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

What It Covers

Dan Shipper, CEO of Every, shares predictions for how AI will reshape work over the next year. Drawing from running a 30-person AI-native company, he argues that SaaS is not dying, the AI job apocalypse is overstated, and that work will bifurcate into two modes: company-wide super agents and codex-style environments replacing traditional desktop workflows.

Key Questions Answered

• •Company Super Agents Over Personal Agents: Early enthusiasm for personal AI agents like OpenClaw collapses in practice because agents require a dedicated human to maintain them. The model that works at scale is one company-wide super agent — Shopify and Ramp both run this model — managed by a forward-deployed engineer. Teams then layer specialized sub-agents beneath it. Personal agents will return as models become less maintenance-heavy, but the near-term architecture is centralized, not distributed.
• •Codex and Claude Code as the New OS: Most professional knowledge work will migrate inside agent environments like Codex or Claude Code, which embed a browser alongside the AI. This means SaaS tools get accessed from within the agent, not the other way around. Users bring their own tokens, which eliminates AI cost burden for SaaS vendors. Shipper runs email, documents, and analytics entirely inside Codex with the in-app browser, achieving inbox zero for ten consecutive days.
• •SaaS Is Not Dying — Buy the Stocks: Agents increase SaaS usage rather than replace it. Every's internal SaaS spend has grown year-over-year despite heavy AI adoption. Agents become high-volume users of existing SaaS products, creating infrastructure demand spikes. The strategic shift for SaaS builders is designing for simultaneous human and agent use: simpler UI, agent-friendly HTML, rollback logs, and approval inboxes — not building a competing AI layer on top.
• •Automation Paradox — More AI Means More Work: Shipper's senior engineer benchmark scores most coding models at 30 out of 100 against human engineers. GPT-5.5 reached 62, a 30-point jump, but still falls short. The gap is not raw capability but judgment: models fix individual issues when told to, while senior engineers recognize when the entire codebase needs a rewrite. This gap means human oversight remains essential, and Every doubled headcount to 30 people over the past year despite full AI adoption.
• •PMs and Full-Stack Designers Are the Power Roles: A PM at Every named Marcus, formerly at Axios, now ships product faster than most engineers by pairing product instincts with Claude Code and Cursor. No engineering handoff required. Similarly, designers who learn to build in agent environments can execute their own interactions without waiting on engineers. Both roles benefit because AI handles execution while human judgment on what to build and how it should feel remains the scarce, non-commoditized skill.
• •CLIs Are Already Over: The terminal-first era of Claude Code was brief. The reason Claude Code succeeded was not the CLI itself but the agent's access to the full computer environment. Once that same access moves into GUI environments like Codex Desktop, most workers — including technical ones at Every — stop using the terminal as a primary surface. The prediction is that within a year, GUI-based agent environments will be the standard, with CLI access retained underneath but rarely touched directly.
• •Models Commoditize Yesterday's Competence: AI models compress prior human expertise into cheap, widely available outputs, making default-quality work indistinguishable and low-value. The structural response is to use those frozen competencies as raw material for novel combinations. Benchmarks measure framed, scorable tasks, but the act of identifying what question to ask or what problem to reframe cannot yet be benchmarked. Riding new models — testing each release against your specific workflows — is the concrete habit that keeps individuals ahead of commoditization.