The AI tools assessment checklist

Yes, you can do this yourself. An AI tools assessment is not a proprietary methodology — it is a week of honest observation, a scoring rule, and the discipline to stop researching once you have a shortlist. What you pay someone for is the filtering. Here is the whole thing, in order.

What do you need before you start?

Almost nothing: a way to log what you did, and a rule that you will not look at a single AI tool until step four. The second one is the hard part.

• A log. A spreadsheet with four columns — time, task, minutes, how it felt. A notes app works too.
• Five working days. Not a quiet week and not your worst week. A normal one.
• No tool research until step four. Browse tools now and you will reshape your problems to fit whatever you found. That is how people end up with six subscriptions and no time back.

How do you track where your week actually goes?

Log what you are doing every time you switch tasks, for five days, in a sentence or less. Do not reconstruct the week from memory on Friday afternoon.

When researchers compare what people say their week looks like against what the same people record in a time diary, the two disagree predictably: long-hours workers overestimate, short-hours workers underestimate. Walthery and Gershuny's study of UK working-time estimates found the error is systematic rather than random — so you cannot correct for it by being thoughtful. You have to write it down as it happens.

1. Log the switch, not the block. Every time you change what you are doing, write one line. Switching is the signal.
2. Record interruptions separately. Microsoft's telemetry across Microsoft 365 users found people are interrupted roughly every two minutes during core hours — about 275 times a day — and receive 117 emails and 153 chat messages daily. If your log shows only clean 90-minute blocks, you are not logging honestly.
3. Add a one-word feeling. Dread, fine, or enjoy. You will use it later and cannot reconstruct it afterwards.
4. Do not tidy it up. The scrappy entries are the useful ones.

How should you categorise the tasks you logged?

Sort every line into one of three buckets: communication, repetitive admin, or knowledge work. No sub-categories — a finer taxonomy takes an hour and changes nothing.

The bucket sizes often settle it. In the 2026 SME Business Barometer, 1,000 UK small business owners estimated they spend 11 hours a week on administrative or finance-related tasks — about six working days a month, nearly double what they spend on sales. If your log looks like that, you know which bucket to attack first.

How do you score each task?

Score every recurring task on three factors and multiply them: how often it happens, how long it takes, how mechanical it is. A task has to score decently on all three to be worth automating.

18 or 27 is a strong candidate. 12 or above is worth a look. Below that, leave it — the setup will cost more than the task does.

How do you shortlist tools without falling down a rabbit hole?

Take your top three scoring tasks and find at most two candidate tools for each. Then stop. The failure mode of a DIY assessment is not picking the wrong tool; it is spending eleven hours comparing tools and adopting none.

1. Start with what you already pay for. Your accounting software, CRM and email client almost certainly shipped AI features in the last eighteen months. No new cost, no new login.
2. Then try a general assistant. Much of the communication bucket is handled by one general-purpose assistant and a saved prompt.
3. Only then look for a specialist tool, and only for what the first two steps could not cover.
4. Apply the 30-minute rule. If it cannot be set up and doing something useful inside 30 minutes, it will quietly never happen.
5. Timebox research to 20 minutes per task. When the timer goes off, pick the better of the two and move on.

A short shortlist is not modesty, it is what adoption looks like. The ONS found UK business AI use has risen from around 12% to 35% since late 2023, but the average adopting business uses only about 1.6 AI technologies, barely up from 1.4. Adoption is wide and shallow. Two tools used properly beats seven signed up for.

How do you know whether it actually worked?

Re-run the log for three days, four weeks after you adopt the tools, and compare the minutes on those specific tasks. Do not rely on how it feels.

METR ran a randomised trial with 16 experienced open-source developers across 246 real tasks and found they took 19% longer when AI tools were available. They had predicted a 24% speed-up and, after finishing, still believed AI had made them 20% faster. METR says the result does not generalise to everyone, and it does not. But that perception gap is why you want a number.

Two hours a week is about two and a half working weeks a year. If your re-measurement lands well under that, the usual explanation is that you picked a knowledge-work task when you should have picked a communication or admin one.

The point of tracking twice is that you find out you were wrong while it is still cheap to change your mind.

Is doing it yourself worth it compared to paying someone?

If you have the five to eight hours and will genuinely spend them, do it yourself — the method above is the method, and it costs nothing. Pay someone if your constraint is time rather than capability.

[Steve — add a line here about what you typically find in the first 45 minutes of a real assessment that people miss when they self-audit.]