How to choose AI tools for your business

Choose the task first, then the tool. The process that works is five steps: name something you repeat at least weekly, count what it actually costs you in minutes, check the tool reaches into software you already use, trial it on a real job you have already done, then measure the difference two weeks later. Most bad AI purchases skip steps one and five.

The usual sequence is the opposite: see a demo, sign up, use it twice, forget it, notice the charge four months later. That is what happens when a decision starts with a tool that has no job to do.

What is the five-step framework for choosing an AI tool?

Five steps, in order. Each is cheap, and each is designed to kill candidates before you spend money on them.

1. Name the repeated task. Not "marketing" or "admin" — one thing you do over and over: follow-up emails after a call, voice notes into quotes, chasing unpaid invoices.
2. Quantify it. Minutes per run × runs per month, written down before you look at any software.
3. Check it reaches your stack. It has to touch the inbox, calendar or system where the work already happens. A tool that needs a new place to work needs a new habit.
4. Trial it on real work you have already done. Not the vendor's sample data. Three jobs from last fortnight, where you know what a good answer looks like.
5. Measure again at two weeks. Same arithmetic as step two. If the number has not moved, cancel while it is still one small line on the card.

Steps one and two take ten minutes with a pen. Step four takes half an hour. Step five takes five minutes, and is the one almost nobody does.

Why should you start with the task instead of the tool?

Because the size of the benefit is decided by the task, not the software. Point the same model at two different jobs and the returns are an order of magnitude apart.

Stanford's 2026 AI Index puts measured productivity gains at roughly 14–15% in customer support, 26% in software development and about 50% for marketing output, and notes that gains shrink on work needing deeper reasoning. Stanford HAI, 2026 AI Index

Getting that order wrong is also where failures start. MIT's Project NANDA report found only about 5% of custom enterprise AI tools reach production, and put the cause in workflow integration rather than model quality. Preliminary findings from a corporate sample, not peer-reviewed research, so hold the figure loosely — but nothing there failed because the model was not clever enough. Reported by Virtualization Review, August 2025

How do you work out how much time a task is really costing you?

Minutes per run times runs per month, timed once rather than guessed. Ten minutes a day is about 3.5 hours a month; at $80 an hour that is $280 of your time, which comfortably justifies a $20 subscription. Ten minutes a week is 45 minutes a month, and does not.

• Minutes per run. Time it once with a stopwatch. Memory runs high on tasks you dislike and low on the ones you do on autopilot.
• Runs per month. Count from your sent folder or calendar, not from your impression of a busy week.
• What it costs when it is late. Some tasks are cheap in minutes and expensive in consequences — quotes that go out three days after the call.

Then set a realistic ceiling. In the Federal Reserve Bank of St. Louis survey of US workers, 20.5% of people who had used generative AI at least once in the past week reported saving four hours or more; among daily users that rose to 33.5%. ITIF, May 2025

Most people save less than four hours a week, and savings scale with frequency — another argument for a task you do daily over an impressive one you do quarterly. Business.com's 2026 survey of 1,009 people at US businesses with 2–250 employees found the same by role: managers saved 7.2 hours a week against 3.4 hours for individual contributors, on the same tools. Business.com

How do you check an AI tool fits the tools you already use?

Open the integrations page before the pricing page. If your email, calendar, accounting software or CRM is not named on it, the tool means copying and pasting — and copying and pasting is where adoption quietly dies in week three.

• Is your actual software named? "Integrates with 5,000+ apps via Zapier" is not a built connector, and usually means another subscription.
• Where does the output land? A tool that writes a good draft into its own dashboard has moved the work, not removed it.
• Does it work where the task happens? If that is between jobs or in the car, a desktop-only tool solves nothing.
• Can you get your data out today? Not on a higher tier, not by emailing support. Today, in a format something else reads.

Be honest about which stack you actually have. The same MIT research found only around 40% of companies had bought an official subscription to a large language model, while staff at more than 90% were already using personal AI tools for work. What a business really runs on is whatever people open without being asked.

What should you actually score an AI tool on?

Five things, each with a condition that ends the trial rather than starting a debate. Write them down before you sign up, while you can still be unimpressed.

How should you trial an AI tool?

On work you have already completed, so you can compare the output against an answer you know is correct. Pick three jobs from the last fortnight and mark the results against what you actually sent. You cannot judge quality on a task you have never done — anything fluent reads as competent when there is nothing to check it against, which is precisely the condition a demo creates.

A demo shows you the tool at its best on the vendor's data. You need to see it at its worst on yours.

How do you know whether an AI tool is actually working?

Re-measure the number from step two, two weeks in. Do not rely on how it feels, because how it feels is measurably unreliable.

The sharpest evidence on this comes from a METR study of 16 experienced open-source developers working on 246 real issues in codebases they knew well. With AI tools they took 19% longer. They had forecast a 24% speed-up beforehand — and afterwards, having actually been slower, they still believed the tools had sped them up by about 20%. METR, July 2025

METR is careful about the limits: small sample, expert developers on familiar code, tools from early 2025 the researchers now call out of date. It is not evidence that AI makes people slower in general. It is evidence that a 39-point gap can open between how fast you feel and how fast you are, which is why step five needs a number rather than an impression.

• Minutes per run, timed again. Same task, same stopwatch.
• How many days out of ten you opened it. Fewer than five is the answer on its own.
• What you stopped doing. If nothing came off the list, the time went somewhere else rather than back to you.

Cancelling at two weeks is not failure, it is the framework working. A tool you drop after a fortnight cost you $20 and half an hour. A tool you keep out of optimism costs that every month for two years.

Where does this framework fall down?

• It only evaluates work you already do. If the real opportunity is a service you have never offered, task-first filtering will not surface it.
• Two weeks is too short for anything involving other people. A tool needing three colleagues to change a habit needs six weeks and a named owner, or a no.
• It biases towards small, measurable wins. Mostly a feature — small wins stick — but it will pass over a tool that pays off slowly.

[Steve — add a short example of a tool that looked like a failure at two weeks and came good later, or the reverse.]

The version I run for clients is in the five filters I use, the money side in what AI tools cost a small business, and more about how I work on the about page.