Enterprise AI Is Headed Toward Autonomy, Says NTT Data’s AI Chief

Wendy Collins- Chief AI Officer at NTT Data NTT Data Most enterprise AI is still stuck in assistant mode, with tools that help write emails, summarize documents, or suggest next steps, but only when asked. What it’s not doing, at least not at scale, is taking meaningful action on its own. But that reality is beginning to evolve in subtle, yet important ways.

In a recent interview with Wendy Collins, chief AI officer at NTT DATA — a part of the global conglomerate NTT Group and innovator of IT and business services — she noted that this is a trend that we will begin to see more and more across the enterprise, adding that the future of AI in the enterprise isn’t just about intelligence, but about autonomy. Collins believes we’re entering a new phase where AI is starting to operate more like an agent than an assistant, moving beyond informing decisions to initiating them. And while the hype is still loudest around GenAI and copilots, these autonomous AI systems that take action are now seeping their way into the heart of the AI conversation.



Agentic AI — a term still finding its footing across the industry — describes systems that don’t simply return answers but complete tasks. In Collins’ words, it’s the difference between an AI telling you a return policy and one that can issue a return authorization, log it in the system of record and notify the customer. Rather than involve a single AI model that just generates results by pulling answers from a vast pool of pretrained datasets, AI agents depend on a coordinated stack of technologies — including language models, decision engines, integrated tools and real-time data access — to execute tasks.

“Agentic AI,” Collins told me, “is bigger than generative AI.” It sits at the intersection of multiple capabilities, and its strength lies in doing, not just knowing. Collins was, however, candid about the current limits.

In industries like insurance and financial services, agentic AI is already reducing cost and latency in call centers and procurement. These domains work because the processes are predictable and well-documented. If a task can be reduced to rules and data flows, it can be delegated to an agent.

Where it falters — for now — is in high-context workflows like underwriting or complex manufacturing, where much of the decision-making still lives in the heads of experienced employees. "We’re not seeing agentic AI take hold in those environments yet," she said. "Because the knowledge hasn’t been captured.

" But that doesn’t mean these industries are excluded from progress. It just means that they're currently focused on using GenAI to collect and structure the very knowledge needed to support agentic systems down the line. What Collins is most bullish about isn’t GenAI alone.

It’s what happens when GenAI is fused with traditional AI techniques like optimization, forecasting and rule-based systems. She called it “ hybrid AI ,” noting that it’s the most overlooked and under-discussed area of enterprise AI transformation right now. “GenAI is a hammer,” she said.

“But some problems need a wrench.” For many business challenges, multiple tools working together are what unlock the most value. And while GenAI can generate, recommend and personalize, it still relies on classical AI to drive precision, consistency and integration.

Many enterprises are stuck in proof-of-concept purgatory — touting dozens of pilots but deploying none. According to Collins, the gap between POC and production is much larger than most leaders anticipate. “It’s not linear,” she explained.

“It’s exponential.” Her advice is that business leaders must “stop trying to boil the ocean.” Instead, they need to start with one or two high-value, internal use cases.

Focus on workflows where the AI can succeed quietly and quickly, not in customer-facing experiments that risk brand equity before the technology is ready. And, maybe most importantly, was what Collins said about building with measurement in mind. “ROI needs to be planned for from the beginning, not retrofitted at the end,” she noted.

Perhaps the most underrated variable in any AI deployment is people. Collins emphasized the importance of enterprise-wide AI literacy, especially among enterprise executive teams. “Companies that invested in executive AI literacy outperformed their peers financially by 40%,” she said, citing recent research.

While Collins didn’t mention the exact research she quoted, it’s an assertion that some other studies support. For example, one MIT CISR research found last year that companies with advanced enterprise AI — which often prioritizes AI literacy as a key component — outpace industry peers in financial performance For Collins, AI adoption is more about comfort, confidence and context, rather than being about a lofty desire to build out infrastructure. While there’s growing talk of AI fatigue, Collins believes much of that fatigue stems from underwhelming outcomes from how AI is used.

When AI doesn’t deliver transformation — when it’s used only to shave seconds off a task — teams lose interest. What that means is that AI’s promises must be felt, not just marketed. As AI becomes more capable of making decisions, enterprises will need stronger governance models to match.

Already, the concerns about AI safety are at an all-time high, with the World Economic Forum noting that “the increasing autonomy of AI agents introduces both immense opportunity and considerable risk . Without proper oversight, such systems may behave in unexpected ways or even undermine intended goals.” The OECD AI Observatory’s 2024 policy report also details how autonomous AI systems challenge existing governance frameworks and increase the urgency for risk mitigation strategies, particularly as these systems begin to operate independently.

While some often think of AI governance as an innovation blocker, Collins described it as “a strategic enabler”. Her team at NTT Data also developed a “payoff matrix" to help clients identify where to begin, how to align value with feasibility and where the biggest traps lie. “It’s not about waiting until all your data is perfect,” said Collins.

“It’s about knowing which parts of your data are good enough to begin capturing value now — and building toward the rest.” From all indications, the future of enterprise AI won’t be decided by the next viral chatbot demo or shiny new app. It’ll unfold quietly — inside workflows, behind dashboards, — where AI stops waiting for instructions and truly starts working on its own.

Collins also urges caution, noting that if someone tells you they know what AI will look like in five years, they’re either lying to you or trying to sell you something. Still, she remains clear-eyed about where we’re headed: “Every new incremental development is going to unlock new problems in the same way that it unlocks new opportunities.” That's the future enterprise AI seems to be inching toward.

.