How I choose AI tools for a business

Five filters, applied in order. Does it fit a workflow you already have? Can it be set up in under 30 minutes? Does it return more in hours than it costs in money and attention? Does it survive the company behind it going away? And can you buy it off the shelf today? Most candidate tools fail the first question, which is why the list at the end is short.

The order is deliberate. The filters run cheapest-test-first, so a tool that does not fit the way you already work never costs me the twenty minutes it would take to price it up properly.

Do you start with the tools or with the business?

With the business, every time. I do not open a directory until I know which three parts of someone's week are costing the most hours, because a tool chosen before the diagnosis is just a guess with a monthly fee attached.

This is also where most people get stuck, and the data backs that up. Across the EU in 2025, only about 20% of enterprises used AI technologies at all — and among those that considered it and then decided against, the single most common reason was a lack of relevant expertise, cited by 70.9%. Not cost. Not regulation. Not knowing what to do with it. Eurostat, data extracted December 2025

So the first half of the job is not selection at all. It is working out what the tool is supposed to be for. Everything below assumes that part is done.

Does the AI tool fit a workflow you already have?

This is filter one and it eliminates more candidates than the other four combined. If using the tool means opening a new tab, remembering a new habit and doing the work somewhere it has never happened before, it will be abandoned within a month — no matter how impressive the demo was.

There is a real tension here worth being honest about. McKinsey's State of AI survey found that the organisations getting genuine value are 2.8 times more likely to have fundamentally redesigned their workflows — 55% versus 20% of everyone else. Redesign wins at enterprise scale. Reported by CX Today

But a five-person business has no change management function and no slack. So the practical version of that finding, for a small operation, is the opposite instruction: put the tool where the work already happens and let any redesign come later, if it comes at all.

• Where does this work live today? Inbox, phone, spreadsheet, WhatsApp, a notebook. The tool has to reach into that place.
• Who has to change behaviour? If it is more than one person, the odds drop sharply.
• What breaks if someone forgets to use it? A tool that only works when remembered is a tool that stops working.
• Does it replace a step or add one? Adding a step is fine only when it removes two.

How long should an AI tool take to set up?

Under 30 minutes from signing up to first useful output. That is a hard line, not a preference. Anything longer is competing directly with paid work, and paid work wins — the setup gets postponed to the weekend and the weekend never arrives.

The evidence for what happens to software that never gets properly adopted is not subtle. Zylo's 2026 SaaS Management Index found organisations leave an average of 36% of their SaaS licences unused. That is money already spent on tools that were bought with good intentions and then quietly ignored. Zylo, January 2026

How do you decide whether an AI tool is worth the money?

By arithmetic, and with a wide margin. Hours returned per month, multiplied by what an hour of your time is worth, has to be roughly five times what the tool costs to run. Not double. Five times — because the estimate of hours saved is always optimistic and the cost is always the floor.

Be sceptical about the hours figure — including your own. Business.com's 2026 Small Business AI Outlook Report, surveying 1,009 people at US businesses with 2–250 employees, put average savings at 5.6 hours a week for workers using AI tools. Business.com The economists measuring the whole workforce, users and non-users together, landed far lower: self-reported time savings equivalent to 1.4% of total work hours, which is well under an hour a week. Bick, Blandin and Deming, NBER

Both numbers are real. The gap between them is the gap between people who adopted something that fit and people who did not. I plan against the low end.

What happens if the AI tool company shuts down?

Assume some of them will, and choose so it does not matter much. CB Insights' March 2026 analysis of 431 venture-backed companies that shut down since 2023 found 70% ran out of capital and 43% had poor product-market fit. CB Insights A small business does not get advance warning of any of that.

So the filter is not "will this company survive?" Nobody can answer that. It is "what does it cost me if it doesn't?"

1. Can I export everything today, in a format something else can read? If the export button is hidden or the format is proprietary, that is the answer.
2. Does it hold data or just do a job? A transcription tool disappearing is an inconvenience. A CRM disappearing is a crisis.
3. How many other things depend on it? A tool wired into five other systems is a much bigger bet than a standalone one.
4. Is there an obvious replacement? Categories with three credible competitors are safe. Categories with one clever startup are not.

In practice this pushes recommendations towards boring, established products for anything that holds data, and allows more adventurous picks where the tool is doing a job rather than storing your business. [Steve — add an example of a tool you recommended that later shut down, and what the switch actually cost the client.]

Why only off-the-shelf tools?

Because anything custom needs an owner, and in a small business that owner is you. A bespoke automation is fine on the day it is built and a liability the first time an API changes, a login expires or the person who understood it goes on holiday.

Off-the-shelf means you sign up, you pay a subscription, someone else maintains it, and if it breaks there is a support address. The constraint rules out a lot of clever ideas. It also means the recommendations still work six months later without anybody being called.

The best tool you will never set up is worth exactly nothing. The second-best tool running by Thursday is worth the whole engagement.

Where does this method fall down?

It is conservative, and conservatism has a cost. Three places it gets things wrong:

• It under-rates tools that need a habit change but genuinely pay off. Some friction is worth it. This method passes over those in favour of something easier and smaller.
• The 30-minute rule penalises anything that touches your existing data. Connecting a tool properly to an inbox or an accounting system can be worth an hour. The rule says no; occasionally the rule is wrong.
• It cannot tell you what you are not doing. These filters judge candidates against work that already exists. If the real opportunity is a service you have never offered, no filter will surface it.

And the field moves. A list written today holds up for six to twelve months, after which the categories are the same but some of the specific picks have been overtaken — which is exactly why the method matters more than the list. The longer version of what a full engagement looks like is in what an AI tools assessment is, and there is more about how I work on the about page.