Just because AI can, doesn’t mean it should

I met an operations manager from a construction company this week. Let’s call him Darren (because that’s his name). Darren had been a prospect for a few weeks and we’d sussed a time to meet at his office. Unlike a normal prospect, however, Darren had tried automation before, and tried it on a bunch of different tasks across different software. He had it parsing CAD pdfs, drafting quotes, scraping pricelists and drafting emails, among a bunch of other tasks. He hated it. He turned it off pretty soon after turning it on.

Here’s where Darren went wrong: he had jumped straight to using AI for the automations. He had MS Copilot doing all sorts of stuff purely from chat prompts and Claude acting as an integration between the different pieces of software. The software wasn’t actually integrated, it was just pretending to be (using AI). It was incredibly expensive, more than a new employee he reckoned. It also didn’t work that much - most of the time, AI would complete the tasks correctly a couple of times and then find something to cock up on the next run. Darren had identified some really good tasks to automate, put in the automations, and somehow ended up spending more time fixing the automations than working on business growth like he intended.

AI is incredibly powerful, and it can do anything online, but it’s not usually the best solution. If a task happens the same every time with different data, you should automate it using programming and/or middleware instead. This is a fraction of the cost of AI tokens and you can trust it to do the same thing every time it runs.

So, if we shouldn’t use AI for recurring work, where is it best used?

Firstly, what you’re probably doing already - use it as an advisor. Bounce ideas off your chatbot of choice, make it suggest things to automate, and make it come up with the best solution for the automation ideas. It’s also excellent for grunt work: writing code, debugging code, spreadsheet formulas, transcribing, data manipulation and more. Stuff that used to take days of messing around can now be done in an hour.

Furthermore, it can build apps. You might hate your CRM (most people do). Try asking Claude Code or Lovable or any other AI app builder to make you a new CRM, with all the good bits of your current CRM and all the bits your current CRM is missing. It’s surprising and exciting how much it can build, entirely referencing plain text prompts. The new app won’t have any existing integrations, but it will have an API endpoint to enable them.

AI also does a great job of web scraping, but keep your prompts as specific as possible and make sure to double-check the links it references. I find it too much of a yes-man; it’ll explain how to do something, I’ll explain that isn’t possible, then it’ll go round in circles on a handful of ideas without just telling me that the thing isn’t possible. Ask it where it’s getting its information from, and have a look at the source to cut down wasted time on misunderstanding a problem.

Darren was pretty switched off the idea of automation, because his idea of automation was completely reliant on AI prompts. AI is capable of anything, but that doesn’t mean you should use it for everything.

-Fred