Back

Dan Shipper Says AI Automation Creates More Human Allocation Work

Automation is a lie. Every time you automate something, in order to make sure the automation is working well, you need a human on top of it.

Watch the recap video here

Lenny's PodcastLenny's PodcastsourceDan ShipperDan ShipperspeakerWatch full clip

Context

Automation shifts scarcity toward human judgment, supervision, trust, product coherence, and the people who decide what agent work should exist.

Big Ideas

  • The scarce role is moving from "do the task" to "allocate and govern the task." Shipper's version of the AI future needs people who can maintain agents, decide what work matters, review higher volumes of output, and keep products coherent as automated capacity rises.
  • SaaS may gain leverage rather than lose it if agent workspaces become the user's operating system. In Shipper's model, apps still matter, but they need to support humans and user-provided agents collaborating on the same object, which changes margins, UX, rate limits, logs, approvals, and pricing.
  • PMs and full-stack designers may become more valuable because AI turns taste and problem selection into directly executable capacity. The model commoditizes yesterday's implementation skill, but the human still has to decide what is worth making and why it should look or behave differently from default model output.

Supporting Context And Sources

  • Lenny's official episode page frames the episode around the same twelve predictions surfaced in the transcript: Codex or Claude Code as the future work surface, one Slack super-agent per company, SaaS resilience, PM and designer upside, forward deployed engineers, CLIs fading, and "automation is a lie." Direct link: https://www.lennysnewsletter.com/p/the-ai-paradox-dan-shipper
  • Apple Podcasts' episode listing repeats the official summary and chapter structure, including the automation paradox at 39:01, the benchmark discussion, the "ride the models" advice, and Lenny's disclosure that he may be an investor in companies discussed. Direct link: https://podcasts.apple.com/us/podcast/the-ai-paradox-more-automation-more-humans-more-work/id1627920305?i=1000769334366
  • Shipper's own Every essay "After Automation" is the strongest corroborating companion piece: it argues that Every automated heavily with Codex and Claude Code while growing from a small team to almost 30 people, because agents still create human work around management, review, and differentiation. Direct link: https://every.to/chain-of-thought/after-automation
  • Shipper's earlier Every essay "The Knowledge Economy Is Over. Welcome to the Allocation Economy" gives the conceptual backstory for the episode's manager analogy: AI changes makers into managers who allocate machine capability rather than personally performing every task. Direct link: https://every.to/chain-of-thought/the-knowledge-economy-is-over-welcome-to-the-allocation-economy
  • Podwise's episode notes interpret the interview as a shift from replacement toward orchestration: agents handle rote work while humans focus on system architecture, creative direction, and higher-level decision-making. Direct link: https://podwise.ai/episodes/8063100
  • Dealroom's note on Shipper's "After Automation" highlights a similar read: the paradox is not just that automation leaves humans around, but that it creates new work in agent maintenance, judgment, and original differentiation after models commoditize default competence. Direct link: https://app.dealroom.co/news/note/after-automation-why-ai-progress-creates-more-work-for-humans-not-less

Full Recap

00:00-02:56 - The setup: Dan Shipper's next predictions - Lenny frames the episode around Shipper's previous call that people were sleeping on Claude Code for non-engineering work, then says this return appearance is about what else Shipper thinks will happen next (00:00-02:33). - The cold open previews the core claims: the AI job apocalypse is not really a thing, Shipper is bullish on PMs and full-stack designers, automation still needs humans, SaaS is not dead, and CLIs are over (00:10-01:31).

02:56-09:17 - Every as a "pocket of the future" - Shipper says Every tries to predict the future by living in it together, with an almost 30-person team of AI early adopters across engineering, design, writing, editing, sales, and customer support (04:48-05:20). - He says the company gets early access to models, beta-tests tools, writes about what it notices, and shifted from people looking at code to people talking to computers in English through Claude Code and similar systems (05:38-07:39). - He describes a practical "reach test": a tool matters when people naturally reach for it in the morning without being told to use it (07:43-07:48).

09:17-16:39 - Work splits into company super-agents and personal work surfaces - Shipper's first work prediction is a bifurcation: employees will delegate to at least one company agent, and much of the rest of their work will happen inside environments like Codex or Claude Co-work (11:15-12:33). - He initially expected personal agents for everyone, but now thinks the near-term pattern is one company-level "super agent" maintained by someone responsible for making it useful across the company (13:07-15:29). - The reason is operational, not philosophical: agents like OpenClaw break, need setup, need context, and need a human who cares enough to keep them useful (14:12-15:09).

18:13-25:39 - Codex and Claude Code become the work operating system - Shipper says putting an agent on a user's computer matters because the agent has access to what the user has access to, can operate through the terminal, and can work across local context (18:26-19:30). - He argues that coding agents escaped programming because once an agent can build anything on a computer, it can help with almost any knowledge-work task (19:36-19:56). - He says Codex is his current daily driver and describes writing documents in an in-app browser while Codex watches, researches, edits, and uses his computer alongside him (20:58-22:20). - The product implication is that instead of AI being baked into every SaaS tool, SaaS tools may increasingly run inside an agent workspace where the user's own AI can see and operate them (23:48-24:42).

24:22-31:13 - SaaS has to serve humans and agents together - Shipper says products should become friendly to agents by making HTML usable, keeping CLI and web states in sync, and allowing humans and agents to use the same software surface at the same time (24:37-25:33). - He says this could change SaaS economics because users bring their own AI tokens into apps like Proof rather than forcing every vendor to pay for all AI usage (25:18-25:39). - He argues that human-agent collaboration creates new UX and infrastructure requirements: approval inboxes, logs, fast rollback, concurrency displays, and capacity for agents making many requests quickly (28:26-30:10). - He gives a support-loop example where an agent sends a bug report with exact repro steps, suspected code context, and enough information to become a GitHub issue that another agent can help fix (30:22-31:13).

31:13-39:01 - CLIs fade and two agents are better than one - Shipper says the CLI era was speed-run, not because CLIs vanish, but because GUI workspaces can preserve the same agent benefits while being better for non-programmer work (31:13-32:35). - Lenny summarizes the two work modes as a company super-agent in Slack plus a desktop agent surface like Codex or Claude Code where normal computer work happens (32:43-33:11). - Shipper adds that two agents can be better than one because a user's base agent can carry far more context into another app or agent than the user would type manually (33:34-36:22).

36:22-48:36 - The automation paradox and benchmark caveat - Shipper's SaaS view is contrarian: he says SaaS is not dead, Every still pays for a lot of SaaS, and agents may increase demand for SaaS by becoming high-volume users of those products (36:42-38:46). - Lenny then points out that Every doubled headcount despite being AI-forward; Shipper answers that automation still needs humans because every automated system needs someone ensuring the work is good (39:01-39:37). - Shipper connects this to his earlier allocation-economy framing: humans working with AI look more like managers, and managers still spend a lot of time working (39:28-39:55). - He cautions that benchmarks overstate autonomy because they measure framed and scored problems, while humans still do the harder unmeasured work of deciding the right frame, prompt, rewrite, or higher-level objective (40:12-45:47). - His Proof rewrite benchmark is the example: earlier models scored around 30 out of 100, GPT 5.5 scored about 62 with an Opus 4.7 plan, while human senior engineers using AI scored in the high 80s or low 90s (42:05-42:44).

48:36-1:08:28 - The shape of work changes toward review, agent operations, and AI-written internal docs - Lenny summarizes the practical advice: use Codex or Claude Code more, let agents use your products, experiment with company agents, and expect SaaS demand to persist (48:36-50:15). - Shipper says the shape of work changes because more people can submit pull requests and technical artifacts, which pushes engineers and product people toward coherence, review, integration, and deletion decisions (50:15-52:49). - He says forward deployed engineers are becoming important because every agent needs a human, including agents running inside model companies and inside Every's own consulting practice (53:56-55:46). - Shipper reframes "babysitting" agents as building systems that let less technical people use AI without doing something dumb, while experts get pulled toward deeper and more generative questions (1:02:17-1:03:09). - He predicts people will read much more AI-generated internal writing and like it, as long as the sender stands behind the document and knows what is in it (1:03:11-1:05:10).

1:08:28-1:21:02 - PMs, designers, and the anti-apocalypse case - Shipper is "super bullish" on PMs because AI lets lightly technical product people combine product taste, user sense, and enough technical context to ship without organizing a full team for every change (1:08:28-1:10:28). - He is also bullish on full-stack designers because AI helps them turn taste, interaction ideas, and differentiation into working pull requests rather than static handoffs (1:11:05-1:12:49). - His broader anti-apocalypse claim is that models make yesterday's human competence cheap, but that cheap default competence becomes commoditized and creates more value for people who can make something new and specific with it (1:13:11-1:15:55). - His survival advice is to "ride the models": keep trying new models on real work, turn over old rocks again when capabilities improve, and treat the frontier as wherever AI meets a real human doing a real task (1:16:00-1:19:35).

1:21:02-1:34:03 - Final advice and lightning round - Lenny observes that much remains familiar, including SaaS, Slack, email, and jobs, even as every role's workflow changes around the edges (1:21:02-1:22:11). - Shipper says people should try their workflows in Codex or Co-work, get comfortable with agent products, and find enjoyable use cases instead of only acting from fear or FOMO (1:23:35-1:24:31). - In the lightning round, Shipper recommends books, names Codex as both a product he loves and an underrated AI tool, and says people can be useful by putting their hands on AI, finding ways to use it well, and sharing what they learn (1:25:24-1:33:39).

00:00-02:56 - The setup: Dan Shipper's next predictions

  • 00:00-02:33 - Lenny frames the episode around Shipper's previous call that people were sleeping on Claude Code for non-engineering work, then says this return appearance is about what else Shipper thinks will happen next .
  • 00:10-01:31 - The cold open previews the core claims: the AI job apocalypse is not really a thing, Shipper is bullish on PMs and full-stack designers, automation still needs humans, SaaS is not dead, and CLIs are over .

02:56-09:17 - Every as a "pocket of the future"

  • 04:48-05:20 - Shipper says Every tries to predict the future by living in it together, with an almost 30-person team of AI early adopters across engineering, design, writing, editing, sales, and customer support .
  • 05:38-07:39 - He says the company gets early access to models, beta-tests tools, writes about what it notices, and shifted from people looking at code to people talking to computers in English through Claude Code and similar systems .
  • 07:43-07:48 - He describes a practical "reach test": a tool matters when people naturally reach for it in the morning without being told to use it .

09:17-16:39 - Work splits into company super-agents and personal work surfaces

  • 11:15-12:33 - Shipper's first work prediction is a bifurcation: employees will delegate to at least one company agent, and much of the rest of their work will happen inside environments like Codex or Claude Co-work .
  • 13:07-15:29 - He initially expected personal agents for everyone, but now thinks the near-term pattern is one company-level "super agent" maintained by someone responsible for making it useful across the company .
  • 14:12-15:09 - The reason is operational, not philosophical: agents like OpenClaw break, need setup, need context, and need a human who cares enough to keep them useful .

18:13-25:39 - Codex and Claude Code become the work operating system

  • 18:26-19:30 - Shipper says putting an agent on a user's computer matters because the agent has access to what the user has access to, can operate through the terminal, and can work across local context .
  • 19:36-19:56 - He argues that coding agents escaped programming because once an agent can build anything on a computer, it can help with almost any knowledge-work task .
  • 20:58-22:20 - He says Codex is his current daily driver and describes writing documents in an in-app browser while Codex watches, researches, edits, and uses his computer alongside him .
  • 23:48-24:42 - The product implication is that instead of AI being baked into every SaaS tool, SaaS tools may increasingly run inside an agent workspace where the user's own AI can see and operate them .

24:22-31:13 - SaaS has to serve humans and agents together

  • 24:37-25:33 - Shipper says products should become friendly to agents by making HTML usable, keeping CLI and web states in sync, and allowing humans and agents to use the same software surface at the same time .
  • 25:18-25:39 - He says this could change SaaS economics because users bring their own AI tokens into apps like Proof rather than forcing every vendor to pay for all AI usage .
  • 28:26-30:10 - He argues that human-agent collaboration creates new UX and infrastructure requirements: approval inboxes, logs, fast rollback, concurrency displays, and capacity for agents making many requests quickly .
  • 30:22-31:13 - He gives a support-loop example where an agent sends a bug report with exact repro steps, suspected code context, and enough information to become a GitHub issue that another agent can help fix .

31:13-39:01 - CLIs fade and two agents are better than one

  • 31:13-32:35 - Shipper says the CLI era was speed-run, not because CLIs vanish, but because GUI workspaces can preserve the same agent benefits while being better for non-programmer work .
  • 32:43-33:11 - Lenny summarizes the two work modes as a company super-agent in Slack plus a desktop agent surface like Codex or Claude Code where normal computer work happens .
  • 33:34-36:22 - Shipper adds that two agents can be better than one because a user's base agent can carry far more context into another app or agent than the user would type manually .

36:22-48:36 - The automation paradox and benchmark caveat

  • 36:42-38:46 - Shipper's SaaS view is contrarian: he says SaaS is not dead, Every still pays for a lot of SaaS, and agents may increase demand for SaaS by becoming high-volume users of those products .
  • 39:01-39:37 - Lenny then points out that Every doubled headcount despite being AI-forward; Shipper answers that automation still needs humans because every automated system needs someone ensuring the work is good .
  • 39:28-39:55 - Shipper connects this to his earlier allocation-economy framing: humans working with AI look more like managers, and managers still spend a lot of time working .
  • 40:12-45:47 - He cautions that benchmarks overstate autonomy because they measure framed and scored problems, while humans still do the harder unmeasured work of deciding the right frame, prompt, rewrite, or higher-level objective .
  • 42:05-42:44 - His Proof rewrite benchmark is the example: earlier models scored around 30 out of 100, GPT 5.5 scored about 62 with an Opus 4.7 plan, while human senior engineers using AI scored in the high 80s or low 90s .

48:36-1:08:28 - The shape of work changes toward review, agent operations, and AI-written internal docs

  • 48:36-50:15 - Lenny summarizes the practical advice: use Codex or Claude Code more, let agents use your products, experiment with company agents, and expect SaaS demand to persist .
  • 50:15-52:49 - Shipper says the shape of work changes because more people can submit pull requests and technical artifacts, which pushes engineers and product people toward coherence, review, integration, and deletion decisions .
  • 53:56-55:46 - He says forward deployed engineers are becoming important because every agent needs a human, including agents running inside model companies and inside Every's own consulting practice .
  • 1:02:17-1:03:09 - Shipper reframes "babysitting" agents as building systems that let less technical people use AI without doing something dumb, while experts get pulled toward deeper and more generative questions .
  • 1:03:11-1:05:10 - He predicts people will read much more AI-generated internal writing and like it, as long as the sender stands behind the document and knows what is in it .

1:08:28-1:21:02 - PMs, designers, and the anti-apocalypse case

  • 1:08:28-1:10:28 - Shipper is "super bullish" on PMs because AI lets lightly technical product people combine product taste, user sense, and enough technical context to ship without organizing a full team for every change .
  • 1:11:05-1:12:49 - He is also bullish on full-stack designers because AI helps them turn taste, interaction ideas, and differentiation into working pull requests rather than static handoffs .
  • 1:13:11-1:15:55 - His broader anti-apocalypse claim is that models make yesterday's human competence cheap, but that cheap default competence becomes commoditized and creates more value for people who can make something new and specific with it .
  • 1:16:00-1:19:35 - His survival advice is to "ride the models": keep trying new models on real work, turn over old rocks again when capabilities improve, and treat the frontier as wherever AI meets a real human doing a real task .

1:21:02-1:34:03 - Final advice and lightning round

  • 1:21:02-1:22:11 - Lenny observes that much remains familiar, including SaaS, Slack, email, and jobs, even as every role's workflow changes around the edges .
  • 1:23:35-1:24:31 - Shipper says people should try their workflows in Codex or Co-work, get comfortable with agent products, and find enjoyable use cases instead of only acting from fear or FOMO .
  • 1:25:24-1:33:39 - In the lightning round, Shipper recommends books, names Codex as both a product he loves and an underrated AI tool, and says people can be useful by putting their hands on AI, finding ways to use it well, and sharing what they learn .

Technical Need To Knows

  • Codex: Shipper uses Codex as a desktop-style agent work surface where he can write, browse, research, manage email, and operate tools in one environment. In the episode, Codex is not just a coding assistant; it is a candidate operating system for knowledge work (18:13-22:20).
  • Claude Code and Claude Co-work: Claude Code is the coding-agent paradigm Shipper says helped reveal that terminal-native agents could do broader non-engineering work. Claude Co-work is described as Anthropic's nicer wrapper around that idea, part of the same broader shift toward agentic work surfaces (18:13-19:56).
  • Company super-agent: A shared company-level agent, likely in Slack, that employees can ask to do work or answer questions. Shipper now thinks this model is more practical than every employee maintaining a personal agent because one responsible operator or team can keep the shared agent useful (11:15-16:39).
  • OpenClaw: An agent harness used as an example of both promise and friction. Shipper says people get excited about it, then discover that setup, server access, breakage, and ongoing maintenance make personal-agent adoption hard without dedicated human care (13:07-15:09).
  • Forward deployed engineer: A human operator-engineer who makes agents useful inside an organization or for customers. In Shipper's framing, this person builds the system, monitors failure modes, updates context, and makes AI usable for less technical teams (53:56-55:46, 1:02:17-1:03:09).
  • Agent-friendly SaaS: Software designed so both humans and agents can operate it. This includes usable HTML, CLI and web synchronization, clear logs, approvals, rollback, concurrency handling, and infrastructure that can tolerate agent request volume (24:22-31:13).
  • Bring-your-own tokens: Shipper's SaaS economics claim that users may increasingly bring their own AI capacity through Codex or Claude-style work surfaces, so SaaS companies do not have to pay for all embedded AI usage themselves (24:22-25:39, 36:22-38:46).
  • Benchmarks and autonomy: Shipper argues benchmarks can mislead because they evaluate already-framed tasks. The unmeasured human work is deciding the frame, recognizing a bad premise, changing the benchmark, or forcing a rewrite instead of accepting a local patch (40:12-45:47).
  • GPT 5.5 and Opus 4.7 references: Shipper uses these as current model examples in his internal senior engineer benchmark. The important point is not the exact model leaderboard; it is his claim that even strong model progress still depends on human framing, review, and higher-level judgment (42:05-43:04).
  • Proof, Spiral, Cora, Claudie, and Every: These are Every's products and internal systems used as evidence. Proof is Shipper's markdown/editor example, Spiral is a writing app, Cora is an email agent, and Claudie is an internal consulting-practice agent (21:35-22:48, 42:33-43:19, 53:56-55:26, 1:08:55-1:10:28).
Back to allocation feed