Most AI startups automate the tasks people didn't want automated

Two complaints keep showing up in the same threads. One: a striking fraction of AI startups are automating tasks nobody wanted automated. Two, said more bluntly: here comes another AI app nobody asked for. Both are pointing at the same mistake, and it has nothing to do with the model quality.

The mistake is automating the wrong half of the job. Every role is a mix of work people resent and work people value. The resented half — the copy-paste, the reformatting, the chasing, the manual reconciliation — is where automation is a gift. The valued half — the judgment, the relationship, the creative call, the thing they'd put on a résumé — is where automation feels like a threat or an insult. Too many products confidently automate the second half because it's the more impressive demo.

The two failure modes

• Automating the part people like. If your product does the satisfying, identity-defining piece of someone's work, you're not saving them time — you're taking the good part and leaving them the drudgery. They won't thank you. They'll resent you, quietly, and not adopt.
• Automating the part people won't trust a machine with. Some tasks carry consequences — legal, financial, reputational — where a human will always want the final call. Automate that and adoption stalls at the exact moment the stakes get real, which is the only moment that matters.

The tedious, low-stakes, high-volume work is the opposite on both axes: nobody's identity is tied to it, and nobody minds if a machine does it. That's the target. It's just a less exciting pitch, so founders skip past it toward the flashier task and build something technically impressive that nobody asked for.

How to find the task people actually want automated

• Ask what they'd happily never do again. Not "what's hard" — what's tedious, repetitive, and beneath them. The eye-roll tasks. That's where gratitude and adoption live.
• Watch where the manual workarounds already are. If people have built spreadsheets, macros, or a "we just do it by hand every Friday" ritual, they've already told you this is worth automating. They voted with effort.
• Check what they'd be uncomfortable handing off. If the honest answer is "I'd never let a tool do that unchecked," that's not your wedge — that's a place to assist, not replace.
• Separate impressive from wanted. The most demo-able task and the most wanted-automated task are rarely the same one. Build for wanted. Demo the impressive part later, if at all.

What the AI-skeptic crowd gets right and wrong

Right: a huge amount of what ships is a solution hunting for a problem — capability-first, need-never-checked. "Nobody asked for this" is often a precise diagnosis, not just snark.

Wrong: the conclusion isn't that AI products are pointless. It's that the ones that work started from a task people genuinely wanted off their plate, confirmed that before building, and resisted the pull toward the flashier demo. The technology was never the problem. The skipped validation step was.