What AI still can't do in 2026 (and what I spend my meetings explaining)

I spend half my meetings explaining what AI doesn't do. It always surprises people, coming from someone who installs it for a living. But that's exactly my job: saying where it genuinely helps, and where bringing it in would be a mistake. Many owner-managers arrive thinking it will do everything. I'd rather they leave with the accurate picture.

In 2026, it writes better than ever, it summarises, translates, sorts, tidies up. On all of that, it's impressive and I use it every day. But there are three things it still can't do, and those are precisely the three that decide whether a project holds up or falls apart.

1. It doesn't know when it's wrong

I'd set up a system for price monitoring that would read dozens of competitor pages and pull out the rates. It had been running quietly and well for a while. Then several of those pages changed structure. The system, for its part, carried on as if nothing had happened, returning false figures without lifting a finger to flag it. No alert, no hesitation. I had to notice the anomaly myself, work out what had changed, and rebuild the feeds so everything worked properly again.

The lesson applies to every use case: AI will never tell you "I'm not sure any more here". It executes with exactly the same confidence whether it's right or wrong. A machine doesn't doubt itself. It's up to you, or someone working for you, to doubt on its behalf and to check what matters at the source.

And it isn't limited to the big systems you keep an eye on. Ask a mainstream assistant for a figure, a legal reference, a service's opening hours: it will answer with the same self-assurance whether it's correct or entirely made up. It has mastered the text. The fact, it guesses, and it sometimes fabricates one from thin air without the slightest sign of hesitation. That's the number one trap everyone falls into at the start: copying a figure straight out of a chat without ever double-checking it.

A machine doesn't doubt itself. It's up to you to doubt on its behalf.

2. It can't read the person in front of it

An unhappy, worried or hurried customer, you can hear it in the first word, in a silence, in a tone that tightens. A tradesperson spots in three seconds that they've got someone in front of them who needs reassuring before any talk of technicalities. AI doesn't. It processes the request, not the person. That's why, on projects where I could have automated after-sales support, I deliberately kept a human at 100%, even though the automation was ready. Someone with a problem wants to feel a presence on the other end, not a fast, polite reply that misses their worry entirely. Automating that moment saves a bit of time and loses the customer.

3. It carries no responsibility at all

The day a decision goes wrong, AI doesn't call the customer back, doesn't issue a refund, doesn't lose sleep over it. You do. So the decision has to stay yours. I apply one simple principle: AI proposes, a human approves, especially when it's irreversible or involves money. On my own system, I deliberately capped the part that could change prices: it suggests, I decide. That's not old-fashioned caution, it's the only division of labour that actually holds up. The machine takes the repetitive part, you keep everything with consequences.

AI is often sold to us as an intelligence. It's a formidable executor, not a conscience. It doesn't know when it's wrong, it doesn't sense who it's talking to, it doesn't own anything it produces. The day a provider swears to you that it will handle everything on its own, they're selling you the dream, not the reality. The right approach fits in one sentence: give it exactly what it does better than you, and jealously keep the rest. That's the whole job when we install automation built onto your existing tools: we fix and frame the process before handing it to the machine. And if your project touches on monitoring your prices or your figures, that's exactly where the safeguard matters most.