Software has been able to seem intelligent for a long time. The abandoned-basket email that lands three hours after you closed the tab, the “customers also bought” strip that knows you’ll want the filter to go with the hoover, the chatbot that greets you by name and seems to follow your problem right up to the moment you ask something that isn’t on its menu. Each of these feels as if something is thinking on the other side of the screen, and each is a rule that a person wrote down in advance. Some of them are older than the people now selling them to you as AI.

The trick itself is older still. In 1966 Joseph Weizenbaum at MIT built ELIZA, a program that played a therapist by spotting keywords in whatever you typed and turning your sentence back into a question. Tell it you were unhappy and it would ask why you were unhappy. There was no understanding anywhere in it, just pattern-matching and a few hundred canned phrases, and yet his own secretary asked him to leave the room so that she could talk to it privately. Weizenbaum spent the rest of his career unsettled by how readily people credited a script with a mind. Sixty years on, the people writing the brochures have worked out the same thing he did, and they’ve found it sells.

I’d call what ELIZA did, and what most business automation still does, algorithmic intelligence. If this, then that. Decision trees, triggers, scheduled jobs, scoring rules. Somebody sits down, anticipates every situation the software will meet, and writes out what it should do in each one. Done well it’s quick, consistent and cheap to run, and for the situations it was built for it will beat a person every time, which is exactly why businesses have been buying it for forty years. In 1980 Digital Equipment Corporation put an “expert system” called XCON to work configuring its computer orders; it grew to about 2,500 rules and was credited with saving the company millions of dollars a year, and it was called artificial intelligence at the time. The limit hasn’t moved since. The moment a situation arrives that nobody wrote down, the system has nothing to offer, and the chatbot asks you to rephrase your question.

What has changed with the large language models is that nobody writes the answer. The developers set instructions and guard rails, but the response to your particular question is produced on the spot, not looked up. That’s the line between the two, and it’s a much cleaner one than the marketing lets on: could a person have written this behaviour down in advance? If they could, it’s an algorithm, however impressive it looks. If the system is producing sensible responses to situations nobody anticipated, it’s AI, with the flexibility that brings and the unreliability that comes with it, because an answer nobody wrote is also an answer nobody checked.

The confusion costs money in two directions. The first is the one most of us have met: the booking system that has had “smart scheduling” since 2015 relaunches as AI-powered on a dearer tier, and what you’re paying extra for is a rule you could have described yourself over coffee. I wrote last month about how to test whether a feature’s AI is doing any work at all, and the test applies here too. The second direction is quieter and I think more damaging. Owners look at the automation they’ve built up over a decade, the CRM triggers and the Zapier workflows and the auto-replies, decide that this counts as “doing AI”, and conclude either that AI has done nothing much for them or that they’re already ahead of the field. Both conclusions rest on the same mistake, and the second one shapes where the next year’s investment goes.

None of which makes rebadging dishonest in every case. Sometimes a vendor has put a model behind the old rules, and you can tell, because the product starts coping with requests it used to fall over on. So when the badge appears, ask what the thing can do now that it couldn’t do before, and ask to see it handle a case that isn’t in the demo script.

Then put the same question to your own systems.

The rules we’ve accumulated are worth keeping, and they remain the right tool wherever we want the same answer every time. They just aren’t intelligence, and calling them that will only stop us seeing what the real thing could do for us.