A Project Manager Walks Into a Terminal

Wednesday morning, 8:47am. My IDE is still dormant from yesterday. The only two things open on my screen are a Claude Code terminal and a GitHub tab — three pull requests waiting for review, a kanban board of tickets I briefed yesterday, and a Zoom transcript I'm about to chop into user stories for the agents to pick up after lunch.

I work the transcript for twenty minutes. Requirements, acceptance criteria, a couple of bugs I noticed the client hasn't logged yet. Each becomes a ticket. Each ticket gets a label, a priority, the context an agent will need to do the work without asking me again. Handoff goes out at 9:14am.

Then the PR reviews. Two are fine — clean diffs, tests green, the kind of output that used to take a junior engineer a week and now takes an agent an hour. The third has a subtle regression in how it handles a timezone edge case. The agent didn't catch it. I only catch it because I've shipped that kind of bug myself, in production, at 2am, in 2009. I leave a review comment. Back to the queue.

If this sounds like a development workflow, you've missed the point. This is project management. I am the PM, the QA lead, the escalation point, and the stakeholder liaison. The agents are the development team. I still have a Cursor subscription. I rarely open it anymore.

And here's the part that should give every knowledge worker pause: it's not just developers.

It's Not Just Developers

Walk into any Morgan Stanley branch and the same inversion is happening, in a different accent. The firm deployed an OpenAI-based assistant to 16,000 financial advisors — document access rates jumped from 20% to 80% almost overnight. The job didn't disappear. It moved up a layer. Advisors now review AI-generated meeting notes rather than write them; they apply firm knowledge rather than hunt for it. In 2024 Morgan Stanley rolled out a second tool, "AI @ Morgan Stanley Debrief", which drafts post-meeting emails and files them to Salesforce automatically. The advisor's role contracted to the thing machines can't do yet: judgment, relationship, recommendation.

Same pattern inside the Big Four. A junior analyst's first three years used to be bent over decks and data rooms. Now they review AI-built decks and validate AI-analysed data rooms. Deloitte's own research estimates executives expect 10-30% productivity gains from AI. What they don't always notice is that the gain comes by compressing the work of five people into the review queue of two.

Duolingo's translators went through the same door. In May 2025 the company announced it would phase out contractor translation work that AI could handle. The full-time "Duos" stayed — but their job changed from producing translations to supervising them. New title, same business card. No raise.

Singapore's version plays out in DBS branches. A Straits Times feature profiled a branch banker in her 60s who uses ChatGPT to answer customer product queries. She isn't answering from memory. She is supervising the machine's output and adding the human layer that keeps it honest.

The interface differs — Claude Code, Copilot, ChatGPT Enterprise, Morgan Stanley's internal assistant, DBS's CSO Assistant — but the role shift is identical. IBM names it plainly: curation over creation, direction over execution. The uniformity of the pattern across industries isn't a coincidence. There's a structural reason knowledge work keeps bending the same way.

The Scarcity Just Flipped

Ethan Mollick at Wharton wrote a piece in January 2026 called "Management as AI superpower" that I've read four times. The core move is compact enough to fit on a napkin:

Now the "talent" is abundant and cheap. What's scarce is knowing what to ask for.

The classic management scarcity equation just inverted. For a hundred years, executives were rare because talent was expensive to find, develop, and direct. The management tax — all that coordination and review and quality control — paid for itself. But if the execution layer is now a cheap, fast, competent agent that can be spun up on demand, the bottleneck is no longer hiring and doing. It's specifying, reviewing, and deciding. That was always management. It has now been distributed to everyone.

Erik Brynjolfsson's Fortune 500 call-centre study in the Quarterly Journal of Economics makes the mechanism concrete: novice agents with AI performed as well as agents with six months more experience without it. AI didn't raise the ceiling — it raised the floor. Once the floor is high enough, the differentiator shifts from execution capacity to evaluative capacity. Which is what managers have always done.

BCG put a number on this one. In their 2026 workforce report, only 10% of AI's value comes from algorithms, 20% from technology, and 70% from rethinking the people component. The bottleneck moved from compute to coordination. We solved the easy part and are now stuck on the hard one.

Andrej Karpathy, the former OpenAI co-founder, told Fortune in March 2026 that he hasn't typed a line of code since December. A few weeks earlier, in a February 2026 post on X, he'd given his new default mode a name — agentic engineering — to describe the shift to orchestrating agents and acting as oversight. Jensen Huang put the institutional version on the CES stage in January 2025:

In a lot of ways, the IT department of every company is going to be the HR department of AI agents in the future.

Read that line twice. IT — the classic infrastructure function — is being reframed as HR for digital workers. At organisational scale, you'll need to hire, onboard, evaluate, and occasionally fire your agents. At individual scale, that's exactly what each of us is already doing at our desks.

I wrote a few weeks ago about Block's 60/40 split — 60% logistics, 40% judgment — as an organisational architecture for AI-era work. What I didn't say loudly enough is that the split isn't just organisational. It's also personal. Your own workday is quietly dividing along the same line.

Management used to be an unusual role for unusual people. It's now the default posture of anyone whose work is knowable.

The Bill Nobody's Paying

The promotion is already happening. The onboarding is not.

Deloitte's 2025 human capital research found that 82% of employees received no generative AI training, even as their executives forecast productivity gains of 10-30%. BCG found 89% of leaders say AI skills are a critical gap, and only 6% are doing anything meaningful about it. The awareness is universal. The action is rare. Everyone got promoted. Nobody got the manual.

What does that look like in practice? Harvard Business Review's researchers gave the phenomenon a name last year: workslop. In their September 2025 piece, the BetterUp Labs and Stanford Social Media Lab team reported that 40% of workers received AI-generated low-quality content in the prior month, and each incident took an average of 1 hour 56 minutes to resolve. For a 10,000-person firm, that translates to an estimated $9 million a year in lost productivity — the invisible tax of shipping output you haven't actually reviewed. And it's worse than wasted time: 42% of recipients trusted the sender less afterward. Workslop doesn't just burn hours. It corrodes working relationships.

Zoom out and the enterprise numbers are uglier. MIT's NANDA initiative reported in August 2025 that 95% of enterprise GenAI pilots showed zero measurable P&L impact. The failure isn't the model. The model is fine. The failure is the absence of workflow redesign, oversight discipline, and the management skills to evaluate what an agent just handed you.

Klarna is the canonical cautionary tale. The fintech famously replaced the work of 700+ customer service agents with an AI that handled 2.3 million conversations in 35+ languages. The volume numbers were glorious. By May 2025 the company was quietly rehiring. CEO Sebastian Siemiatkowski publicly acknowledged that cost had become too dominant a factor when the company organised its AI rollout — and that what you end up with, when cost crowds out everything else, is lower quality. Investing in human support, he told reporters, was "the way of the future" for Klarna.

Cost was measurable. Quality was not — until it became unrecoverable. This is the workslop story at enterprise scale.

And there's the failure mode that ends up in court. In Mata v. Avianca, New York attorneys filed a ChatGPT-generated brief citing six cases that didn't exist; when one of them got nervous he asked ChatGPT if the cases were real, and the model reassured him they were. The firm was sanctioned $5,000. By October 2025, legal researchers had catalogued over 1,000 similar AI-hallucination cases across US courts. It's easy to mock. The mechanism is universal: the manager stopped checking the work. He trusted the output instead of supervising it.

The training gap is fixable. The deeper problem is not.

The Judgment Paradox

Here is the thing I don't know how to solve.

David Duncan, writing in HBR this February, framed it more precisely than I've heard anywhere else. AI now handles the messy, repetitive tasks that once built judgment. Junior employees miss the chances to develop it. Organisations risk ending up with managers who've never done the underlying work. Good AI oversight requires domain expertise. Domain expertise, historically, was built by doing the work AI now does.

This is not a hypothetical. SignalFire's 2025 data shows new graduates' share of tech hiring fell to under 6% in 2025 — less than half of pre-pandemic levels. Big Four graduate intake is down double digits. The on-ramp that used to build the reviewers is being dismantled at exactly the moment review becomes the only differentiated human skill.

Inside Anthropic — the company that makes the model I've been briefing all morning — engineers are already describing this as the paradox of supervision. In the company's December 2025 research on how AI has transformed its own engineers' work, one engineer frames the worry bluntly: the real problem isn't the shape of their own skill set, it's oversight. Skills that atrophy or never develop, they note, become a problem not in the abstract but specifically for your ability to safely use AI. The oversight job gets harder as the hands-on job you used to do gets rarer.

Effective AI use requires supervising AI. Supervising AI requires the skills that atrophy when you stop doing the work. Carnegie Mellon and Microsoft Research put a finer point on it in their CHI 2025 paper: higher confidence in AI was directly associated with less critical thinking. The workers who trusted the model most applied the least scrutiny. That's the falling-asleep-at-the-wheel effect, formalised in data.

I was lucky. I came up through a generation that racked servers in London Docklands at 3am, wrote shell scripts by hand, SSH'd into production when nobody else was awake to help. The judgment I deploy against Claude's pull requests was paid for over twenty years of shipping exactly the kind of code Claude now produces in twenty minutes. I'm not sure where the next generation of reviewers comes from. And I'm not sure anyone is.

The paradox is real. But the fatalism it invites isn't.

What Actually Works

Singapore's three banks — DBS, OCBC, UOB — are jointly retraining 35,000 bankers on AI fluency over the next two years. Not a PR campaign. A structural response to the universal promotion, funded at national scale.

DBS alone reports S$1 billion in audited economic value from AI in 2025 — from 1,500+ models across 370+ use cases. The playbook is not mysterious. It was articulated by Tan Su Shan, who became CEO in April 2025:

Let them own the model. Let them own the feedback loop. Let them own the outcomes.

That's not a motivational quote. It's the thesis, operationalised: distribute AI ownership instead of deploying AI from the top. Piyush Gupta, her predecessor, is worth naming honestly too. From Fortune Asia earlier in 2025:

In my 15 years of being a CEO, for the first time, I'm struggling to create jobs.

The playbook isn't painless. It is specific. For individuals, three moves:

• Manage your own AI use deliberately. Gallup's data is unambiguous: employees whose managers actively support AI are roughly 7x more likely to say AI helps them do their best work. If your manager won't coach you, coach yourself. Pick one workflow — PR reviews, meeting notes, research briefs. Study it for a month. Measure what actually changed.
• Invest in what AI can't give you. Domain expertise. Taste. Judgment under ambiguity. The ability to tell when an AI-generated plan is subtly wrong. These are the expensive skills now. They compound. Your five-year-old bug scars are worth more than they've ever been.
• Refuse workslop. The hardest discipline in a manager-mode world isn't producing more — it's refusing to ship output you haven't actually reviewed. HBR's workslop data found more than half of recipients admitted to sending workslop themselves. Don't be that person. It's the cheapest way to torch your reputation in 2026.

And then there's the Shopify framing, which deserves honest engagement rather than dismissal. Tobi Lütke's April 2025 memo — reflexive AI usage is now a baseline expectation — is provocatively framed. It also happens to be directionally correct. If you can't articulate what AI cannot do in your specific role, you haven't thought hard enough about your role. That's not a performance mandate. It's a diagnostic question. The answer tells you where your next five years of salary come from.

Back at My Desk

Back to Wednesday morning. Claude Code still open. GitHub still open. Three PRs reviewed, two merged, one kicked back with comments. The kanban board has moved. Lunch soon.

The regression I caught in PR #2 — the timezone edge case that would have surfaced in production in Singapore at 2am on the first Sunday after daylight-saving shifted in New York — I only caught because I've shipped that kind of bug myself, at exactly that hour, in a different decade. The judgment came from the work I no longer do.

The skills that make my current workflow any good — crisp requirements, clear acceptance criteria, thoughtful PR comments, knowing when to overrule the agent, knowing when to ship — are not developer skills. They were always management skills. I just acquired them by accident, because the toolchain quietly rotated underneath me.

I still have a Cursor subscription. I rarely open it. The tool I pay for isn't necessarily the tool I use. That sentence describes my IDE. It also, if I'm being honest, describes my job title. I am paid to write code. I mostly don't anymore. I am a project manager for agents, whether my business card admits it or not.

Mollick ends his piece with a question I can't improve on:

I don't know exactly what work looks like when everyone is a manager with an army of tireless agents.

Neither do I. But I know my Wednesday morning looks like the first draft of the answer. And yours probably does too.

A little earlier in the same piece, Mollick writes a line worth keeping on a sticky note: the skills that are so often dismissed as "soft" turned out to be the hard ones. The soft skills were never soft. They were just under-priced. They are not under-priced anymore.