When AI is not a good idea

Most articles about AI are about what's possible. This one is about when you shouldn't do it. That's more useful, because a project that should never have started costs you money, time and your team's trust. Below are six situations where AI is the wrong tool, and what works instead.

When is AI the wrong tool?

AI is good at tasks with a lot of repetition, a lot of variation in the input and an acceptable margin of error. Think of reading text, summarising, classifying and making suggestions. It's bad at tasks that must be precise and predictable, where the rules are fixed, or where a single mistake can't be reversed.

So first ask yourself: is this a task that needs judgement, or a task that follows rules? With rules, ordinary automation is faster, cheaper and more reliable. It's less exciting and it works the same way every time.

1. Your process isn't a process, it's a habit

Ask three colleagues how they handle a request. If you get three different answers, you don't have a process. You have habits, and there's nothing there to automate.

What to do instead: write the process down first and resolve the differences. Often half the steps turn out to be unnecessary. Cutting that half delivers more straight away than any model. Automation comes after that.

2. The volume is too low to earn anything back

A task that happens twice a week and takes ten minutes adds up to just over sixteen hours a year. That won't pay for an integration, a model or maintenance. And maintenance never stops: systems change, APIs change, your own process changes.

What to do instead: leave that task as it is, or solve it with a template, a form or a setting in software you already have. Look for the high-volume processes. How to rank them is in our article on the AI business case and under what manual work costs.

3. The mistake must not happen

A model gives an outcome with a probability, not a guarantee. For a misclassified email, that's not a problem. For a medication dose, a payment or a decision with legal consequences, it is.

What to do instead: use the model to prepare, not to decide. Let it summarise, queue and flag. The decision stays with a person, with the reasoning visible on one screen. That's also what the EU AI Act expects for high-risk applications.

4. Your data lives in six places, in six formats

The same customer spelt three ways. Product codes that don't match between the online store and the ERP. Time sheets where half the descriptions are empty. A model let loose on this produces neat nonsense, and that's harder to spot than an empty cell.

What to do instead: get the source in order first and decide which system holds the truth for each piece of data. It's dull work and it's the precondition for everything that follows. With us, that's part of the first phase of AI implementation: access, data, integrations and ownership.

5. There's already a package that does this

Many companies want something built for a problem their current software already solves. It's just never been set up, or nobody knows the feature exists.

What to do instead: first take a hard look at the licences and settings you're already paying for. Only build once there's a demonstrable gap. When your own software does make sense is explained under custom software.

6. Nobody in the organisation owns it

An AI application isn't a project with an end date. Someone has to review the outcomes, make corrections and decide when something changes. Without that person, quality slowly slips without anyone noticing.

What to do instead: appoint the owner first, with time in their diary. No owner, no project. It's a hard rule and it saves you a lot.

How do you spot a project that will fail before it starts?

There are a few signs you almost always see before it goes wrong.

• The trigger is the technology: someone saw a demo and wants to do something with it.
• There's no baseline, so afterwards there's no way to tell whether things got better.
• The scope is a department instead of a task.
• The people who do the work today aren't involved in the design.
• There's budget for building, but not for running it.

If you recognise three or more, stop and go back to the question of what needs to get better.

So what should you do?

Put your processes side by side and score them on four things: how often it happens, how much time it takes, how bad a mistake is and how hard the systems are to integrate. The processes that happen often, take a lot of time, have a small error impact and are easy to reach come first.

That's the core of the AI Opportunity Audit: five days, one number, three opportunities. And if the outcome is that AI won't deliver anything here right now, that's in the report too. What the audit involves is explained on the AI Opportunity Audit page.