Scaling AI from Experimentation to Impact with IBM Consulting

I had a great conversation at IBM Think 2026 with Javier Cassini, Global Managing Partner for Hybrid Cloud and Data at IBM Consulting. We'd just come off Constellation's Futures Forum conference, where I'd spent two days with 120 CEOs and board members. The theme was unmistakable: the metrics are real now, and the pressure to scale AI is no longer theoretical.

Javier confirmed what I was hearing on the ground. The question has fundamentally changed. Clients aren't coming to IBM asking for use cases anymore. They want to capture the dividend, and they want to do so at an industrial scale.

Three Patterns Driving Real Results

Javier outlined three profiles he's seeing consistently across IBM's customer base.

The first is productivity at scale. Not 10-20% efficiency gains... we're talking 40-60% reductions in operating costs, depending on the domain. The key insight here is that you can't get there by layering AI on top of existing steps. You have to re-engineer the entire workflow from first principles. Finance transformation, order-to-cash, supply chain, and software development lifecycles. All of it gets rethought around AI, not retrofitted with it.

The second is decision velocity. I love this framing because velocity isn't just speed; it has direction. It means collapsing decision trees, automating steps that previously required a committee meeting every Friday, and doing it with better context and data than a human could have managed alone. So many processes aren't expensive in terms of headcount, but they're slow due to friction at specific points. That friction is where the opportunity is.

The third, and arguably the most strategic, is net-new revenue. Pearson was one of the examples on stage at Think: how do you build entirely new AI-native products and business models? Javier is clear that you typically need the second pattern to fund the third. But boards increasingly see this as existential. If you're not learning and building the flywheel now, you may be behind your competitors before you realize it.

Why Sophisticated Teams Still Struggle

This is the part of the conversation that I think gets glossed over too often. Even teams with strong data foundations and clear executive sponsorship hit walls. Javier's take: it starts with security, trust, and governance. If you don't invest in those early on, you'll hit problems at scale that you didn't anticipate in a clean MVP environment.

He also walked through a data readiness framework that I thought was genuinely useful. Most teams think about data in terms of cleanliness — is it accurate, is it structured? That's layer one. But you also need a semantic layer (what does the data actually mean?), a context layer (what does it mean in this specific situation, in this market, with these rules?), and a decision layer where feedback loops capture what actually happened and improve the model's autonomy over time. That last piece, the learning loop, is what separates organizations that get compounding returns from those that plateau.

The Operating Model Is the Real Bottleneck

Javier said something near the end of our conversation that I've been thinking about since. The bottleneck has moved. It's no longer a technology problem. It's an operating model problem. The question isn't whether the models are good enough, they are. The question is how quickly you can transform your organization, how quickly you can bring your team to new ways of working, and whether leadership understands the art of the possible well enough to provide real direction.

CEOs are feeling this acutely. At our FastForward conference, two things were top of mind beyond the standard profitability and productivity goals: do I have the right people, and can my organization actually evolve? Javier sees a lot of use-case proliferation that never scales because there's no single top-down commitment to a single vision. Organizations get attached to what they've already built, even when it hasn't reached scale, and it waters everything down.

What Success Actually Looks Like

One of the metrics Javier tracks is the level of agency: how much work can you delegate to agents versus keeping a human in the loop? The model he described is humans at the edge: setting direction, defining what they want, delegating execution to agents, and expanding that surface over time as trust and observability increase. Autonomy isn't a switch. It's a dial. And you build toward it through evals, feedback loops, and deliberate governance — not by hoping the models figure it out.

Interestingly, the organizations getting there fastest tend to be in regulated industries. They already know exactly what humans can and can't do. That clarity turns out to be a competitive advantage when you're designing AI systems with the right guardrails.

The conversation could have gone another two hours. We barely scratched the surface of governance trade-offs and the shift from automated decisions to trusted execution. More on that soon.