The most useful AI tool I’ve watched someone use at work does exactly one thing: it turns a rambling 45-minute call recording into six bullets and a list of who promised what to whom. The person using it is a project coordinator who used to spend an hour every Tuesday afternoon writing those notes by hand. She has no opinion about model benchmarks. She has a strong opinion about getting the hour back.
That’s the test worth applying to every AI tool for work you’re considering. Not “is this impressive,” but “did this hand me time this week, and can I point at exactly where.”
Start with the job, not the tool
Most people go wrong in the same way. They read a list of forty apps, sign up for nine, and by week three they’re using one and paying for four. The order is backwards.
Spend ten minutes in your calendar and your sent folder instead. Look for a task that repeats, that you don’t enjoy, and that produces something with a clear shape: a summary, a reply, a table, a set of notes. That’s your first candidate. One task, one tool, one week of testing. There’s a decent argument that starting with a single job rather than twenty apps is the only onboarding advice that reliably survives contact with a real workday.
Where AI tools for work actually earn their keep
After watching a lot of teams try this, the wins cluster in four places.
First drafts of things nobody wants to write
Status updates, project briefs, job descriptions, the “here’s what changed” email after a messy meeting. A model won’t produce a final version you’d put your name on, but it will hand you 300 words to react to in fifteen seconds. Editing is faster than staring at a blank page, and that gap is where the hours hide.
Meeting capture and the follow-up
This is the most measurable win, and the easiest to defend to a manager. A notetaker that joins the call and returns decisions plus action items saves somewhere around 20 to 30 minutes per meeting. Two meetings a day, five days a week, and you’ve clawed back most of a full working day. The catch: someone still has to skim the output before it reaches a client, and names get mangled more often than the demos suggest.
Reading long material quickly
That 60-page vendor agreement, the 200-message support thread, the report that landed at 6pm on a Friday. Asking “what are the five things here that would make me say no” is a far better use of a language model than reading cover to cover when what you need is orientation. It is not a substitute for a lawyer, and it never will be.
Spreadsheets, formulas, and messy cleanup
Describing what you want in plain English and getting a working formula back is a real shortcut for anyone who isn’t a spreadsheet person. Same for splitting full names into two columns, standardising dates, or writing the regex you’d otherwise spend twenty minutes searching for. Check the output on three rows before you drag it down four thousand.
Where AI quietly costs you more time
The failure modes are less dramatic than the hype and less obvious than you’d think. Three patterns come up again and again:
- Numbers you can’t verify quickly. Anything involving totals, dates, or figures pulled from your own systems is a bad fit. You’ll spend longer checking the arithmetic than you saved generating it.
- Wording that carries risk. A reply to a difficult client, a performance conversation, a clause in a contract. The cost of a slightly-off tone here is far higher than the ten minutes you saved drafting it.
- Work that needs your judgement, not your throughput. If the value you add is deciding what matters, outsourcing the thinking removes the part people are actually paying for.
There’s a useful longer breakdown of which AI tools hold up under real workloads and which quietly fall apart, including the ones that are better built in-house in an afternoon than bought.
A short list that covers most office work
Here’s the thing about tool sprawl: a general assistant plus one or two specialists handles roughly 80% of what a typical desk job throws at you. If you’re building a shortlist, think in categories rather than brands.
- One general assistant. ChatGPT, Claude, or Gemini. Pick one, learn its quirks, stop switching.
- One meeting recorder. The single highest-return purchase for anyone in back-to-back calls.
- Something that lives inside the app you already use. An assistant built into your docs or inbox beats a separate tab you forget to open.
- Transcription. For interviews, voice memos, and the five-minute brief you recorded in the car.
Notice what’s missing: no dashboard, no workflow builder, no platform that promises to connect everything. Those are the tools that get bought and then quietly ignored after the trial.
Do the arithmetic before you add another seat
A twelve-person team on six tools at $25 per seat per month is spending about $21,600 a year. That’s a salary. It might be worth it, but it deserves the same scrutiny as a hire.
Ask for one number per tool: hours saved per person per month. A $30 tool that saves 90 minutes pays for itself several times over. A $30 tool where nobody can name the 90 minutes should be cancelled this week, not renewed on the promise of a rollout. The discipline of buying fewer tools and measuring the ones you keep is unglamorous and it’s where most of the actual savings sit.
If you want a repeatable way to test a candidate before you commit, this five-step workflow for putting an AI tool through its paces walks through it with real time figures rather than vibes.
Set the rules before someone pastes a contract into a chatbot
Every company I’ve seen adopt these tools learns the same lesson in the same order: enthusiastic trial, quiet individual usage, then a near-miss involving client data. Get ahead of it with a one-page policy that answers four questions.
What can never go into a public model (client names, personal data, unreleased financials)? Which accounts are company-owned versus personal, because work done in a personal account is work you can’t audit? Who reviews anything customer-facing before it ships? And what happens when a model produces something confidently wrong in a document that’s already been sent?
That last one matters more than the rest. Build the habit of a second pair of human eyes on anything external, and the rest of the policy gets easier to live with.
A 20-minute audit to run this week
You don’t need a strategy deck. You need three columns on a sheet of paper and twenty minutes.
First, list the five tasks you did most often last week. Don’t rank them by importance, rank them by repetition. Second, mark the ones where a draft, a summary, or a cleanup would have saved you at least fifteen minutes. Third, pick the single highest scorer and run it through a real tool for five working days, timing yourself before and after.
At the end of the week you’ll have one number, one habit, and one honest answer about whether the tool earned a second week. That’s more than most teams get from a six-month procurement process. Repeat the loop once a month and within a quarter you’ll have a small set of tools you actually reach for, a clear sense of what to ignore, and the time back that was the point all along.