Should you automate it? A checklist to run before you buy any AI tool 

Automate a task only when it's repetitive, follows a stable process, works from inputs you already have in a usable form and has a named person who checks the output. If any one of those is missing then fix the process first, because automating a broken process makes the mess faster.

Why do most AI projects fail before they start?

Most AI projects fail because of what surrounds the tool and not the tool itself. The task wasn't clear, the process was still changing, the data was messy or nobody owned the result, and you can check for every one of these in an afternoon before you spend a rupee. That's why I think this list beats any vendor demo.

Run the six questions below on each task you want to automate and answer them honestly. A task that scores well is a good candidate, and if it doesn't you still win because the answers show you what to fix.

What is the checklist?

Is the task repetitive and does it happen often?

Count how often it happens in your business. A task that happens many times a week is a good candidate, but a task that happens twice a month rarely repays the effort of building anything and my rule is to leave it alone. If the answer is yes then go on, and if it's no then a written checklist is probably enough.

Is the process stable, or is it still changing?

Ask yourself whether the steps have stayed the same for months and whether everyone does them the same way. If the process changes often then your automation will always be out of date, and if two people do it differently then agree one way first.

Is your input clean enough for a tool?

Are your documents readable, the fields filled in and the names spelled the same way each time? AI copes with some mess, but missing or contradictory data gives you wrong results, and poor input is the most common hidden problem so I'd check it first.

What happens when the tool gets it wrong?

Picture the worst mistake it could make for you, like a customer getting a wrong quote or a payment going to the wrong account. If the mistake is cheap and easy for you to catch then go ahead with care, but if it is costly or hard to reverse then add a strong check or keep the task manual.

Who approves the output before it goes out?

Name one person, and I mean one person rather than a team or whoever is free. That person checks the result before it matters and needs the time to do it. If you can't name anyone then nobody owns the result and errors will slip through.

Can you actually measure what it saves you?

Write down how long the task takes you today, or how many enquiries you miss, using your own records. Measure again after the change, because if you can't measure it then you can't tell whether the project paid for itself.

How do you score it?

Give one point for each yes and the score tells you what to do next, but treat it as a guide to your own judgement and not a rule that decides for you.

A no on the question about mistakes or the one about approval matters more than the rest, so treat either as a stop sign until you have an answer.

What does a scored example look like in practice?

Say your business types supplier invoices into your accounts by hand every week. The task is frequent and the invoices arrive as clear PDFs, so that's two yeses. A bookkeeper reviews every entry and the hours are logged, which means yes on approval and measurement. A wrong entry gets caught in that review, so the mistake question is a yes. But two of your staff enter invoices in different ways, so the process isn't stable and that's a no. The score is five, so agree one way of entering invoices and then build.

What should you do when you get a no?

A no is useful because it shows you where the work is, and each one has a plain fix -

• If it's not repetitive then use a checklist or template and do not build anything.
• If it's not stable then agree one way of doing the task, write it down and test it for a few weeks.
• If the input is poor then clean the data at the source, for example with required fields in your forms.
• If mistakes are costly then add a human check or keep the task manual.
• If there's no approver then decide who owns the result and give them time to do it.
• If you can't measure it then start counting now, before you change anything.

None of this is wasted effort and I'd say it's the best part of the exercise. A written procedure and a named owner improve your business even if you never automate anything.

Frequently asked

My short answer to the questions owners ask most is to pick a frequent task with a clear owner and fix the process before you build, and I hope the longer answers below help!