The AI paradox: More automation, more humans, more work

ELI5/TLDR

Dan Shipper runs a small, very AI-heavy company. His big claim: even as the AI gets better at doing work, humans end up with more work, not less, because someone always has to babysit the machine. He thinks the people who win the next year are product managers and designers who learn to build things themselves, and he is weirdly bullish on boring business software surviving the whole thing. The whole conversation is a set of bets he wants graded a year from now.

The Full Story

Dan Shipper is the founder of Every, a company of about thirty people where everyone — sales, writers, customer service, not just engineers — drives AI coding tools all day. A year ago on this same podcast he said people were sleeping on Claude Code for non-engineering work, and that turned out to be right. So Lenny invited him back to make more bets. The format is explicit: three buckets of predictions, to be scored in May 2027.

Two shapes of work

Shipper thinks daily work splits into two modes. First, everyone gets at least one agent they can hand tasks to, probably living in Slack. Second — and the one he is more excited about — most of your actual work starts happening inside a coding agent like Codex or Claude’s “co-work,” which becomes a kind of operating system for everything: email, documents, research.

“Most of the work that you do is actually going to happen on your computer in an environment like Codex or Cloud Co-work.”

On the first mode, he changed his mind in public. He used to think everyone would have a personal agent — a little reflection of you, like the soul-companion daemons in The Golden Compass. Then reality hit. Personal agents break constantly and need babysitting, and most people won’t put in that effort. So companies are landing on one big “super agent” for the whole company (Shopify and Ramp both have one), maintained by a dedicated person. The underlying rule he keeps repeating: an AI agent is only useful right now if a specific human cares about it and keeps watching it. Sever that connection and the agent goes dead.

Put the browser inside the agent, not the other way round

Here’s the genuinely surprising bit. For years the assumption was that the future is AI baked into your web apps — a chatbot in the corner of every product. Shipper says it’s flipping. Instead of putting AI into your browser, you put a browser inside your AI agent, so the agent can see whatever you’re looking at and act on it.

He works in Codex with an in-app browser open on his document, and the agent watches him write and can jump in. Think of it like a colleague sitting next to you who can also see your screen and grab the keyboard. He says he’s been at inbox-zero for ten straight days because the agent gathers his emails, renders them as a page, and he just talks at each one.

The SaaS apocalypse is “dumb”

A second-order effect: if your software runs inside the user’s agent, the user brings their own AI tokens. The software company doesn’t pay for the compute. This, Shipper argues, saves the margins of business-software companies rather than killing them.

“I would buy SaaS stocks right now. Um I would I think the SaaS apocalypse is done… What agents do is increase the number of users of SaaS. Not get rid of it.”

The logic: agents are just more users, hammering your product at high volume. The job of a software company shifts from “bolt an AI chatbot on” to “make a product that humans and agents want to collaborate on together.” Simpler products, because the agent handles formatting and tables; but new demands too — approval inboxes, logs, one-click rollback, because an agent can make a thousand changes in three seconds. (This, he notes, is why GitHub is straining: the traffic spike is mostly people’s agents.)

He also declares the command-line moment over. Claude Code’s early popularity got misread as “the magic is the terminal.” Shipper says no — “we made GUIs for a reason” — and most of his technical team has already moved off the raw terminal back into graphical tools.

Automation is a lie

The emotional core of the episode. Every doubled headcount in a year — not what you’d expect from an AI-forward company. Shipper’s explanation: every time you automate something, you need a human on top making sure the automation works. He compares the AI-era worker to a manager, and points out managers aren’t lounging on a beach — they’re constantly checking in. More automation, and he personally works more.

“We have so much automation, so much AI, and I also work way more.”

Why do benchmarks make AI look more independent than it is? Because, he argues, benchmarks only measure problems we’ve already framed and can score. He built his own “senior engineer benchmark” by vibe-coding an app called Proof that kept crashing on launch (he gave himself “vibe coder elbow” from bursitis). He had two real senior engineers rewrite it properly, then scores each new model against those rewrites. Models scored ~30 out of 100 until GPT-5.5 jumped to ~62 — clearly heading toward senior-engineer level.

But here’s the catch. Given a list of bugs, every model dutifully tries to fix the bugs. A real senior engineer looks at the codebase and says, this is garbage, we need to rewrite it, I know you don’t want to hear that. The act of reframing the problem — deciding what the real task is — is human work that can’t be scored until someone writes it down. So benchmarks can saturate without senior engineers being replaced.

New jobs, fuzzy jobs

A side effect of everyone being able to do everything: people are confused about what their job even is. Engineers can design, PMs can ship code, marketers can build pages. Shipper thinks this settles down — marketing people still do marketing, they just touch the website now.

A genuinely new role is emerging: the “forward deployed engineer,” whose whole job is making sure the company’s agents work. One of his engineers, Nitesh, spends most of his day in Slack arguing with an agent called Claudie that runs their consulting practice. The framing he prefers over “babysitting”: you’re building a system so that less-technical people can do formerly-technical work without screwing up.

He’s also bullish on AI-written internal documents — plans, strategy memos, most email. He says we already happily read AI-written engineering plan docs, and the aversion to AI prose for internal work is silly because most humans write bad strategy docs anyway. His one rule: you must stand behind every line. The tell for slop is “it took them less time to make it than it takes me to read it.”

Who wins: PMs and full-stack designers

The payoff bucket. Shipper is “super super bullish” on product managers. His example is Marcus, a PM by training, only “lightly technical,” who took a year off to get deeply fluent with coding tools. Now he ships faster than almost anyone, pairing modest technical knowledge with sharp product and user sense — a hire that would have been impossible a year ago.

Equally on full-stack designers. Designers have always been frustrated handing beautiful interactions to engineers who don’t build them right. Now they make the pull requests themselves. And because everyone’s using the same models, default AI output all looks the same — slop — so a designer’s eye for making something different becomes more valuable, not less.

“What models do in general is they make yesterday’s human competence cheap… it becomes commoditized. It’s not valuable anymore. What humans do is we go in there and we’re like, yeah, how do I use this to make something new and interesting?”

That line is his whole theory of why mass unemployment won’t happen. Models freeze and cheapen yesterday’s skill; humans push to the next edge; that new edge eventually gets absorbed into the models, which opens another edge. The advice for not getting laid off is two words he likes: ride the models. Try every new model on whatever you do, stay curious, keep “turning over the rock” to see if it can do the thing yet.

He makes a nice point that the edge of AI isn’t San Francisco — the people there build the models but don’t know all the ways to use them. The edge is wherever AI meets a real person doing real work. And the remarkable thing is that the best model in the world is available to anyone with a little money the day it ships; if IBM had invented AI it would cost a fortune and only big companies would have it.

Key Takeaways

An AI agent is only useful today if a specific human keeps watching it; sever that connection and it goes dead. This is why companies are converging on one maintained “super agent” rather than a personal agent per employee.
The work surface is flipping: instead of AI inside your apps, a browser goes inside your AI agent, so it sees your screen and acts alongside you.
When SaaS runs inside a user’s agent, the user supplies the AI tokens — the software vendor doesn’t pay for compute, which protects vendor margins.
Agents are just more users. Shipper expects a demand spike for SaaS, not its death — hence “buy SaaS stocks.”
New UX demands for agent-era software: approval inboxes, change logs, fast rollback, because an agent can make thousands of edits in seconds.
Benchmarks only measure problems humans have already framed and scored; the human work of reframing the task (deciding the real job) is invisible to benchmarks, so they can saturate without replacing senior people.
GPT-5.5 was the first model to confidently rip out and rewrite bad code from scratch; earlier models patch around the edges even when told not to.
“Automation is a lie” — every automation needs a human on top, so an AI-forward company can grow headcount, not shrink it.
New role: the “forward deployed engineer,” whose job is keeping the company’s agents working.
Biggest winners: AI-fluent product managers and full-stack designers who now ship their own work directly.
Default AI output all looks the same (everyone uses the same models), so human taste and originality become more valuable, not less.
The survival move is “ride the models”: test every new release on your own work, stay playful, keep checking what’s newly possible.

Claude’s Take

This is a sharp, useful episode wrapped in a fair amount of insider San Francisco vocabulary. The two load-bearing ideas — “automation needs a human on top, so headcount can grow” and “benchmarks only measure pre-framed problems” — are genuinely good and worth carrying around. The benchmark point in particular is the kind of thing that quietly reframes a lot of doom-versus-utopia AI noise.

Worth keeping the BS filter on, though. Shipper has obvious incentives: he sells AI consulting and runs AI products, so “everything is changing AND your job is safe AND buy the software stocks” is a conveniently comforting message for the people who pay him. His sample size is one unusually self-selected thirty-person company of AI obsessives in Brooklyn — he says himself they’re a “little pocket of the future,” which is another way of saying they’re not representative. The “no job apocalypse” call may well be right, but it’s a bet from someone who’d struggle to say otherwise. And the specific predictions (Slack super-agents, CLIs dead, SaaS stocks up) are time-stamped to be graded in a year, which is admirably honest but also a reminder that these are guesses.

A 7. The mental models are above average and the honesty about scoring is refreshing, but it’s a vendor-adjacent founder making confident calls on a one-company sample, padded with sponsor reads and a long book-recommendation tangent at the end.

The AI paradox: More automation, more humans, more work | Dan Shipper