3 AI lessons from 27 data leaders in 2025

Welcome to the deep dive today we’ve taken on a pretty monumental task. It’s a big one. Our source material isn’t just an article or two, it’s the entire 2025 library of insights from David Sweenor Show databases,

28 separate interviews with some of the biggest minds in AI and

analytics, exactly. So our mission here is to filter all of that expertise and give you the one cohesive story that really defined the year.

And when you look back 2025 really was a tipping point. I mean, we finally moved past the, you know, the speculative hype, and into the messy reality of applying the stuff at enterprise scale, for sure. And what we found looking across all those conversations is that success really boiled down to three core themes, first, strategy over technology, always. Second, navigating the explosive rise of AI agents. And third, a really intense focus on the human element in all of this. So if you’re

looking for the quickest way to understand where top companies actually put their money and their energy this year, this is your shortcut. You’re gonna walk away with the distilled wisdom of dozens of leaders. All right, let’s do it. Okay, let’s unpack this. Let’s start with that first and often most painful lesson,

the first big theme, strategy Trumps technology. It just came up again and again, mostly because of how many projects were failing, right? And the consensus was so clear, the failure point was never the algorithm. It was the organizational planning, the

planning or lack thereof. Eric Kavanaugh had that perfect phrase for it, the ready fire aim approach, exactly.

Companies just get excited, they launch and then they realize they have no idea where they’re even shooting.

And that’s the huge, expensive blind spot, isn’t it? The organizational maturity DAP, we have this amazing technology, but our internal processes just aren’t ready for it. They can’t absorb it. I remember Gail Carlson was incredibly blunt about this. He argued that what something like 90% of Gen AI projects will fail to deliver transformative value.

Transformative. Yeah, that’s the key word.

And not because the math is bad. It’s because the organizations themselves, you know, they lack the people, the processes, the governance, to actually scope and design these things properly.

It’s just a massive resource drain if that scaffolding isn’t there, and it creates a huge accountability problem. How so well, if you start with a cool tool and then try to find a business problem for it, you’re almost guaranteed to fail because you’re spending money without any clear metrics, right? Which is why Robert Lake’s advice was so critical. He said, leadership has to define the strategy first. They have to set the vision and say, Here is your specific problem to solve, instead of the other way around, exactly, instead of having the AI team deliver some amazing solution, and then trying to figure out how to, you know, bolt some business value onto it, after the fact that alignment up front, that’s what separates the winners from the resource burners.

And that actually leads perfectly into our second theme, because if strategy was the challenge, then the tools leaders were wrestling with that defined the solution. Which brings us to the rise of agents 2025.

Was the year we stopped thinking about agents as, you know, little bots that

schedule meetings, and started seeing them as foundational infrastructure.

Rich Mendes was really insightful on this. He compared agentic AI not to a product you plug in, but to electricity or

the web, a horizontal capability, exactly.

It’s something you have to embed into every layer of your workflows. And his key point was that this is not about replacement. It’s ai plus human. It augments, it accelerates, but it doesn’t usually run the whole show on its own.

But okay, once you accept that agents are like electricity. You immediately run into an operational headache, management, management, suddenly, building an agent is easy. It’s democratized, which means companies are facing this problem of overseeing 1000s of them. It’s a compliance nightmare, and

that’s where John Thompson gave some really critical, practical advice. He said, You have to wrap an agent with everything you would wrap around an employee,

all your internal policies, procedures, everything.

But here’s the profound difference. He pointed out, agents are inherently more compliant because, as he put it, they don’t willfully not follow the instructions.

Ah, right. They don’t have bad days. They don’t get distracted or decide to cut corners.

They lack that human capacity for willful negligence. So if compliance

is actually easier, where did these agents deliver the real business value? Because we always see the flashiest examples, but the sources showed the value was often in the the unsexy stuff

totally in the narrow defined tasks. Chelsea wise shared a perfect example. She used an LLM workflow to solve what she called a multi dimensional problem deep inside an HR process, exactly not some big customer facing thing. She had a spreadsheet of boot camp applicants and needed to flag the

high fit candidates. And the agent helped her look past the obvious signals.

It looked for non conventional signals, things like the motivations they. Expressed in their essays, not just the company name on their resume or if they went to an Ivy League school. That’s real value. It’s automating the discovery of nuance that a human reviewer would just miss,

and that focus on nuance leads us right in to our third major theme, the absolute criticality of the human element.

Yes, because you can have the best strategy, the most compliant agents, but if your internal culture is a mess, the whole thing just falls apart.

Danny stout really drove this point home. He said alignment is absolutely vital.

He was so clear on it. He stressed that if the team, data science, business engineering, if they aren’t aligned before you even start. If they don’t agree

on the goal, there’s no way that whatever model you choose is going to be successful.

Exactly. The tech is just an accelerant. If you’re accelerating in the wrong direction, you just fail faster and more expensively.

The human element also means facing up to the complexity of real world ethics. This isn’t just a simple fairness switch you can flip.

No, it’s all shades of gray. Monica Cisneros talked about this, reflecting on her ethics training at Stanford, the world is complex, so the AI you apply to it will be just as complex. You have to navigate those trade offs,

and that tension became painfully real with things like the compass recidivism algorithm.

That case study came up a lot. It showed that when you try to achieve one definition of fairness, you almost always violate another what’s an example of that? Well, if you try to increase what’s called classification parity, so equalizing the error rates between different groups, you actually reduced calibration, which is the consistency of the risk score itself. And the outcome that was, the painful outcome, was that black defendants were twice as likely to get a false positive to be incorrectly labeled as high risk. It shows you have to make these really difficult, real world choices about which definition of fairness you’re even optimizing for.

It’s an incredibly difficult tightrope to walk, and that walk extends to things like creativity too. When you rely on AI to just churn out content, you run into what Melissa burrows called regression to the

mean, where everything just starts sounding the same, bland, safe, corporate,

and that mediocrity comes with its own risks. BURROWS warned about AI laying traps in the language. What does she mean by that tiny, subtle errors that are easy for an AI to make, but that erode customer trust, like using the word complimentary meaning free when you actually meant complimentary meaning, it adds value.

Yeah, the AI doesn’t get the semantic context of your brand, and those little slip ups, they really add up and cost you.

So that gives us our three pillars for 2025 strategy, agents and people. But David’s interviews covered so much more ground, we need to shift gears now for a high speed synthesis of all the other vital insights,

let’s do it time for the lightning round. We can group these thematically. Maybe cover adoption, then governance, then

operations, perfect. Let’s start with the adoption and market shift we saw this year.

Okay, and we can’t forget the insights from Sean Rogers, who told David that Gen AI hopped the chasm into the mainstream,

right because it was consumer LED. Unlike things like Hadoop

exactly, it was understandable. And Judith snowbo added that this levels the playing field, but only if you treat it like a teammate and start experimenting early.

You can’t just sit back. Okay, so that experimentation has to be backed by serious governance. What did the experts say about controlling the chaos?

This was a big shift. Tom has been observed that governance is becoming a value driver, not just a compliance headache, it’s actually shortening model life cycles by up to 70%

so you either fail faster or succeed quicker, precisely.

And Shane Murray added that the best data teams are focusing on platform standards, privacy, security, observability, so they can govern without being seen as pumping the brakes on speed

and for open source. BRUNO trun his point was crucial, absolutely.

He stressed the need for Responsible AI where everything generated, whether it’s code or a SQL query, has to be inspectable and reproducible

for trust. Okay? Moving on to agent and operational management.

Creon Park gave us a pretty sobering look at the future here. He warned the next big challenge is managing the sheer number of agents. We’re talking 10, maybe 20,000 agents that all need constant updates.

Wow, that’s a huge operational burden. And Jawad Rashid pointed out how agents are changing internal roles.

Yeah, he said the finance function is moving from just cost control to being a cognitive agent that actively drives revenue and critically, he said that having imperfect data can’t be an excuse to block that innovation.

I liked rajev kazika todis point here about the irony of it all

the return to high touch human experiences. He said, agentic AI’s real untapped potential isn’t driving that automating all the pre event research and meeting logistics

so the actual in person meeting can be about real, high value connection, exactly. Finally, let’s hit data, practice and team enablement. Automation isn’t just about replacement.

No, it’s about making people better at their jobs. Kamal Maheshwari said AI should automate the tedious tasks people hate, like defining data glossaries free them up for real thinking.

Matt Magney had that great insight on sales training, the AI

driven roleplay reps, especially new hires, prefer it because it’s less stress than practicing a pitch in front of their manager, a low stakes way to build high

stakes skills. But none of that works if you don’t actually understand your data landscape, which

is Tina Chase’s huge blind spot, the misconception that you can’t document end to end data lineage, especially stitching inter system lineage across 20 plus different apps,

believing it’s impossible means you’re always underestimating the complexity,

always and tying this all back to the human element. Gabriela Contreras stressed that you have to adjust your messaging altitude,

right? Developers want the deep technical dive the

lower altitude, but a CFO just wants the high altitude view. What’s the value and Will it break anything you can’t speak the same language to everyone.

That is an incredible amount of wisdom from an amazing year of conversations. So let’s bring it all home. What does this mean for you as we look ahead?

The overarching theme, really is that AI success in 2025 was all about operational maturity and clarity of intent. It wasn’t about raw model power.

You learned that you can’t just outsource a broken strategy or a messy internal process to an algorithm and hope for the best, right?

And that realization is forcing everyone to rethink what is uniquely human about their role. Thomas Ben had a brilliant take on this saying AI is making marketers more human. How? So? By automating all the mechanical stuff, the data entry, the scheduling, we are now forced to focus on the things tech still struggles with, creativity, empathy, relationship building. That’s where the real human value is.

Now that shift from imitation to augmentation is powerful, but maybe the thing to leave everyone thinking about connects back to Monica Cisneros idea of the Turing trap track developers instinctively want to make AI as human as possible, right to perfectly imitate us. Yeah, but we have to decide if that’s really the right goal,

especially if it just serves to displace people from uniquely human, creative or compassionate tasks.

So the question for you to ponder is, is perfect imitation the ambition we should aim for? Or should we define AI by what it enables us to achieve a really important question. If you enjoyed synthesizing all these insights, remember you can go deeper into David sweenor’s work, his books, the PMMs prompt playbook and generative AI business applications are available now through

TinyTechGuides, and thanks for joining us for this year end Deep Dive.

You can catch the full interviews with all of these amazing thought leaders by subscribing to databases on Spotify, Apple podcasts and YouTube. We’ll catch you next time.