Most Companies Are Quietly Unplugging the AI They Just Launched

Here's a number that should make you sit up if you're about to put an AI in front of your customers: three out of four companies that already did it have had to pull it back down.

That's not a doom-blog statistic I'm massaging for effect. It comes from a survey this spring of more than 2,500 senior decision makers — the people who actually signed off on these projects — across ten countries. Seventy-four percent said they'd rolled back or shut off a live AI customer agent after it was already deployed. Not "thought about it." Did it. Yanked a thing they'd launched, in front of real customers, back offline.

And here's the part that stopped me cold. Among the companies that described their safety guardrails as fully mature — the ones who did the most homework, spent the most, built the most careful controls — the rollback rate was higher. Eighty-one percent. The straight-A students were pulling the plug more often than everyone else.

So either the careful companies are doing something wrong, or the thing everyone's selling you as the answer isn't the answer. It's the second one.

What's actually going wrong

When one of these projects dies after launch, it's almost never because someone forgot to add a guardrail. It's because the bot was bolted onto the front of a business that wasn't ready to stand behind it.

Picture the normal version of this. A company buys or builds a customer-service AI. It demos beautifully. Leadership, reasonably, starts thinking about the savings — if the bot handles most of the volume, you don't need as many people answering the phones. So the human team gets quietly sized down to match the new world. The bot goes live.

Then the bot hits the thing it can't handle. Maybe it's an edge case, maybe it's an angry customer, maybe the AI just confidently says something wrong. The conversation kicks back to the humans — except now there are fewer of them than there used to be, because everyone planned for a world where the AI was carrying the load. The queue backs up. Wait times balloon. The very same survey found that when these agents fail, the two biggest consequences are exactly that — support queues spiking and the brand taking a reputation hit. And a damaged reputation doesn't heal the moment you fix the bug.

That's how a project that looked great in the demo becomes a thing you're unplugging four months later.

The boring thing that actually predicts success

The most useful finding in that whole report wasn't about AI at all. When the researchers looked for what separated the companies whose AI stuck from the ones who had to retreat, the strongest predictor wasn't how much they'd spent, and it wasn't how mature their guardrails were. It was how well the AI was wired into their actual communications systems — the plumbing underneath.

That tracks with everything we see. The chatbot is the glamorous part. The unglamorous part — does it actually know your inventory, can it see the customer's real order history, does it hand off cleanly to a human with all the context instead of making the customer start over, does it fail gracefully when it doesn't know — that's the part that decides whether this thing survives contact with your customers. And that's the part that gets shortchanged, because it doesn't demo well and nobody's excited to pay for it.

You can buy the smartest model on earth. If it's stapled to the outside of your business instead of woven into it, you're going to be in the 74%.

When the honest answer is "not yet"

So here's the thing we'll tell you, and it doesn't bill many hours: a lot of companies asking us to build a customer-facing AI aren't ready to put one in front of customers yet — and the fix isn't a better bot. It's getting the systems underneath in shape first, and keeping enough humans in the loop that the bad day doesn't become a bad week. Sometimes that means a smaller, narrower tool than what you walked in wanting. Sometimes it means "let's wire up the integration and the human handoff first, prove that's solid, and then add the AI on top." Less exciting on a slide. A lot more likely to still be running next year.

If a firm tells you the bot is the hard part, they've got it exactly backwards. The bot is the easy part now. Everything it has to touch to be trustworthy is the hard part — and the rollback numbers are what happens when you skip it.

The question worth sitting with

Before you launch anything customer-facing, ask the unglamorous question instead of the exciting one: when this AI hits something it can't handle — and it will, in week one — what happens to the customer on the other end, and is there still a person there to catch them?

If you don't have a confident answer, you're not ready to go live. And that's a much cheaper thing to learn now than four months from now, with your name on it.