The productivity promises attached to AI tools have reached a point where almost anything sounds plausible. Save ten hours a week. Automate your entire workflow. Replace three team members with one subscription.
Some of that is real. A lot of it isn’t. And the gap between the two tends to cost people more time than they saved.
The Things That Actually Work
Drafting Anything With a Predictable Structure
Emails, reports, summaries, outlines, job descriptions, meeting agendas — any task where the output follows a recognisable pattern is where AI delivers consistently. The structure is already solved; the AI fills it with the right content faster than a human typing from scratch.
The practical test: if you could describe what a good output looks like before you start, AI can probably produce a strong first draft. If you genuinely don’t know what good looks like until you see it, AI will frustrate you.
Reducing the Activation Energy of Difficult Tasks
There’s a specific type of procrastination that comes from staring at a blank document. The task isn’t actually hard — starting it is. AI eliminates that friction reliably.
Ask it to produce a rough first version of whatever you’re avoiding. It won’t be good. That’s not the point. Having something to react to — to edit, reject, or build on — breaks the paralysis in a way that’s difficult to replicate any other way.
Processing Large Amounts of Text Quickly
Reading a 40-page report to extract three relevant decisions is genuinely unpleasant work. Feeding it to Claude or ChatGPT and asking specific questions takes three minutes. For anyone whose job involves regularly processing dense documents, this alone justifies the subscription cost many times over.
The Things That Disappoint
Creative Work That Requires a Genuine Point of View
AI produces competent writing. It rarely produces writing with a recognisable voice, an unexpected angle, or an argument that surprises the reader. For content where distinctiveness is the point — where the goal is to say something that couldn’t have come from anywhere else — AI output needs such heavy editing that it stops being a time saver and becomes a constraint.
Anything Requiring Current, Verified Information
AI models have knowledge cutoffs and confabulate confidently when they don’t know something. Using AI to research fast-moving topics — recent market data, current regulations, live product comparisons — produces outputs that look authoritative and contain errors that aren’t obvious without independent verification. The time spent checking often exceeds the time the AI saved.
Complex Decision-Making With Real Consequences
AI is useful for exploring options and stress-testing reasoning. It’s unreliable as a decision-maker on anything that matters. It doesn’t know your full context, doesn’t carry the consequences of being wrong, and optimises for responses that sound reasonable rather than ones that are correct for your specific situation.
The Pattern Behind What Works
Looking across these categories, a pattern emerges. AI performs best on tasks where the quality of the output is easy to assess quickly — where you can read a draft in thirty seconds and know whether it’s usable. It underperforms on tasks where evaluating the output requires as much expertise as producing it would have.
That’s not a criticism. It’s a calibration guide.
A More Useful Frame
Stop asking whether AI is productive. Start asking which specific tasks in your day meet the criteria above, and apply it there first. The results in those areas are reliable enough to justify the rest of the experimentation.