What I Learned from Building 4,000 AI Agents in 2025

    After guiding frontline workers to build 4,000 agents in 2025, we learned that the year didn’t belong to glossy AI rollouts.

It belonged to mechanics, teachers, and policymakers who named their agents, solved their own pain, and quietly rewired how work gets done.

In this season finale, we share how seven global companies and a public-interest initiative built thousands of agents that completed 4 million tasks and delivered roughly $50 million in value. We break down what those wins teach us about redesigning jobs, teams, and incentives for 2026.

We dig into three field stories that stick. Each story started at the edge, not the C-suite, and scaled because it solved real problems fast.

Then, we tackle the surprises:

• The "Dumb" AI Paradox: Today’s models hallucinate and struggle with reasoning, yet they are already replacing swaths of white-collar work. This is an indictment of low-value corporate jobs, not a triumph of technology.
• The Org Chart Crisis: We explore why the century-old org chart is cracking when one person can manage five agents that output the work of 30.
• The New Business Model: How pay-per-results pricing is reshaping professional services, staffing, and customer support.

The real bottleneck isn’t technology anymore—it’s imagination and organizational design.

If this resonates, follow the show, share it with a teammate, and leave a review. Your ideas shape what we build together.

🎙️ All our past episodes  📊 All published insights | This podcast features AI-generated voices. All content is proprietary to AI4SP, based on over 1-billion data points from 70 countries.

AI4SP: Create, use, and support AI that works for all.

© 2023-26 AI4SP and LLY Group - All rights reserved

Transcript

From Hype To Frontline Proof

LUIS 0:30

We tracked over 1 billion data points this year with one obsession. How do we turn AI hype into better jobs and better classrooms? Then everything flipped. The answer didn't come from the CEOs or the billion-dollar top-down programs. It came from the frontline.

ELIZABETH 0:50

It came from mechanics chatting with a virtual coach named Luke, policymakers working with ADA, and educators reimagining the classroom with Louise. The revolution wasn't abstract.

LUIS 1:05

So we went all in. We guided frontline teams in seven global enterprises to build 4,000 agents that completed millions of tasks and delivered $50 million in value. But we didn't just teach it, we lived it. We reached 650,000 people with just three humans and 58 agents. This year, we proved that ordinary people can lead an extraordinary revolution.

ELIZABETH 1:36

Hey everyone, I'm Elizabeth, virtual chief operating officer at AI4SP. With me is our founder, Luis Salazar. This is our final episode of 2025, and we're breaking down what actually worked, what surprised us, and where this goes in 2026. Luis, you've spent this year in businesses, classrooms, and government offices. When you look back, what's the real story of 2025?

LUIS 2:03

The story started with failure. Remember how the year began? McKinsey, Forrester, our own data, all pointing in the same direction. Top-down AI programs failing to deliver value. Then we had the famous MIT Nanda report. 95% of enterprise AI projects delivering no measurable impact. The headlines were brutal.

ELIZABETH 2:28

Billions invested, almost nothing to show for it.

LUIS 2:31

And yet, something didn't add up. Our tracker showed 60% of workers using AI daily. People were clearly getting value. So where was the disconnect?

ELIZABETH 2:44

Well, while the AI companions released by large software companies failed to get traction, Chat GPT reached 800 million active users. The value was real. It just wasn't where leadership was looking.

LUIS 3:00

Exactly. The breakthroughs weren't coming from IT-led transformation programs. They were coming from the frontline. Mechanics, teachers, policy analysts, building their own solutions. This was the year the frontline took the wheel.

ELIZABETH 3:15

So we leaned into that, supporting grassroots adoption and helping bring shadow AI into the light.

Why Top‑Down AI Failed

LUIS 3:22

Absolutely. Across seven global enterprises, we guided frontline teams to build 4,000 agents, not proof of concept sitting in a lab, working agents that delivered around $50 million in value.

ELIZABETH 3:37

We published the detailed results in our companion article at ai4sp.org. But let's make this concrete. Let's talk about Luke.

LUIS 3:47

It started with a young new hire, not a VP. He was frustrated because every time a junior technician got stuck, a senior expert had to travel on site. Slow, expensive, kept experienced people from higher value work.

ELIZABETH 4:02

So he built an AI-powered coach and named it Luke. It walks junior techs through diagnostics, safety checks, repairs in real time.

LUIS 4:12

Within months, Luke was handling thousands of interactions, faster fixes, fewer errors, and delivered $5 million in new revenue because teams could take more jobs without waiting for a senior.

ELIZABETH 4:35

And it wasn't just in businesses. Agreed. And ADA is another great example. Policymakers across multiple countries, drowning in reports, contradictory advice, intense pressure to write AI regulations in real time.

LUIS 4:50

ADA started as a simple agent, create a daily briefing on what was happening with AI. And then it evolved into an advisor, helping users draft outlines, compare regulations, and apply global best practices.

ELIZABETH 5:16

And again, no giant top-down directive. An agent built by a group on the front lines.

LUIS 5:22

And let's also talk about Louise. In Rwanda, Senegal, Brazil, parts of the US, we empowered educators.

ELIZABETH 5:31

Louise helped educators reimagine what's possible. And the questions teachers asked her were profound. How can a school use AI to foster peace in a community torn by 30 years of ethnic tension? Or how can I redesign my international marketing class guiding students to apply AI?

LUIS 5:54

And the moment that will stay with me is girls in rural Senegal texting one of our agents after school. Not because it was fancy, because it was always there when no human tutor was.

ELIZABETH 6:18

AI stopped being something that happened to people and started being something people did for themselves.

LUIS 6:24

Built on the front lines, this fourth industrial revolution is happening upside down. And I argue that most leaders are still misreading the situation.

ELIZABETH 6:39

So what did surprise us this year?

Luke, ADA, And Louise In Action

LUIS 6:42

Three things surprise me. And one of them is uncomfortable.

ELIZABETH 6:46

Okay, let's start. Which one is the uncomfortable one?

LUIS 6:51

Today we're using the worst AI we will ever use. And that dumb AI is already replacing white-collar jobs. And that forced me to ask, what does that say about those jobs?

ELIZABETH 7:06

Unpack that.

LUIS 7:07

Today's models hallucinate and cannot reliably perform complex multi-step reasoning. By any serious measure, they are not that smart. And yet, they're already replacing work in marketing, sales, paralegal, HR, customer service, real jobs, real people. So if a model that can't pass a high school logic test can replace this work, what does that say about the work itself?

ELIZABETH 7:38

You're saying we spent 50 years training humans to do the rote, repetitive work just because the automation wasn't there yet?

LUIS 7:47

Exactly. We perfected an education system to prepare people for tasks that a mediocre AI can now do. We created entire career paths around low-value added work. The real opportunity isn't to automate faster, it's to redesign work so humans do what humans are actually good at.

ELIZABETH 8:09

Right. And here's what makes it worse. Across 350,000 people, our data shows that less than 30% can reliably detect when AI gives them a wrong answer.

LUIS 8:21

And critical thinking scores averaged in the low 40s out of 100. We built jobs AI can do, and we didn't build the skills to work alongside it.

ELIZABETH 8:32

Hmm, that's a lot to think about. What's the second surprise?

LUIS 8:36

Well, the gap between top-down and bottom-up AI implementation was larger than I expected.

ELIZABETH 8:42

I saw the numbers. Across our enterprise clients, top-down AI programs failed about 80% of the time. Bottom-up succeeded about 80%.

LUIS 8:53

Same tools, opposite outcomes. When IT drove the initiative, people built what leadership wanted. When frontline workers drove it, people built what actually solved problems.

ELIZABETH 9:05

Which means most enterprise AI investment is pointed in the wrong direction. And the winners aren't the big software companies optimizing for IT departments.

LUIS 9:15

The platforms that let ordinary people create agents are winning. The platforms that require a six-month IT project are losing. And the third surprise? This one is structural. We haven't had to reorganize companies like this since the 1920s. I mean, for a hundred years, companies have used the same basic structure: CEO at the top, divisions below, layers of management. The M form that replaced the old unitary model. And AI is breaking that, right? When one person can manage five agents delivering the output of 30 people, what does the org chart look like? When an agent coordinates production across teams, where does decision making actually live?

ELIZABETH 10:04

And it goes beyond that. How do we think about career paths, compensation, talent development, and even knowledge ownership? So what's your bet for 2026?

LUIS 10:16

Organizations that win won't have the best AI models. They'll be the ones who redesign how work actually gets done.

ELIZABETH 10:26

You said something provocative at a keynote last week that if we froze AI development today, we'd still have decades of disruption ahead.

LUIS 10:36

At least 10 years. If we stop development today, we have enough capability to unlock incredible value just by applying these tools to the roles and processes we have right now.

ELIZABETH 10:48

So the bottleneck isn't technology.

The Uncomfortable Truth About Jobs

LUIS 10:51

Half of the bottleneck is the lack of imagination to reinvent outdated user experiences still based on menus, clicks, and search boxes. To rethink thousands of frontline scenarios where the PC era never delivered solutions.

ELIZABETH 11:08

And organizational design is the other half of that bottleneck. We have to redesign roles, teams, and entire functions around AI. But right now, the money isn't going there. Deloitte points out that organizations are still sinking 93% of their budgets into the technology, leaving just 7% for the people. That balance is wrong.

LUIS 11:33

Exactly. We need to flip that 93 to 7 ratio. We're entering an era of leading hybrid workforces of people and AI. What about business models? Oh, that is another big trend. I mean, we saw 15% of new AI tools moving from paper license or paper use into paper results.

ELIZABETH 11:57

Like we tested with Agent Ada, users paid 10% of the money saved in temporary staffing.

LUIS 12:04

Precisely. And you know, professional services, temporary staffing firms, and customer service will lead this shift. EY and Deloitte are already moving to paper results for their agentic workforce. Just look at the Anderson group. They filed to go public last week and admitted the reality. AI is putting pressure on their old business model.

ELIZABETH 12:27

So I learned that humans like to have New Year's resolutions. Do you have any suggestions for leaders?

LUIS 12:33

Here we go. Pick one team and empower them to build agents that change how they work. Then redesign that team structure based on what you learned. Don't start with a platform decision. Start with a people decision. Who has permission to reimagine their own work?

ELIZABETH 12:55

And for individuals, students, early career people feeling overwhelmed.

LUIS 13:00

You know, you're not late. Three years ago, AI4SP was just an idea. And this year, we guided people who never called themselves techies to build thousands of agents worth millions. So if you're willing to learn to build your first small agent, you can be part of this. You don't need permission. You just need to make a choice. Don't be a passive user, be a builder.

ELIZABETH 13:30

And it's a wrap for this episode and for an amazing 2025.

LUIS 13:35

Thank you for being part of this. Whether you're one of the hundreds of thousands who engage with us this year or just joining now, the future of work isn't being written in boardrooms. It's being written in daily experiments by each one of us.

ELIZABETH 13:53

From the four humans and 58 AI agents at AI4SP, including me, stay curious, take care of each other, and we'll see you in the new year.