Nerthos home Join the beta

Trends · 5 min read

Sold as a helper, bought as a replacement

AI was sold as a way to get more done. Plenty of companies heard "with fewer people." What Ford and Klarna learned, what the research shows, and a Monday task.

Open the image full size

Out there

You've seen this meeting. Maybe you've sat in it. Somebody puts up a slide with two columns: what AI costs a year, and what the people cost a year. The math looks great. It always looks great on a slide. And nobody asks the one question that matters, which is what those people know.

Maybe I'm wrong, but I think what happened was this. AI got sold to get more done. A lot of leadership teams heard "get it done with fewer people" and stopped listening right there.

Klarna went first, and loudest. Early 2024, they announced their AI assistant was doing the work of 700 customer service agents. Hiring froze for over a year. Headcounts dropped by about a fifth, mostly people leaving and nobody replacing them. Then May 2025 rolls around and the CEO is telling Bloomberg the AI-first approach gave customers lower quality service. So now they're hiring humans again. Quote: "investing in the quality of human support is the way of the future for us."

I don't bring that up to pile on. I respect a CEO who says "we got that wrong" out loud. Most don't.

And Klarna has company. Gartner thinks half the companies that cut customer service staff because of AI will be hiring for that same work again by 2027. Different job titles. Same jobs.

Ford's story is the one I keep coming back to. They expected AI tools to catch quality problems if you fed them the design requirements. The tools didn't. Ford's VP of vehicle hardware engineering said the tools didn't have the experience of the veteran technicians, and a lot of those technicians had already left before anybody captured what they knew. So, Ford brought in around 300 veteran engineers to sit in weekly design reviews and catch problems before they hit the factory floor. This year Ford finished first in J.D. Power's Initial Quality Study for the first time since 2010, and their CEO says lower warranty and recall costs are helping the bottom line. Turns out experience still counts for something. I'm choosing to find that encouraging, for personal reasons.

None of this means AI is useless. I use it. I built it into my own product. The point is that the value was sitting in the people the plan left out.

The research backs that up, and it's specific. Economists at Stanford and MIT followed 5,179 support agents after they got an AI assistant. They resolved 14 percent more issues per hour. The newest agents? 34 percent more, because the AI was basically handing them the habits of the best people on the team. Customers were happier. Fewer agents quit. Nobody got replaced and everybody got better. That's the version I'd want to buy.

Harvard and BCG tried it with 758 consultants. On work an AI is good at, the ones using it finished 12 percent more tasks, 25 percent faster, and about 40 percent better. But on work just outside what an AI handles well, they were 19 percent less likely to get the right answer. Which is kind of my whole point. The tool is great right up until it confidently isn't, and you need a person who can tell the difference.

Then MIT looked at company AI projects last year and found 95 percent of pilots never made it past the pilot. Their explanation was a learning gap, tools that never got worked into how people do their jobs. Ninety-five percent. I'd call that a "nobody asked the people doing the work" problem.

In here

I'll admit up front that I have a horse in this race. Nexus has an AI built in, so take what follows with whatever size grain of salt feels right.

In more than 25 years of IT, I've noticed that the work people complain about is rarely the work they're good at. It's the typing up, the sorting, the third reformat of the same status update. That's where I wanted an AI in Nexus. The decisions stay with people.

So, it starts switched off. An admin turns it on for the workspace, and each person can still hide it for themselves. When it's on, it drafts and you decide. Describe a task in plain words, and it writes one up for you to check before anything is created. Open a contract and ask for a summary, and it writes up the risks, obligations and renewal terms, labeled as AI and flagged once the contract changes. Contract text doesn't go to the AI at all unless your workspace turns that on. Rate your skills, and it can suggest a level, but you or your manager make the call. Nothing in your workspace changes until a person accepts it.

That's Klarna's lesson built into the product: the judgment about whether the work is good enough belongs to the person doing it.

Ford's lesson is the one I think about most, though. The AI couldn't learn from those technicians because their knowledge had already walked out the door. An assistant can only help with what someone took the time to write down. That's part of what Nexus is for: the meeting notes on the project, the decision in the record, the skills on the org chart, kept where the next person, and the AI helping them, can find it.

What to do Monday

Ask one person on your team a question no vendor can answer for you: which hour of your week is purely busy work?

Write the answer down. That's your first place to try AI, with the person who does the job deciding whether the result is good enough. Track two things for a month: the time it saves, and whether the work has gotten better or worse. Leave headcount out of that first conversation entirely.

Then ask your most experienced person to write down one thing only they know how to do. Ford would have paid a lot for that page.

Sources

Liked this one?

Get the next one in your inbox. Notes from the build, what changed in Nexus, and one trend worth your time.

Want to try Nexus yourself? It costs nothing to join the beta