AI Productivity Fails

Shrivu ShankarMay 10, 2026ShareArticle voiceover0:00-13:32Audio playback is not supported on your browser. Please upgrade.Most AI users today get ~10–20% more productive no matter how “game changing” they claim it is or how many lines of code they output. Yet, I still think 2x or even 10x+ is both real and reasonably expected. Real transformation requires two changes at once: personal practice and organizational refactoring. Whether the output lands as 10x leverage or 10x slop depends on the practice and the org around it. So far in 2026, I’ve seen exponential increases in output but linear increases in realized impact. This post covers some of the issues I’ve found “debugging” the issue. I’ve grouped observations and recommendations into: Personal Pitfalls — guidance for individuals Organization Pitfalls — guidance for organizations and leadership Cartoon via GPT Image 2Personal PitfallsYou don’t shift left, so you don’t understand what you ship.AI removes the friction that used to force planning. Without something pushing back on bad abstractions, the upfront thinking gets skipped silently. You ship systems you can’t debug or extend, full of blanks AI quietly filled in. Outline first: headers, structure, audience, plan, principles, what-done-looks-like, then let AI fill it in. Review shifts up the stack: outcome-vs-plan instead of line-by-line. The interrogation belongs inside planning: spawn subagent critics to red-team the plan before you let anything generate against it. The easy review at the end is only easy because the hard review at the start was good. If you find it taxing to review your own AI generated output, you didn’t shift left. Context overhead doesn’t shrink with task size, and small tasks are mostly edges. AI handles the middle 80% of work well but can be brittle on the first and last 10%: setup, edge cases, final-mile review. With AI, often a two-line fix pays the same setup cost as a full feature. So you spend more time briefing the agent than just writing it. Even then it lacks context, ships something subtly wrong, and you redo it. Lift task ambition until the context cost is justified. The heuristic: if it’s smaller than a meaningful unit of work (a PR, a section, a chart, a campaign), it’s probably too small. Starting “small” with AI might be the reason you don’t graduate to larger leaps of AI-driven outcomes. AI scales generation, not your human-bound working memory. The ceiling on parallel agents is how many threads a human can hold without dropping context or understanding, and that number is small. You either single-thread with a coffee break while you watch it code or overshoot and abandon five threads. Stay at ~three or fewer active in your cognitive window, or close the loop (next section) and pull fully out. If you are only running a single session at a time, you are probably not delegating enough. If you are struggling to manage dozens of sessions, consider whether you even need to be “managing” them, whether sessions should take longer leaps (fewer agents that do more work), or to intentionally stay linear until you invest in the right skills and context. Closing the last 1% is verification infrastructure and an incredibly well-defined definition of done: a different discipline than generation, so people skip it and layer AI on top of existing handoffs instead. You become the screenshot human, the agent waiting on your click or upload to validate. Close the loop end-to-end: tests, queries, sandboxes, browser tools, type checks, real API responses. Whatever lets the agent see its own output and iterate without you. Before automating the surrounding process, delete it. Closed loops are how you exceed the ~3-agent ceiling: work runs underneath your cognitive window because it doesn’t need to be in it. Believe it or not, AI can handle the “fuzzy” verifications well too with the right pre-defined context, including the “does this actually make sense for long-term system architecture” and “is this the right feature to be adding to the product” questions that I often hear naively claimed as only something a human can do in the loop. Our chat interfaces often treat every session as ephemeral. Taste is a reusable artifact, but the interface gives you nowhere to deposit it. So encoding never starts, or you overcorrect and write one skill per task, ending up with a directory of one-shots nobody else can use. Build skills for the class of task, not the instance. Some are specs (how to do a thing); others are principles (how to think about a class of things). The durable artifact is the skill (often literally some markdown file), not the prompt. The recognition signal is concrete: any time you find yourself editing or “bullying” the output for a better answer, that’s a rule you can codify once and stop bullying forever. Consider even meta skills that regularly take feedback (from you or across the users of a process) and make the right skill or context modifications from accum

AI Productivity Fails

Key takeaways

More in computer-science