Do AI tools actually make you more productive?

Sometimes, and less reliably than almost everyone believes. In a randomised controlled trial published by METR in July 2025, sixteen experienced open-source developers were 19% slower on real issues when allowed to use AI tools — while believing they had been sped up by 20%. It is a small study on one kind of expert work, so it does not generalise to everything. What it does establish is that self-assessment is not evidence.

Which AI productivity tool should I start with?

Start with whichever of the four jobs you are actually short of. If you lose information, start with capture — transcription. If you drown in material, start with synthesis. If you stare at blank pages, start with drafting. If you repeat the same sequence of steps weekly, start with automation. Most people buy a drafting tool for a coordination problem and conclude AI is overrated.

Are AI meeting assistants worth it?

They are worth it if someone reads the output. Transcription converts attending a meeting into reading about a meeting, which is only a gain when the reading actually happens and replaces attendance. Teams that adopt one and keep the same attendance list have added a cost, not removed one.

What is the highest-value AI automation for most people?

Something small, boring and weekly. The compounding wins come from removing a ten-minute task you do fifty times a year, not from an ambitious workflow you build once and abandon. Automation is also the one category where the time saved is real rather than converted, because no human has to review the output of a rule that either fired or did not.

How do I tell if an AI tool is actually saving me time?

Measure the whole loop, not the generation step: time to write the prompt, plus time to read the output, plus time to fix it, plus the times you threw it away and did the task manually. Compare that against a baseline you actually timed. If you have not timed the baseline, you do not have an opinion — you have a feeling, and the evidence says the feeling runs about 39 percentage points optimistic.

Is AI better for creating or for reviewing?

Reviewing, and it is not close. Models are consistently better at finding problems in a finished thing than at avoiding those problems while making it. Asking one to critique your draft, list what a reader would object to, or find the gap in your plan is reliably more valuable than asking it to produce the draft.

The Complete AI Productivity Guide (2026)

Here is the thing nobody selling these tools will tell you: an AI productivity tool does not remove work. It converts work into a different kind of work.

Drafting becomes reviewing. Attending becomes reading. Doing becomes specifying, then checking. Every one of those trades can be a large win — and every one can be a loss, depending entirely on which kind of work you were actually short of.

Almost all disappointment in this category comes from making a trade you did not need.

The measurement problem

Start here, because it invalidates the way most people choose these tools.

In July 2025, METR published a randomised controlled trial on AI and developer productivity. Sixteen experienced open-source developers worked through 246 real issues — bugs, features and refactors averaging about two hours — in repositories they already knew well. Each issue was randomly assigned to allow or forbid AI tools.

The developers expected AI to make them 24% faster. It made them 19% slower. And afterwards, having actually done the work, they still believed it had sped them up by 20%.

Sit with the size of that gap. The people doing the work, measuring nothing, were wrong by roughly 39 percentage points about the direction of the effect — not the magnitude, the direction.

The honest caveats, because they matter and most citations of this study drop them: sixteen developers is a small sample; they were experts working in codebases they knew deeply, which is precisely the situation where a model’s context disadvantage is largest; and it captures early-2025 tools. It is not evidence that AI slows everyone down at everything. Plenty of other work finds real gains, particularly for less experienced people on unfamiliar tasks.

What it does establish is narrower and more useful: feeling faster is not evidence of being faster. The subjective sense of productivity these tools produce is real, immediate, and uncorrelated with the outcome. Any decision you make on that feeling is a coin flip.

The four jobs

“Productivity” is four different problems. Buying for the wrong one is the most common and most expensive mistake here.

Job	The problem	What it converts	Tools
Capture	Information is lost as it happens	Attending → reading	Otter.ai, Descript — see best AI for note-taking and meeting transcription
Synthesis	Too much material, too little time	Reading everything → reading a summary	NotebookLM, Claude
Drafting	Blank pages, slow first versions	Writing → editing	ChatGPT, Notion AI, Gamma — decks: best AI for presentations
Automation	The same steps, over and over	Doing → specifying once	Zapier — SMB automation plan, Zapier vs n8n

Read down the “converts” column and the strategy becomes obvious. Automation is the only row where the work does not come back to you. A rule that moves a file or files a ticket either ran or did not; nobody reviews its prose. Everything else hands you a second task and asks you to believe it is smaller than the first.

That does not make the other three bad. It makes them conditional.

Where the wins actually are

Automation, on small boring things. The compounding gains come from removing a ten-minute task you do weekly, not from an ambitious workflow you build once and never trust. Fifty repetitions a year of something dull is a genuine recovered day, and it never needs proofreading. This is the least glamorous row in the table and the only one with unconditional value.

Synthesis, when you have too much material and know what you are looking for. Twenty PDFs and a question is a job NotebookLM does extremely well, at zero cost. The research guide covers the mechanics and the verification you still owe.

Review, not creation. The single most underused move in this category: stop asking for drafts and start asking for critique. “What would a sceptical reader object to here?” “What is missing from this plan?” “Where does this contradict itself?” Models are markedly better at finding faults in a finished artefact than at avoiding those faults while producing one — and critique has no review cost, because you are already the reviewer.

Capture, if and only if someone reads it. More on this below.

The meeting-notes trap

AI meeting assistants are the clearest illustration of the conversion principle, so they are worth being specific about.

Transcription converts attending a meeting into reading about a meeting. That is an enormous win when the reading genuinely replaces the attending: five people stop sitting through a call, one person skims the summary, four hours become twenty minutes.

It is a pure cost when attendance does not change. The meeting still happens, the same people still sit in it, and now a document exists that nobody opens. You have added storage and a subscription.

So the question to ask before buying Otter.ai or anything like it is not “is the transcription accurate?” It is: which meeting will I stop attending? If there is no answer, the tool cannot pay for itself, however good it is.

The same test applies to summaries generally. A summary nobody reads is not a time saving; it is a time saving someone else was supposed to collect.

What to automate first

If you take one action from this guide, take this one. Look at your last two weeks and find something that is:

repetitive — same shape every time
frequent — at least weekly
rule-shaped — you could describe it to a new colleague in three sentences
low-stakes if wrong — because occasionally it will be

Then automate exactly that with Zapier or its equivalent, and nothing else, for a month. Resist the ambitious version. Elaborate automations fail quietly, and a broken automation you have stopped checking is worse than the manual step it replaced. When task cost or self-hosting becomes the real question, compare Zapier vs n8n for AI automation.

Measuring it properly

Given the METR result, the only defensible way to know whether a tool helps you is to measure the whole loop:

real cost = prompting + reading the output + fixing it
          + the attempts you abandoned and did by hand

Against a baseline you actually timed. Not remembered — timed.

Three practical rules:

Time one week without the tool first. If you skip this you will never have a comparison, and the feeling will fill the gap.
Count the throwaways. The attempts that produced nothing usable are part of the cost and are the ones memory deletes.
Re-check quarterly. These tools change monthly; a verdict from six months ago is about a different product.

The honest summary

The gains in this category are real but narrower and more specific than the marketing suggests. They concentrate in automation of dull repeatable steps, in synthesis when you already know your question, and in criticism of work you have already done.

They are thinnest exactly where the advertising is loudest: generating first drafts of things you will have to check line by line anyway.

And the most reliable finding available is that you will not be able to feel the difference. Measure it, or accept that you are guessing — see what you actually get on free AI plans before paying for the guess, and the prompt engineering guide for getting more out of the tools you keep.

Task deep-dives on the same theme: best AI for note-taking, best AI for meeting transcription, best AI for presentations, best AI for email, AI automation for small business, and Zapier vs n8n.