What can go wrong with AI automation, and what happens when it breaks?

Nobody selling automation wants to lead with the failure modes, which is exactly why you should ask about them. Plenty can go wrong. The honest version is that the difference between a minor blip and a lost customer is decided before launch, in how the thing is built.

Four failures cover most of what we see.

You automated a broken process. Automation is an amplifier. If your quoting process loses jobs because nobody confirms the site visit, automating the quote email just loses jobs faster and more politely. The process has to work manually before it deserves a machine.

The data was a mess. Duplicate contacts, stale prices, three different spellings of the same customer. AI does not clean this up for you, it builds on top of it. Feed a follow-up system a contact list where half the emails are dead and it will diligently follow up with nobody.

The AI guessed. A chatbot asked a question it cannot answer will often answer anyway. Wrong prices, invented delivery dates, made-up policy. The fix is structural, not hopeful: the system should only state facts it can look up, and hand anything else to a person. If a supplier cannot explain how their bot is stopped from guessing, it is not stopped.

There was no fallback. This is the Saturday problem. Most automations are chains of third-party tools, and every link is something that can go down, change its format, or quietly start rejecting requests. A chain with no fallback fails silently. The form looks fine, the enquiry evaporates, and you find out when the customer rings someone else.

So what happens when it breaks on a Saturday?

That depends entirely on whether the builder planned for it. The pattern we use on every system is capture first, clever later. A lead form saves the raw submission to the database before anything else runs. Then the enrichment, the CRM push, the auto-reply. If any of those later steps dies, the enquiry is still sitting safely in the database and an alert goes to a phone. Worst case is a reply that goes out later than usual, not a customer who got silence.

A fragile build wires it the other way: form straight into a chain of tools, no record kept, no alert configured. Both versions look identical in the sales demo. Only one of them survives a bad weekend.

I build on the assumption that something will fall over, because something always does. APIs go down, emails bounce, a supplier changes a format without telling anyone. The system's job is to make that boring.

What should you ask before you buy?

Make a supplier walk the chain with you and answer, step by step: what happens when this one fails? Where do errors go, who gets told, and how quickly? Can a human pick up the work mid-process? What does the customer see while it is down? Vague answers here are the real red flag, not the failure itself. Everything breaks eventually. Good systems break loudly, in front of the right person, with the work intact.

If lead handling is the bit you are worried about, our guide to AI for lead follow-up shows what the capture-first pattern looks like in practice.