing for won’t look the same by the time they graduate. Not because the jobs will vanish, but because the task mix inside those jobs is already shifting. The Burning Glass data shows we can measure this shift now, just three years after ChatGPT launched.

Content knowledge still matters. But it’s not enough on its own.

We need to build human skills alongside the content. The Burning Glass report shows that skills requiring judgment, context, and relationships are seeing increased demand. Communication. Collaboration. Critical thinking. Ethical reasoning.

These have always mattered. Now we have data showing they matter more, not less, as AI becomes more capable.

But here’s the challenge: these are suitcase terms. They sound clear until you try to teach them. What does “critical thinking” actually mean in practice? What specific behaviors make up “collaboration”? These words carry a lot of cargo, and if we don’t unpack them, we end up with vague learning objectives that are hard to measure and harder to develop.

This is a time that calls for precision. If we’re going to help students build human skills, we need to get specific about what those skills actually are and how they show up in real work.

We also need to build human agency. Students need the ability to direct intelligence that isn’t their own. That’s what productive collaboration with AI actually requires.

This means helping students develop intention, so they can clarify what they’re trying to accomplish before they reach for any tool. It means building self-determination, so they shape how AI gets used in their work rather than accepting whatever default appears. And it means developing judgment, so they can tell when AI output serves their purpose and when it falls short.

We’re not preparing students for a destination that will stay fixed. We’re building their capacity to work with tools and systems that will keep evolving. The curriculum needs to reflect that. Not by dropping content, but by adding the human capabilities that make content knowledge useful when AI handles more of the routine work.

Where This Leaves Us

The Burning Glass report gives us something useful. Early evidence that AI’s impact on work is real, measurable, and more complicated than the simple stories we’ve been telling.

Jobs aren’t vanishing wholesale. They’re changing from the inside. Some tasks are getting automated. Others are getting amplified. Often in the same role, at the same time.

The question isn’t whether your job will exist in five years. The question is what your job will look like, and whether you’ll have a hand in shaping it.

That’s where human agency comes in. The ability to set your own intentions, to determine how you work with AI rather than having it determined for you, to apply judgment that the machine doesn’t have.

Building that capacity is the work in front of us. In ourselves, in our teams, in our students.

The transformation is already underway. The Burning Glass data confirms it. What we do next is up to us.

About the Author:

Todd McLees is co-Founder of humanskills.ai

External Links: Full Report:

Read the full Burning Glass Institute report: "Beyond the Binary: How Automation and Augmentation Are Combining to Reshape Work"