Disclosure: some links on this page are affiliate links. If you sign up through one, I may earn a commission at no extra cost to you. I only link tools I actually use.

Where AI Actually Saves Time (and Where It Quietly Wastes It)

workflowsframeworks

After a couple of years of using AI tools daily, the pattern that predicts whether they’ll save me time has almost nothing to do with which model I’m using. It’s a property of the task: how expensive it is to check the output.

Call it the verification tax. Every AI output arrives unverified, and you pay the tax before you can use it. On some tasks the tax is a glance. On others, verifying the output costs as much as doing the work yourself — and on those, AI is a treadmill that feels like progress.

Here’s the taxonomy I actually use.

Where it reliably saves time

Transformation with the source in hand. Reformatting, restructuring, converting notes into prose, prose into bullet points, a transcript into minutes. You already have the truth in front of you, so verification is a skim: does the output match the source? These tasks are AI’s home turf, because nothing new has to be true.

First drafts you were dreading. The blank page is a real cost, and an imperfect draft demolishes it. The key discipline: treat the draft as scaffolding, not as text to lightly edit into shape. You’re allowed to keep its structure and throw away every sentence.

Summaries for your own eyes only. Condensing a long document so you can decide whether to read it properly. The stakes are low by construction — the worst case is you read the document anyway, which was the default.

Rubber-ducking and option generation. “Give me five angles on this problem” costs seconds and sometimes surfaces the angle you’d have found an hour later. You’re not trusting the output; you’re using it as a brainstorm partner who never gets tired of your topic.

Code with a built-in checker. Code is unusual: it comes with cheap, partial verification — it compiles or it doesn’t, the test passes or it doesn’t. That discount on the verification tax is a big part of why coding assistants became the first AI category that professionals genuinely pay for. Partial is the load-bearing word, though: tests catch wrong behavior, not wrong design.

Where it quietly wastes time

Facts you’d have to look up anyway. If you must check every claim against a source, the AI step added a middleman. Worse, fluent wrong answers are more dangerous than no answers — they arrive pre-formatted as confidence. Anything involving current prices, current features, or current versions of fast-moving products is in this bucket by default; ask the model’s own training-date lag to vouch for it and it can’t.

Final text in your own voice. Getting a model to sound like you takes longer than sounding like you. The edit loop — “warmer,” “shorter,” “less like a press release” — is entertaining and slow. Draft with it, sure. But the last mile of anything with your name on it is faster by hand.

Judgment calls dressed as questions. “Should I take this job?” “Is this design better?” A model will answer, fluently, and the fluency will feel like insight. But you can’t verify a judgment call — there’s no source to check it against — so what you’re really doing is laundering your own hunch through a confident third party.

High-stakes anything without review. Legal, medical, financial, security. The verification tax here isn’t just high, it’s mandatory and professional. AI can help you prepare better questions for the expert; it cannot be the expert.

The 15-minute audit

Take your last work week and list the tasks that recur. For each one, ask two questions:

  1. How expensive is verification? A glance, a skim, a line-by-line read, or an expert review?
  2. Do I have the source of truth in hand, or would the AI be generating claims I’d have to check elsewhere?

Glance-or-skim and source-in-hand: automate aggressively, this is free time lying on the floor. Line-by-line without a source: leave it alone, or restructure the task so the truth is in the prompt. That single sort, applied honestly, beats any tool recommendation I could give you — and it keeps working when the tools change, which is roughly monthly.

If you’re deciding whether a specific tool deserves money for one of the tasks in your “automate” pile, I wrote a separate framework for that: How to evaluate an AI tool before you pay for it.