AI Made You Faster. Did It Make Your Company Better?

A conversation with Marko Horvat, SVP of Business Transformation at ELB Learning

Marko Horvat has spent his career at the intersection of finance, strategy, and organizational transformation: as a CFO and controller across healthcare and consulting, as a VP at Gartner where he advised hundreds of CFOs on transformation priorities, and now as SVP of Business Transformation at ELB Learning, a 25-year enterprise learning and development company that has increasingly focused its work on helping organizations navigate AI. He's also a Forbes Finance Council contributor and a board member.

On a recent episode of Building For Others, he put a name on something most enterprises are living through but struggling to articulate.

Individual productivity and firm productivity are not the same thing

Individual productivity and firm productivity are not the same thing. It's the frame Marko keeps coming back to, and it breaks most enterprise AI strategies once you test it against real work.

Here's the example he used to illustrate how this works: Let’s say he spent time setting up Copilot to write his emails, going from his natural three-sentence, occasionally misspelled style to polished, multi-paragraph responses. He is proud of the output. Then he sends one to Sachin Kamdar, elvex's CEO, who tells him he gets hundreds of emails a day and would very much prefer the shorter version. Sachin in fact builds his own bot to compress long emails back down to something readable, so Marko's productivity improvement required Sachin's productivity improvement to undo it: two people working harder, in opposite directions, on the same exchange.

"Did I get more productive? I guess," Marko said on the episode. "Did you get more productive? I guess. But we didn't really solve for a business problem that helps the business get better."

This is the failure mode of individualized AI. When there's no shared context, no common layer of knowledge about how work gets done, what outputs actually look like, what the next person downstream needs, the productivity improvements dissipate at the handoff instead of compounding. The handoff is where value actually transfers between people, and that's exactly what most AI platforms leave untouched.

Marko pushes this further: AI has made it very easy to start things and much less clear that it helps people finish them. A four-hour task now takes 45 minutes, which means people have three and a half extra hours to pick up the next thing, creating more work in progress that stalls whenever it reaches a collaboration point. "You can be incredibly individually productive," he said. "But unless you're the only stakeholder and sole decision maker in that loop, you can only get so far." At some point something requires other input, and that's where the bottleneck has always been.

So what do you do about it?

Better prompting and more training don't address the structural problem. Individual workflows need to become team workflows.

When someone figures out a better way to do something, that pattern needs to surface to the people around them rather than staying in their personal chat history. The context that makes AI useful (how your company talks about its work, what good output looks like, what the next person in the chain actually needs) has to be shared and maintained at the organizational level, not rebuilt from scratch by each person every time.

The interaction points themselves need to be multiplayer: the places where one person's output becomes another's input are exactly where AI should be present, not just the solo stretches in between.

Both of these require a platform that applies across the whole company.

Enterprise AI has delivered real productivity gains. They just aren't converting into organizational results, because the tools were built for individuals and deployed at companies.

The Model Isn't the Moat

Marko also brings a structural observation from his finance background: the underlying model is rapidly becoming a commodity.

He cites MIT research from last summer that tracked frontier versus open-weight model performance on a common benchmark. When the study started, open models were six months behind. By the end of the summer, they were three months behind, and the economics of model pricing suggest the convergence will continue. "The LLM behind it is rapidly becoming like a commodity," he said. "It's all about how you use the tool."

For organizations, this means the competitive advantage lives in the harness around the model: the institutional knowledge, proprietary workflows, and company-specific context that make AI outputs reflect how the organization actually operates, rather than how the average user prompts. If everyone sends the same prompt to Claude, they get roughly the same output. The companies that pull ahead will be the ones feeding their own way of doing things into the AI layer.

This is why Marko is skeptical of the "prompt engineer" framing that dominated 2023 and early 2024. AI gets easier to use with every iteration, and the usability challenge is largely being solved by the technology itself. The harder work, the part most organizations are still deferring, is the governance and context layer: deciding what AI should be used for, in what structure, according to what institutional rules.

The Implied Bet Nobody Is Naming

When companies fund AI initiatives through headcount reductions or hiring freezes, they're making an implicit wager that AI-enabled productivity will arrive fast enough, at scale, to cover the reduction. Marko calls that an implied bet, and the early returns aren't encouraging.

McKinsey research shows that high performers outperform average performers by an exponential margin, not a fixed one. So when you project revenue or output based on an AI-enabled capability level that never materializes across your workforce, the organization finds itself without the assets it thought it had. "That's a balance sheet liability," Marko said, drawing on his Forbes Finance Council work. "It means you do not have the assets you thought you were going to have to achieve the results you thought you were going to achieve." The $500M AI budget that burns in a month is a spending story. The organization not being able to produce at the level it planned for is a strategic one.

You Can't Hire Your Way Out

The solution most companies reach for is consultants and specialist hires. Marko thinks this is the wrong move, and he's direct about why.

AI is moving faster than the labor market can respond. Someone who has been building agentic enterprise finance workflows with real production deployments and audit history is largely a theoretical hire at this point. The handful of people who do have that experience are either too valuable to their current employer to leave or already building their own thing. "You have to train your own people," he said. "This is not a problem you can hire out of, because the pace of technology is increasing faster than the workforce can keep up."

ELB Learning's work is built on exactly this premise. The company has spent 25 years building organizational capability to help companies learn new ways of working, and Marko's playbook centers on building ownership of AI solutions inside the organization itself: in the workflows, onboarding processes, and career development paths that make capability self-sustaining rather than dependent on a champion who might leave or a consultant who will.

He's also seen what happens when companies skip this step. Clients come in asking for use cases and agent builders, and Marko redirects them. "What is it exactly that you want them to do?" he asks. The answer is usually something they heard at a conference. His response is consistent: identify the workflow problems first, then figure out where AI helps solve them.

Stop Obsessing Over Data

Marko closed with a take that lands differently coming from a CFO: companies are too focused on data and not focused enough on decisions.

Moneyball came out in 2003 and The Signal and the Noise in 2012; we've been having essentially the same conversation about data-driven culture for two decades. Meanwhile, data has become its own end, with organizations building lakes and warehouses full of information and no clear picture of what decisions they're supposed to enable.

AI has compounded this. Because AI is good at querying data and surfacing patterns, it's now easy to find data that supports whatever conclusion you already wanted to reach. "You can get the data to fit whatever decision you want," Sachin noted during the conversation. The real discipline is knowing what decisions you need to make and working backwards to the data those decisions require.

Marko's prescription: instead of asking what data you have, ask what you're trying to decide and what you need to know to decide it well. In a business environment changing this fast, historical data from five years ago has limited bearing on what you're doing today. "You're generating data every day," he said. "By the time your initiative goes live, you'll have six months of data designed for exactly what you're building."

Marko Horvat is SVP of Business Transformation at ELB Learning and a Forbes Finance Council contributor. You can find him on LinkedIn and follow his ongoing writing on AI, finance transformation, and workforce strategy. Learn more about ELB Learning at elblearning.com.

Building For Others is elvex's podcast about the people actually building AI-powered things that other people use. No hype, no vaporware: just practitioners sharing what they've made and what they've learned.