One year ago, enterprise AI was defined by experimentation. Companies were building pilots, testing large language models, and trying to determine where generative AI fit inside their organizations. Success was measured by whether the technology worked.
Today, the conversation has fundamentally changed.
At Worth’s second annual AI reception with IBM during the Ai4 conference in Las Vegas, I sat down with Sunil Murthy, IBM’s AI Field CTO, to discuss what has changed over the past twelve months. His answer was immediate.
“The rate and pace of innovation is pretty rapid,” Murthy said. Organizations have moved from pilots into production much faster than many expected. The progress, he said, has been “exhilarating.”
Yet beneath the enthusiasm lies a more complicated reality.
IBM recently surveyed approximately 1,000 C-suite executives about their AI initiatives. Average AI return on investment came in at roughly 55 percent, well below the triple-digit returns many companies expected only a year ago. The shortfall is not a reflection of disappointing technology. It is evidence that enterprise transformation remains far more difficult than deploying a new model.
“The challenge with data,” Murthy told me, “has stayed the same.”
That deceptively simple observation may be the defining lesson of enterprise AI’s second year.
The technology is improving at extraordinary speed. Organizations are not.
One of the most interesting themes throughout our discussion was how dramatically executive priorities have evolved. Twelve months ago, nearly every AI conversation centered on productivity and cost reduction. Companies wanted to know whether AI could automate repetitive work, reduce headcount, or improve efficiency enough to justify the investment.
Those questions have not disappeared, but they are no longer the primary focus.
Increasingly, CEOs are asking how AI can create new revenue.
“The conversation has shifted,” Murthy explained. Organizations are using AI to reach customer segments they previously could not serve, accelerate product development, optimize supply chains, and shorten the distance between an idea and the marketplace.
That distinction may ultimately define the next phase of enterprise AI.
Throughout our conversation, Murthy returned repeatedly to one idea. The biggest obstacle to scaling AI is rarely the AI itself.
IBM’s research found that approximately 21 percent of potential productivity is lost to friction between technology teams and business leaders. Additional friction exists inside technology organizations themselves, between architects, developers, and operations teams. The technology often works. The organization surrounding it does not.
Murthy believes a consistent set of issues prevent companies from moving successfully from pilot to production: architecture, data, decision rights, and operating models.
IBM’s Institute for Business Value has been studying those same fault lines in its report Redesign for enterprise AI, and its research adds detail to what Murthy described in our conversation.
On architecture, the report found that most enterprises run hybrid environments by default rather than by design. Without a consistent execution layer tying multiple clouds and on-premise systems together — what Murthy calls an agent control plane — every environment boundary becomes a friction point.
That friction carries a price tag. Architecture economics, as the report frames it, increasingly determine whether an AI initiative pays for itself, and the same multi-vendor, multi-model sprawl that creates the friction also creates an opening: model arbitrage, routing each task to the most cost-effective model rather than defaulting to the priciest one, keeps costs in check as agent usage scales.
On data, the report found that the majority of high-value process data sits outside the public cloud, stored in formats built for retention rather than action. Agents end up reconstructing context at every step instead of pulling real-time context directly from where the work happens.
On decision rights, agentic AI forces organizations to answer questions they have mostly avoided: who owns an agent’s decision outcomes, what level of autonomy is acceptable, and how performance gets measured and corrected.
On business and IT operations, the report found that renovating an IT estate for AI carries real setup costs. The payoff depends on continuity, teams that build on prior work and deepen context over time, rather than the ramp-up, deliver, disband pattern most transformation projects follow.
As Murthy put it during our discussion, organizations need to stop simply recalibrating business processes and begin redesigning them from end to end.
That distinction is critical.
Many companies are automating yesterday’s processes rather than inventing tomorrow’s.
One of Murthy’s more provocative observations was his belief that “AI becomes the new UI.”
Instead of employees navigating dozens of enterprise applications, future workers may simply interact with intelligent agents capable of retrieving information, executing workflows, and generating answers regardless of where the underlying data resides.
Imagine asking a single AI assistant to generate a global financial report instead of logging into ten separate reporting systems.
Murthy believes that future is arriving sooner than many expect.
“The same architecture that lets AI act as the new UI also lets you retire redundant backend systems as agents mature,” he writes. Rather than maintaining multiple procurement systems, reporting tools, or workflow platforms, organizations will increasingly consolidate complexity behind intelligent interfaces.
That represents far more than adding AI features to existing software. It represents a new way of thinking about enterprise software itself.
Governance has also evolved considerably over the past year.
Last year, organizations were still trying to understand AI risk. Today, they are beginning to operationalize AI governance across increasingly complex environments populated by hundreds, and eventually thousands, of autonomous agents.
Murthy pointed to one practice that separates organizations that scale agents quickly from those that stall: evaluations first. Rigorously testing how an agent behaves before it goes into production builds the track record needed to expand its responsibility later, with far fewer surprises.
Even so, Murthy argues that traditional model evaluations are no longer enough.
His framework is remarkably straightforward: visibility, control, accountability.
“If something is not visible,” he told me, “you cannot control it. If you cannot control it, you cannot drive accountability.”
As enterprises deploy agents across multiple cloud platforms and software ecosystems, governance becomes less about evaluating individual models and more about observing, securing, and auditing an entire network of autonomous systems. His Ai4 paper makes the same point succinctly: “If you can’t see an agent’s behavior, you can’t control it. If you can’t control it, nobody can be accountable for it.”
For highly regulated industries, that may prove to be one of AI’s most important competitive differentiators — it’s part of why IBM was recently named a Leader in the inaugural Gartner® Magic Quadrant™ for AI Governance Platforms, recognition tied to its approach for helping organizations manage AI risk, oversight, and compliance at scale.
Perhaps the most memorable example Murthy shared came not from a customer deployment but from a simple thought experiment.
“If something’s 100 meters away, should I walk or drive?”
The obvious answer is to walk.
Unless, he points out, you’re headed to the car wash.
“The model isn’t missing intelligence,” he writes. “It’s missing context.”
That simple analogy perfectly captures the challenge facing enterprise AI today.
Foundation models continue improving. Costs continue falling. Standards continue maturing. Agentic systems are becoming increasingly capable.
None of those advances eliminate the need for organizations to provide the context, governance, architecture, and operating models that allow AI to make consistently intelligent decisions.
In many ways, the constraint has shifted.
For decades, technology limited business ambition. Companies wanted to move faster than their systems allowed.
Murthy believes that equation has flipped.
“The model was never really the constraint,” he concludes. “The constraint is whether the organization around it, its architecture, its data, its decision rights, and its operating model, gets redesigned to match.”
The race is no longer about building better AI.
It is about preparing organizations to maximize the benefit of AI.
