Don’t put AI in everything

Originally published in AI in Practice on LinkedIn on 4 October 2026.

Some of the AI pitches doing the rounds are giving me flashbacks to the SEO cowboys of the early 2010s. “You’ll be left behind” did a lot of heavy lifting back then. Apparently, it still has a bit of mileage left in it. 😅

There’s a polished demo, a promise to transform the business and, somewhere towards the end, a subscription. What’s sometimes missing is a decent conversation about what actually needs fixing. Which feels like a sensible place to start this newsletter.

AI can do some incredible things. But before we put it into another workflow, we should understand the work. Sometimes the fix is a better form or getting two systems you already pay for to talk to each other. Sometimes someone simply needs to own the process. Admittedly, “we sorted out the handover” is harder to dress up as a revolution. Still useful, though.

Start with what’s getting in the way

“We need to use more AI” doesn’t give a team much to work with. “Our customers wait too long for an answer” does. It gives you something to investigate: where the delay happens, what causes it and what a better result would look like.

My starting point is the problem and what it costs the business. Then I look at whether a clearer process or something the business already uses could solve it.

A task with clear, stable rules may suit rules-based automation. A reminder or an approval threshold doesn’t automatically need an AI agent. AI might earn its place when there’s varied information to interpret or a first draft that takes someone a lot of time to prepare. A workflow can combine both approaches. The decision is which parts belong where.

Take customer enquiries

Say your team handles emails about deliveries and returns. AI could help summarise a long exchange and draft a response using approved information, particularly when someone would otherwise have to piece the whole thing together manually.

Delivery information needs a reliable source, such as your order or tracking system. Refund limits need to follow the business’s rules. An unusual complaint may need someone with the authority to resolve it. A wrong answer is still wrong, however confidently it’s delivered. Getting it sooner doesn’t make it right.

And if customers keep asking because the delivery information on your website is unclear, there’s an obvious place to look. Fixing the page could prevent some of those enquiries before anyone needs to answer them.

Count the whole job

Producing a draft in ten seconds looks good in a demo. If someone then spends ten minutes checking and correcting it, that time needs to be counted. So does the work involved in connecting the tool to existing systems and keeping it working, including the exceptions someone still has to sort out manually.

Time saved doesn’t automatically mean money saved, either. Be clear about whether the team can use that time elsewhere, whether it reduces spending or whether it avoids a future cost. A collection of spare minutes scattered across the week may be harder to put to work than the business case suggests.

That alone doesn’t tell you whether the change is worthwhile. Faster customer resolution or more capacity in an overloaded team can be worth paying for. Just be clear about which benefit you’re claiming and how you’ll know it happened.

Give it one job and set the boundaries

Pick one recurring problem, give it an owner and agree on what improvement would look like. For customer enquiries, measure the time to resolution alongside the quality of the answers. If a faster response leaves the customer chasing a correction, count that additional effort too.

The controls should reflect what a mistake could do and how easily it can be detected and corrected. Suggesting a refund for someone to approve and issuing it automatically involve different authority. Decide what the system is allowed to do, when a person must step in and which errors should stop the trial.

“Someone will check it” needs more detail, too. Name the person, make clear what they’re checking and give them enough time to do it properly. They also need the authority to stop the process if something looks wrong.

Use what you learn to decide whether to expand, adjust or stop. Stopping is a reasonable option when the evidence doesn’t support proceeding. A well-designed trial can give you evidence to stop before taking on an ongoing cost.

There’s plenty of worthwhile work AI can help with. We’ll explore that here, alongside the practical decisions that make it work in a business. My preference is to be a bit choosy about where AI goes.

What’s one problem in your business that a better process could solve?

If you want to work through where to start in your own business, try my 3-Minute Discovery. You can also get in touch to discuss what’s getting in the way.