How to Adopt AI at Work Without Breaking Your Business

I called my doctor’s office the other day to leave a message. The new AI chatbot picked up and launched into a two-minute uninterruptable speech about how it was going to help me. When that finally ended, it started a one-minute one.

Then more after that. Five minutes every call, every time, before I could get to a human or even leave a message.

The old phone tree was clunky. This was worse. And somebody at that medical practice spent real money to make it worse.

That clinic isn’t the exception. It’s the pattern. Companies are rolling out AI everywhere, and a lot of those rollouts are quietly hurting the businesses paying for them.

The fintech giant Klarna laid off 700 customer service workers and replaced them with an OpenAI-powered chatbot, then started rehiring humans when the CEO publicly admitted the AI produced “lower quality” service. Same disease, different scale.

Not because AI is bad. Because nobody asked the boring question first: where does this actually belong in our work, and where does it ruin things if we put it there?

I’ve been around long enough to watch this happen before. Different technology, same mistake.

The short answer

Adopting AI at work without breaking your business comes down to four moves:

• ► Use AI to amplify the boring, repetitive work that drains your people, not to replace the judgment they get paid for.
• ► Identify the edge cases before you deploy, not after. For more, see six places AI will break your work. AI handles the routine. It breaks on exceptions.
• ► Keep a human in the loop wherever a mistake costs money or trust.
• ► Retrain your people on the new tools instead of replacing them. For more, see AI can write a business book. it can't write yours.. The augmented human beats the replaced one every time.

That’s the framework. The rest of this piece is forty years of receipts on why those four moves are the difference between AI saving your business and AI quietly killing it.

What most rollouts get wrong

The doctor’s chatbot failed because somebody confused two different jobs. AI is good at handling routine questions at scale. It is not good at knowing when to get out of the way and let a human take over.

The clinic didn’t think about the edge case, which in this instance was me. A patient who already knew what he wanted and just needed to leave a message. The system had no off-ramp, so every patient calling in got the long-form AI experience whether it served them or not.

That’s the whole disease in one phone call. Smart tool, deployed without thinking about where it belongs, ends up making things worse than what it replaced.

The same mistake plays out at the top of the Fortune 500. Klarna, the buy-now-pay-later fintech, fired roughly 700 customer service workers and handed their job to an OpenAI-powered chatbot. The CEO publicly claimed AI could do every job humans do.

The company reported around $10 million in initial savings and a market story that fueled their IPO. Then customer satisfaction collapsed on anything complicated, and the same CEO was on Bloomberg admitting the AI produced “lower quality” service and that Klarna was hiring humans back. The reversal cost more than the original savings ever delivered.

Same problem as the clinic chatbot, with two more zeros on the budget. Smart tool, deployed in a job it couldn’t actually do, with no humans in the loop where the customers needed them most.

I see the same mistake in writing. Someone hands me chapters their AI wrote and they want to know what I think. The chapters are technically fine.

Clean grammar, no spelling errors, every paragraph sits politely next to the last one. And they say nothing. Nobody is home.

The author published the machine’s first try because the surface looked clean, and now they have a book that reads dead. Same disease as the chatbot. Tool used in the wrong place.

Where AI actually works

Back in the eighties I led a project to automate the control system for a water district in a region that gets buried under snow every winter. The math was brutal. Tanks and pumping stations scattered across hundreds of square miles, and for a month at a time during heavy weather the remote sites were physically cut off from central command.

Roads closed. Lines down. Crews couldn’t reach the equipment.

The water had to keep flowing anyway.

What we built, well before anybody used the term AI, was a distributed system of remote controllers that could think for themselves. Each remote station learned the patterns from years of weather and flow data, knew what to do when the central system went dark, and knew its own limits well enough to fail safely when conditions went outside what it had ever seen. The whole thing was written in Pascal because Pascal solved the problem.

That project worked for one reason. We didn’t ask the algorithm to replace the people who understood water. We asked it to handle the routine when the people couldn’t physically get there.

The humans stayed in charge of the system. The system handled the grunt work and bought the humans time to deal with the parts only humans could deal with.

Forty years later, the right uses of AI follow the same rule. I worked with a client whose warehouse staff wear augmented reality glasses on the floor. A picker walks into the aisle and sees a glowing arrow that nobody else can see, pointing at the exact box they need.

They can wave a hand and “open” the box to see what’s inside without touching it. If they grab the wrong one, the system tells them in real time. Their pick rate doubled.

Nobody lost a job. The same number of people now do far more interesting work because the boring search-and-verify part is handled by the system.

That is what good AI adoption looks like. The humans got more capable.

The work got less tedious. The system handled what it’s good at and stayed out of what it isn’t.

Where AI breaks

Now the warning side. I’ve watched smart people get hurt because they didn’t understand where the machine fails.

AI fails on edge cases. Anything outside the pattern it learned. The chatbot in the medical office is a tiny version of this.

A self-driving car is the big version. When the bicycle wobbles unexpectedly, when the pedestrian steps off the curb early, when the traffic backs up for a reason the model has never seen, the system has to make a call it was never trained to make.

Most of the time it gets it right. Some of the time it doesn’t, and the consequences of that small percentage are enormous.

AI hallucinates. It invents facts confidently. For anyone publishing in their own name, this is brutal.

The machine will cite a study that doesn’t exist, attribute a quote to the wrong person, get a date wrong by two decades, and present all of it in clean, authoritative prose. If you’re not checking every claim, you’re going to put your name on a lie. That happens to writers daily right now, and most of them haven’t realized it yet.

AI drifts. Over a long task it loses the thread.

Whatever you asked it to do at the start bends slightly with every response, the way a tree bends in wind, and twenty pages later the document has wandered off the original ask. The longer the project, the worse the drift.

AI is shallow. It writes what most people would write about the subject. The surface looks intelligent.

The depth isn’t there. It hits the common facts, makes the common arguments, lands the common conclusion, and never says the one specific risky thing only someone who actually lived the work would know to say.

It can fake the texture of expertise. It can’t produce expertise.

And AI has no lived experience. It has read about love and grief and combat and divorce. It hasn’t felt any of them.

For fiction or memoir, anything that asks a reader to feel real emotion, this is the fatal limit. The machine can describe the shape of an emotion. It cannot write from inside one.

Why edge cases are the whole game

Run those failure modes at scale across a whole industry and you get the long-term version of the problem. I wrote a book on it called The Death of Thinking. The short version is what happens when we trust machines to do the parts of our work that used to make us sharper, and end up dumber on the other side.

Every example above comes back to the same problem. The thing that breaks an AI rollout is almost never the routine case. It’s the exception.

The blind pedestrian. The patient who needs to leave a message.

The chapter that needs a moment of real grief on the page. The interview transcript that contains the one detail that should change the whole story.

The work of adopting AI well is the work of mapping the edge cases before you deploy. Where does this system meet a human who doesn’t fit the pattern? What does it do then?

Who catches it? Where does the cost of a single failure land?

If you can’t answer those questions, you don’t have an AI strategy. You have an AI deployment, and you’re going to find out the hard way.

The retraining piece

The other half of this is what AI does to the people doing the work. This is part of my Technology of Writing Hub, where I collect everything on the topic. The honest answer is that some jobs are going away. Not all of them.

Not most of them. But the parts of jobs that were always boring and repetitive will get automated, and the people who only did those parts will need new work.

This isn’t new. It’s the industrial revolution again, faster. Blacksmiths didn’t all become unemployed when factories arrived.

They retrained and moved into the parts of metalwork that machines couldn’t do. The ones who refused to adapt had the hardest time. The ones who learned the new tools came out ahead.

I have a client in his seventies who runs AI on his own manuscript drafts. He asks the machine to flag weak chapters and find inconsistencies. He doesn’t let it write his book.

He uses it to make himself a sharper reviewer of his own work. He’s better at writing now than he was five years ago, partly because of the tool.

He’s not afraid of being replaced. He’s busy being augmented.

That’s the model. Whatever your age, whatever your industry, the move is the same.

Find the boring repetitive parts of your work and put AI on them. Free up the time those parts used to take. Put the time into the work only you can do.

Get better at the human part. That’s the augmented human, and that person is going to be more valuable in five years, not less.

What to actually do this week

If you run a business or own an AI rollout, here’s the short list.

Pick one boring task that drains your people. One. A repetitive, time-consuming, emotionally numbing task.

That’s your first AI use case. Not the customer-facing thing. Not the public-facing thing.

The internal grunt work nobody likes doing. Get a win there before you touch anything risky.

Before deploying anything that touches a customer, write down every edge case you can think of and ask what the system does in each one. If you can’t answer for the top ten, you’re not ready to deploy.

Wherever a mistake costs you a customer or could trigger a lawsuit, keep a human in the loop. Not a human reviewing logs after the fact. A human with veto power in real time.

Train your people on the tools instead of around them. Show them what AI is good at, what it isn’t, and how to be the boss of it instead of being intimidated by it. The companies that win the next decade are the ones with augmented humans, not the ones that fired everyone and trusted the chatbot.

That’s the framework. The rest is showing up every day and doing the work of integrating it carefully, one process at a time, with your eyes open.

The takeaway

Adopting AI at work isn’t about whether to use it. That argument is over.

The companies that figure out where AI amplifies their people and where it doesn’t belong are going to outrun the ones still arguing whether to allow it at all. And the people who learn to be augmented by it are going to outrun the people who treat it as the enemy.

The doctor’s chatbot will get fixed eventually, when enough patients complain and the practice realizes that making it harder to leave a message wasn’t the upgrade they thought they were buying. The warehouse with the AR glasses already figured this out. The water district figured it out forty years ago in Pascal.

The pattern is older than the buzzword. The work is figuring out which side of it you’re on.

Frequently Asked Questions

Related: why most AI rollouts fail