Ava’s older cousin Marcus worked the customer support phones at a big software company. He’d been there three months, and he hated it.

“Every call is the same nightmare,” he told Ava over dinner. “Someone calls in furious because their printer won’t print, or their invoice looks weird, or their account got locked. I’m flipping through a 400-page manual trying to find the right answer while they’re yelling at me. The people who’ve been there ten years? They just know. I’m drowning.”

A month later, Marcus came back for dinner — and he was a different person.

“They gave us an AI assistant,” he said. “It listens to the call, looks at the customer’s account, and suggests what to say. It’s like having the best person on the team whispering in my ear. My numbers shot up. I’m closing tickets faster than guys who’ve been there for years.”

Ava asked the obvious question. “Did it help those experienced guys too?”

Marcus shrugged. “Honestly? Not really. They were already good. The AI didn’t make them much better. It made me much better.”

That right there — Marcus’s shrug — is one of the most important things happening in the AI agent revolution. And it’s the opposite of what most people expected.

The Surprise Nobody Saw Coming

When AI tools first showed up in workplaces, almost everyone assumed the same thing: the smartest, most experienced workers would get even further ahead. They’d use the new tools best. The gap between top performers and everyone else would grow into a canyon.

Then the data came in. And the data said: nope.

In 2023, three researchers — Erik Brynjolfsson of Stanford, Danielle Li of MIT, and Lindsey Raymond — studied 5,179 customer support agents at a Fortune 500 software company over the course of a year. Some agents got an AI assistant that read along with their customer chats and suggested responses in real time. Some didn’t. The researchers measured how many customer problems each agent could solve per hour.

Here’s what they found:

• The agents who got AI help resolved 14% more issues per hour on average.
• But that 14% average hid something stranger. Novice and low-skill workers improved by 34%. Experienced, top-performing workers improved barely at all.

The AI didn’t lift the rocket scientists higher. It pulled the rookies up toward the level of the veterans.

The paper is called “Generative AI at Work,” and it’s become one of the most cited studies on this topic. Its core finding has a name now: AI agents are equalizers. They flatten the gap between the best and the rest.

The Training Wheels That Think

Think back to learning to ride a bike.

If you had training wheels, you remember the feeling. They didn’t pedal for you. They didn’t steer for you. They didn’t decide where to go. All they did was keep you from falling. They held you up in the wobbly moments while your body learned what balance felt like. They turned a terrifying impossible task into something you could practice without bleeding.

The fastest kid on the block didn’t need training wheels — they were going to figure out balance no matter what. But for the rest of us? Training wheels were the difference between giving up and learning to ride.

An AI agent at work is like training wheels that think.

It doesn’t do your job for you. It catches you when you’re about to make a mistake. It whispers, “Hey, the answer to this kind of question is usually X.” It pulls up the right document before you finish typing the search. It reminds you of the policy you forgot. For the new person, that’s the difference between drowning and swimming. For the veteran who already knows everything? The wheels are barely touching the ground.

Marcus’s “best person on the team whispering in my ear” — that’s exactly what’s happening. The agent has absorbed the patterns of the best workers and is sharing those patterns with the ones who need them most.

The Headline Numbers

Studies and company reports keep finding the same pattern. Here are the most-cited figures — and where each one actually comes from.

Finding	Number	Source
Customer support productivity boost (average)	+14%	Brynjolfsson, Li & Raymond, Generative AI at Work, NBER Working Paper 31161 (2023); published in the Quarterly Journal of Economics, 2025
Same study — novice and low-skill workers	+34%	Same paper, same 5,179-agent study
Per-worker productivity boost in human-AI teams vs. human-human teams	+50%	Ju & Aral, Collaborating with AI Agents: A Field Experiment on Teamwork, Productivity, and Performance, arXiv 2503.18238 (2025); 2,234-participant field experiment
IBM productivity gains from AI across 70+ business areas	$3.5 billion (by end of 2024)	IBM, Embracing the future of HR by becoming an AI-first enterprise (2025); CIO magazine, “IBM agent AI delivers $3.5 billion in productivity impact” (2025)
IBM workforce size when this was achieved	~270,300 employees	IBM 2024 Annual Report
Routine HR questions handled by AskHR with no human involved	94% (of 11.5M interactions in 2024)	IBM, Embracing the future of HR (2025)
Speed boost for managers doing HR transactions at IBM	75% faster	IBM, Embracing the future of HR (2025)
Companies using AI agents reporting higher productivity	66%	PwC AI Agent Survey (May 2025), 308 US executives
…reporting cost savings	57%	Same PwC survey
…reporting faster decision-making	55%	Same PwC survey
Developers using or planning to use AI tools in their work	84% (up from 76% the year before)	Stack Overflow 2025 Developer Survey

A few honest caveats before we go further:

• These numbers come from different industries, different tasks, and different AI tools. They don’t all add up into one tidy claim.
• The gains depend heavily on the type of work. Repetitive, knowledge-heavy tasks (like answering common customer questions) see big lifts. Highly creative or deeply technical tasks often see smaller, messier results.
• Bigger numbers in press releases often shrink under independent study. Trust the peer-reviewed papers more than the marketing slides.

Even with all those caveats, the direction is unmistakable. AI agents are making real, measurable differences in real workplaces — and they’re making the biggest difference for the people who started out behind.

Why the Less-Experienced Gain More

Why does this happen? Why doesn’t the AI just turbo-charge the experts?

A few reasons, all pretty intuitive once you hear them.

The experts already know the answers. A senior customer support agent has handled thousands of “my printer won’t print” calls. They don’t need an AI suggesting what to say — they could write the suggestions themselves. The AI’s whispered advice is just confirming what they were going to do anyway. There’s not much room left to improve.

The novices don’t know the answers. They’re flipping through manuals, asking coworkers, putting customers on hold. The AI hands them, in three seconds, what would have taken them ten minutes to dig up. The gap to close is huge, and the AI closes most of it.

The AI learned from the experts. This is the beautiful part. The AI assistants in these studies were trained on transcripts of the best workers — the people whose chats ended in happy customers. So the AI is essentially bottling up what the best workers do, and pouring it into the ear of the ones still learning. It’s expert knowledge, redistributed.

Confidence matters more than you’d think. A new worker hesitates. They second-guess. They put customers on hold to double-check. The AI doesn’t just give them the answer — it gives them the certainty to act. Cut out the hesitation, and the work speeds up dramatically.

The researchers found something else too: customers were more polite to workers using the AI and less likely to demand a manager, and those workers were less likely to quit. When you stop drowning, you stop wanting to escape.

A Bigger Study, A Bigger Number

In 2025, another team of researchers ran an even larger experiment. They recruited 2,234 people to work on a real project — designing ads for a think tank — and split them into two kinds of teams: human-human pairs, and human-AI pairs (where a person was paired with an AI agent instead of another person).

The results were striking. Workers paired with an AI agent produced 50% more ads per person than workers paired with another human. Their conversations were more focused on the task — about a quarter more task-oriented messages — and they made 73% fewer direct edits to the ad text themselves: the AI handled the typing while the human steered.

The study, Collaborating with AI Agents, came out as a working paper (arXiv 2503.18238) in March 2025. The researchers were careful, though, about what their numbers do and don’t mean. Each human-AI team had half as many people, so whole teams with an AI partner actually turned in fewer ads in total than pairs of humans. The per-person productivity went up; the per-team output went down. In other words: the AI made each person more productive, but it didn’t fully replace a second person. Whether that’s good news or bad news depends on whose job we’re talking about.

Real Companies, Real People

This isn’t theory. Companies are reporting real money saved and real time freed.

IBM’s AskHR — the same AI assistant we met in Chapter 1 — handled more than 11.5 million employee interactions in 2024. Ninety-four percent of those got resolved without a human HR person ever stepping in. That left IBM’s actual HR specialists free to focus on the hard cases — promotions, conflicts, complicated benefits questions — where human judgment really matters. IBM’s leadership credits AI tools across the company with about $3.5 billion in productivity savings by the end of 2024. Its managers can now complete HR transactions like transfers and promotions about 75% faster than before.

PwC, one of the world’s biggest consulting firms, surveyed 308 senior US executives in April 2025. Two-thirds of the companies using AI agents reported higher productivity. More than half reported cost savings. More than half reported faster decisions.

Software developers — the people who write the code that runs everything — are some of the heaviest users. In Stack Overflow’s 2025 Developer Survey, 84% of developers said they use or plan to use AI tools in their daily work, up from 76% the year before. (Interesting twist: their trust in AI accuracy actually dropped over the same period. We’ll get to that in the next chapter.)

The pattern is the same across all of these stories: agents aren’t replacing the humans. They’re amplifying them — and amplifying the ones who needed help the most.

What This Means for You (and for Marcus)

If you’re someone who’s just starting out — at anything — this is the best news in the book. The tools that used to be available only to the experienced are now available to you on day one. The decade of experience that used to separate the new hire from the veteran can be partly closed by an AI agent whispering in your ear.

If you’re someone who’s already an expert, this is interesting news too. Your unique advantage — your hard-won judgment, your gut instinct for the weird cases — is still yours. The agents can teach the basics. They can’t yet teach the wisdom. But they can free you from the boring parts of your job so you can spend more time on the parts that need you.

And if you’re a parent or a teacher? The shift is that the floor is rising. The minimum competence in many jobs is going up, fast. Schools and workplaces that figure out how to teach kids to work with agents are going to launch their kids further than ever. The kids who learn only to compete with agents are going to struggle.

Marcus, by the way, got promoted six months later. He’s training new hires now. He tells them, “Don’t be afraid of the AI. Use it like training wheels. Eventually you won’t need them as much — but they’re how you get going.”

Key Takeaways

• AI agents tend to help the least-experienced workers the most — they close the gap between novices and veterans, rather than widening it.
• The most famous study (Brynjolfsson, Li & Raymond, 2023) found a 14% average productivity boost in customer support — but a 34% boost for new and low-skilled workers, and almost nothing for top performers.
• Real companies are reporting real money saved. IBM credits AI with about $3.5 billion in productivity savings by the end of 2024. PwC finds two-thirds of AI-agent-using companies report higher productivity.
• The training-wheels analogy fits. Agents don’t do the job for you. They keep you from falling while you build skill.
• The biggest caveat: results vary wildly by task, industry, and tool. Anyone selling you one universal number is oversimplifying.

---

All of this sounds incredible — almost too good. Cheaper. Faster. More fair. Beginners catching up to veterans. So what’s the catch? Because Ava has been wondering it the whole time, and you probably have too: what happens when an agent gets it wrong? And they DO get it wrong, sometimes spectacularly. In the next chapter, we’re going to look the dark side straight in the face. The agents that hallucinated. The agents that broke things. The agents that lied. Because if we’re going to live with these helpers, we’d better know exactly how — and how often — they let us down.