TL;DR: AI agents made output nearly free, and most companies still measure and reward output. So activity numbers rise while the work that actually ships stays flat, and the gap is paid for in review time by the people who still read, with no report tracking it. Under that incentive, skipping the read is the rational choice. The cost shows up first as a longer review queue, then as rubber stamps. Track review wait time alongside output, and make sending unread work back offend no one.
Want to listen to this article?
Orion Playbook newsletter subscribers can listen to the AI-generated audio version for free. Already a subscriber? Log in here
I keep a monthly review of how I work with AI, and one section counts the decisions I logged. All summer, that count went up. Whole sessions went into tuning a local AI agent and rewriting the instructions my AI tools follow. None of that work reached anyone who reads what I publish. Each figure was honest, and each measured activity, on a dashboard I designed myself.
Lazy work used to look like less. Shopify co-founder Tobi Lütke now says the failure case of lazy work is "over output," and that people have taken to "lobbing 'slop grenades' at each other" (if you haven't yet, go listen to his interview on Shane Parrish's podcast). A slop grenade is work an AI produced, passed on unread, for a colleague to check. The people throwing them are doing what they are paid to do. Lazy work now looks like diligence, and it passes every test a busy leader has time to run.
Output was a fair stand-in for outcome only while producing it cost something. That cost kept every output metric roughly honest, because inflating the number meant doing the work. Lines of code, slide or word counts, and hours logged all worked this way; the strategy-consulting deck managed to maximize most of them simultaneously. AI agents that write a whole document or code change on request removed the cost. Nobody's conscience changed.
The bill arrives first as a queue. A Harvard study of engineering records from hundreds of companies found that agents raised the code submitted, the tasks actually finished stayed flat, and the review queue grew. Reviewers sent more back and took longer to clear it. The time a producer saved landed on whoever still reads, and nobody tracks it. Call it the catcher's tax, since somebody has to catch the grenades. The most careful people pay it twice: once to hit their own AI-boosted targets, once to check everyone else's.
The rubber stamp comes later, when nobody watches the queue. Reviewers under load read less closely, as I argued in an earlier issue. Unread work gets approved, the dashboard counts it as throughput, and throughput justifies asking for more volume. The speed is real: Google's annual survey of software teams links more AI use to faster delivery, and to less stable releases. That is where the bill lands, late and far downstream of the number that ran up the cost.
The fix is structural. Organizations track what is plentiful and ignore what is scarce: whether someone who could defend the work has actually read it. At Anthropic, where daily code changes multiplied over the summer, the Claude Code team rebuilt review around that scarcity. Machines now do most of the reading, the engineer who directed the agent signs off as one of its reviewers, and the team tracks how long each change waits for a human reader. In your next operating review, put that wait beside every output number AI can inflate. Start with code, where review already leaves a timestamp; for decks and documents, count how often work comes back. Make "did you read this?" a reason to send work back that offends nobody, which happens only after you have asked it of your own staff first. Your own part: have your AI send your work back before a colleague has to.
Any new measure becomes the next thing to game, and this one will be gamed too. Two numbers that pull against each other are still harder to game than one.
The busiest-looking teams and the most tired reviewers are the same story, told from opposite ends of the queue.
All the images were generated with AI (ChatGPT Images, Gemini Nano Banana, Claude Opus) by Gérard Métrailler.
Food for your AI
Paste this into your AI's standing instructions, so it becomes the first reader to send your work back.
When I ask you to produce anything another person will read, review, or act on, work against my laziness, not for it. First ask who receives it and what they must decide, then write the shortest version that serves that. Do not lengthen it unless I say why. Before I send it, ask me two questions about its content that I must answer without looking at your draft. If I cannot, tell me plainly that I am about to hand someone work I have not read, and name the section to own first.
Sources
Campbell, Donald T. "Assessing the Impact of Planned Social Change." Occasional Paper #8, Public Affairs Center, Dartmouth College, 1976. Reprinted in Journal of MultiDisciplinary Evaluation 7(15): 3–43, 2011. https://doi.org/10.56645/jmde.v7i15.297.
Chen, Fiona, and James Stratton. "Artificial Intelligence in the Firm: Bottlenecks in Software Production." Working paper, Harvard University, version of August 4, 2026. https://fion.ac/jellyfish.pdf. Accessed 2026-10-26.
Google Cloud. "Announcing the 2025 DORA Report." September 23, 2025. https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report. Accessed 2026-10-26.
Osmani, Addy, Morgan Westlee Lunt, and Qing Wang. "How the Claude Code Team Does Code Review." Claude blog, Anthropic, October 15, 2026. https://claude.dev/blog/solving-code-review-at-anthropic/. Accessed 2026-10-26.
Parrish, Shane, host. "Tobi Lütke: AI Agents, Better Decisions, and the Future of Work." The Knowledge Project, Farnam Street, 2026. https://fs.blog/knowledge-project-podcast/tobi-lutke-3/. Accessed 2026-10-26.
