Most organizations today use AI to make individual human tasks faster, or at most, require humans to monitor AIs performing tasks. A smaller number of organizations are pursuing something fundamentally different– enterprise autonomy. Enterprise autonomy is the ability of an organization’s systems to make consequential business decisions with minimal human intervention.
The distinction matters. Autonomous systems compress decision cycles, eliminate bottlenecks that depend on human availability, improve consistency across high-volume operations, and free skilled professionals to focus on work that requires genuine judgment. In industries where speed, accuracy, and responsiveness define competitive advantage the ability to operate autonomously at scale is becoming a differentiator, not an aspiration.
Yet, most organizations remain far from this reality. Based on Avasant’s research, AI adoption grew by roughly 45% from 2024 to 2025 [1]. But only a few of those funded projects have progressed beyond isolated pilots into production systems capable of autonomous decision-making.
That gap is not a technology problem. The models are capable. The compute is available. The gap is a leadership problem, a measurement problem, a data problem, and, perhaps most underappreciated, a contract problem.
At Avasant’s Empowering Beyond Summit 2026, a panel of senior enterprise leaders from Hertz, IBM, IFF and Prudential gathered to examine precisely this tension: why are so many well-resourced organizations still stuck in a cycle of disconnected pilots, and what does it practically take to leapfrog into genuine enterprise-wide autonomy?
Their answers were direct and consistently pointed toward the same conclusion: the barrier is not what most boardrooms think it is.
This article examines the six barriers that keep enterprises stuck between AI experimentation and scaled autonomy.
Together, these barriers explain why AI budgets continue to rise while enterprise autonomy remains limited in practice.
The language an organization uses to frame its AI initiatives shapes the outcomes those initiatives can realistically achieve.
Rahul Lele, who leads strategy and the technology portfolio at IFF, a $10–12 billion global flavors and fragrances enterprise, made an observation that should prompt reflection across every CIO office: IFF has deliberately stopped calling early-stage AI initiatives “pilots” or “proofs of concept.” They call them “Proofs of Value” or POVs.
“Conceptually, when you say proof of concept, it sounds very experimental,” he noted. “But when you attach a certain value to it, there is automatically a mindset of commitment. You are all in and you are trying to see whether this will actually deliver value or not.”
The semantic shift is not cosmetic. A Proof of Concept invites your organization to test whether something is possible. A Proof of Value demands that your organization define, upfront, what business outcome it is chasing and then assess honestly whether the initiative is delivering it. This reframes failure from “the experiment didn’t work” to “we learned this path does not create value, so let us find one that does.” The fail-fast culture borrowed from Silicon Valley becomes manageable when each iteration is anchored to a defined value, not just a technical demonstration.
“Micro efficiencies, trying to automate or eliminate redundant processes in silos, focusing on tasks rather than purpose, hold us back.”
– Shridar Jayakumar
The stakes of getting this right are significant. S&P Global Market Intelligence’s 2025 survey of over 1,000 enterprises found that the average organization scrapped 46% of its AI proofs-of-concept before they reached production; and 42% of companies abandoned most of their AI initiatives altogether that year, up sharply from 17% in 2024 [2] [3]. MIT’s research found that 95% of generative AI pilots produced no measurable profit-and-loss impact within six months [4] [5]. These are not numbers about bad AI. These are numbers about initiatives that were never grounded in value from the start.
Shridar Jayakumar, who leads AI-first business services at IBM, framed the systemic issue precisely: “Micro efficiencies, trying to automate or eliminate redundant processes in silos, focusing on tasks rather than purpose, hold us back. The AI-native startups entering your market are not shackled by old processes. Customers now have a choice.” IBM itself applied this logic internally, treating itself as “client zero” for an inside-out AI transformation and unlocking $4.5 billion in annualized savings across three years. These were savings that were reinvested into retooling people and refocusing on client value, not simply harvested as profit.
The takeaway for enterprise leaders: do not begin any AI initiative without a named, quantified business outcome that it is being measured against. Not a technology objective. A business outcome.
Data quality is not a prerequisite for an AI strategy; it is the AI strategy, or the absence of one.
Gaurav Rastogi, Senior Director of Data Analytics at Hertz, a company operating across 160 countries, described the foundational realization his team arrived at two and a half years into their journey: “Autonomy is not driven by AI. It is driven by data trust.”
Gaurav described a foundational architectural decision his team made early in their journey. Most large enterprises suffer from the same underlying problem: every department runs on its own copy of the data, with its own definitions of basic concepts like “customer,” “revenue,” or “order.” AI systems built on top of that fragmentation do not synthesize the truth; they just pick one version of it, often the wrong one.
“Autonomy is not driven by AI. It is driven by data trust.”
– Gaurav Rastogi
Hertz solved this by consolidating every fragmented data source across the organization into a single unified repository. Then they built a separate, rigorously cleaned and validated layer on top of it that became the one and only source of truth feeding all AI systems. No autonomous decision at Hertz draws from a local departmental copy. Everything runs from the same validated foundation.
But Gaurav was equally direct about what happens after you build it: “If there is no oversight, data quality will degrade 100%.” Data goes stale. Systems change. New sources get added. To prevent the trusted layer from quietly drifting back into the fragmented mess it replaced, Hertz assigned dedicated data quality champions whose explicit responsibility is to maintain the integrity of that foundation on an ongoing basis. The architecture, he emphasized, is not a project you complete. It is an operating model you commit to.
IBM’s Institute for Business Value found in 2025 that 43% of chief operations officers identify data quality as their most significant data priority, and over a quarter of organizations estimate losses exceeding $5 million annually from it [6].
The implication is uncomfortable for organizations that have invested heavily in AI tooling before addressing their data foundation: the models are not the bottleneck. The data is.
The board will ask about cost. That is expected and appropriate. But if cost reduction is the primary lens through which an enterprise evaluates its AI investments, it has already started on the wrong foot.
Rahul at IFF was clear on this point. For IFF’s make-to-order business units, where clients submit briefs for custom scent formulations and speed of response is a competitive differentiator, the metrics that matter are speed to deliver, first-time-right rate, and the accuracy of promised completion dates. “Cost is always typically a second afterthought. It is more about experience” he said.
This is not an argument against financial discipline. It is an argument about sequencing. AI’s highest-value applications in most enterprises are not in cost reduction but in acceleration, quality improvement, and decision speed. When organizations measure only cost saved, they systematically undervalue those contributions and underinvest in the initiatives that generate them.
Gaurav at Hertz introduced a complementary framework: unit economics at the product level. His team tied every AI data product to a cost function, measuring the unit cost of delivering a result against the strategic value it creates for the organization’s three core goals: revenue, operating expenses, and depreciation. This approach gave Hertz the discipline to abandon pilots where the numbers did not justify scale, and the confidence to double down where they achieved 10x ROI. “If you leave it open,” he noted, “it becomes a very expensive experiment that you cannot control.”
The companies that will scale are not the ones that start with “how much can we save?” They will start with “what business problem are we solving, and what does success look like in operational terms?”
The practical takeaway: define your AI success metrics before you begin and make sure they reflect the actual business goal.
One of the more nuanced tensions in enterprise AI adoption is the instinct to keep a human in every consequential decision loop, a reflex that, when applied indiscriminately, prevents organizations from realizing the value of autonomous systems at all.
Shridar at IBM named this directly: “The third thing to avoid is to be human-centric in the wrong way. Impulsively saying ‘we need a human in the loop’. Where do you draw the line?” His answer: human connection is irreplaceable and should be protected; human approval of every routine decision is a constraint that prevents scale.
Rahul at IFF provided a practical illustration from a highly regulated, sensory-driven industry. IFF uses AI to generate options for scent reformulations or raw material substitutions. But those suggestions are constrained to pre-approved “building blocks,” and the human expert retains the final call on sensory judgment. They cannot let a system decide ‘this smells nice’. The human is not in the loop to slow things down; the human is in the loop precisely where irreplaceable judgment is required. Everywhere else, the system proceeds.
Gagan Puranik at Prudential, operating in a risk-averse financial services environment, described a parallel approach: beginning with “high-frequency, low-regret” decisions as the first candidates for autonomy. This means beginning with work where the cost of an occasional error is recoverable and where volume makes human review impractical at scale. These initiatives build organizational confidence and demonstrate that autonomous systems can be trusted before the enterprise extends that trust to higher-stakes domains.
The design question, then, is not “should humans be involved?” It is “what is the specific contribution that a human makes at this point in this process that cannot be replicated by the system?” Where the answer is genuine judgment, expertise, or accountability, keep the human. Where the answer is habit, organizational inertia, or risk aversion dressed as governance, reconsider.
This point came up repeatedly across the panel.
Rahul made it plainly: “It is not only about how many people I can cut. If that is the question, you are asking the wrong question. At the end of the day, it is: how can I make my people more effective? How can I still have people think and apply AI so that the output is better?”
Gaurav reinforced this from the organizational adoption side: “Make AI a friend to your people that exist within the organization. Give them the path to success so they can participate together with you concurrently to bring your enterprise to the forefront.”
This is not a feel-good abstraction. It is a strategic argument. The organizations most at risk are not those that failed to adopt AI; they are those that adopted it to cut headcount, destroyed the institutional knowledge that made their models work, and then couldn’t recover.
The World Economic Forum’s Future of Jobs Report 2025 found that 85% of employers plan to prioritize upskilling their workforce in response to AI, and that by 2030, 77% are committed to reskilling employees specifically to work alongside AI. Crucially, the WEF projects that 59% of the global workforce will need reskilling or upskilling by 2030 [7]. This means that the question is not whether this investment is necessary, but whether your organization will make it proactively or be forced into it reactively.
Gaurav specifically called out reskilling as one of his three non-negotiables: “Encourage reskilling and upskilling within the group so people are encouraged to adopt new skills; because it is changing really fast. New models keep coming up. They should have the ability to try, learn, succeed, and fail together.”
Organizations that treat their people as a cost to be optimized will optimize themselves out of the human judgment and contextual knowledge that makes AI systems worth building in the first place.
Perhaps the most underappreciated operational blocker in the enterprise AI conversation is the structure of existing vendor and outsourcing relationships.
Gagan at Prudential, who oversees a vendor ecosystem covering approximately 60% of Prudential’s IT workforce, described the core problem: “The current way of engagement with suppliers is task-based. We really have to think about decision-making and shared accountability.”
Prudential is actively working with its strategic partners to reimagine contracts that shift from task execution to decision-driven incentives where the value a supplier brings is measured by the quality and outcomes of decisions supported, not the volume of work completed. “They are no longer just strategic partners,” Gagan said. “Now shared accountability is built into them.”
The vendor ecosystem, which executes a significant portion of the actual work, often gets left behind. Modernizing that ecosystem means engaging providers on the same AI-transformation journey, not just demanding outputs from a legacy task structure.
The market is moving in this direction broadly. Enterprises decision-makers are expanding their use of performance-based pricing, and traditional time-and-material contracts are increasingly being replaced by outcome-based agreements tied to measurable business results.
The action here is not complex in theory, though it requires organizational will: review every major technology and services contract through the lens of whether its incentive structures reward the outcomes you actually want. If your current contracts reward activity rather than outcomes, they are misaligned with an AI-first operating model, and they will quietly resist your transformation efforts from the inside.
Across the panel’s closing rapid-fire exchange, the four executives distilled their combined experience into the following framework. These are not aspirational principles; they are operational orientations drawn from organizations that have been doing this work.
Do these things:
Stop doing these things:
The enterprises that will move from isolated automation to true enterprise autonomy in the next three to five years are not necessarily the ones with the largest AI budgets. They are the ones that have been honest about the real blockers: fragmented data, misaligned metrics, poorly structured vendor relationships, and a workforce that has been told AI is happening to them rather than with them.
The Empowering Beyond Summit 2026 panel made clear that the path is known. American Airlines is already rebooking passengers autonomously at scale [8]. Lemonade processes about half of all claims through to payment without human intervention [9]. IBM has unlocked $4.5 billion in annualized savings by treating itself as its own first client for AI-first transformation.
The question for every enterprise in the room and every C-suite or leader reading this is not whether autonomous enterprise AI is possible. It demonstrably is. The question is whether your organization has the leadership mandate, the data discipline, the measurement honesty, and the supplier relationships to get there before your competitors do. Avasant can help you assess where you stand and guide you in the right direction.
[1] Avasant, “Worldwide Technology Trends 2025,” Avasant, 2025.
[2] S&P Global Market Intelligence, “The Big Picture 2025: Generative Artificial Intelligence,” S&P Global Market Intelligence, 2025.
[3] WorkOS, “Why Most Enterprise AI Projects Fail — and the Patterns That Actually Work,” 22 July 2025. [Online]. Available: https://workos.com/blog/why-most-enterprise-ai-projects-fail-patterns-that-work.
[4] MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025,” Massachusetts Institute of Technology, 2025.
[5] SR Analytics, “Why 95% of AI Projects Fail and How Data Fixes It,” 26 February 2026. [Online]. Available: https://sranalytics.io/blog/why-95-of-ai-projects-fail/. [Accessed 21 May 2026].
[6] IBM Institute for Business Value, “The True Cost of Poor Data Quality,” 2025. [Online]. Available: https://www.ibm.com/think/insights/cost-of-poor-data-quality. [Accessed 21 May 2026].
[7] World Economic Forum, “Future of Jobs Report 2025,” World Economic Forum, 2025.
[8] Travel and The World Tour, “Travel and The World Tour,” Travel and The World Tour, 13 July 2025. [Online]. Available: https://www.travelandtourworld.com/news/article/now-american-airlines-announces-new-ai-rebooking-biometric-screening-and-real-time-app-improvements-to-support-record-summer-travel-demand/. [Accessed 5 May 2026].
[9] R. Ord, “Web Pro News,” Web Pro News, 1 May 2024. [Online]. Available: https://www.webpronews.com/lemonade-ceo-ai-drives-swift-claims-and-surges-profits/. [Accessed 25 May 2026].
[10] Avasant, “EBS 26 | Leapfrogging Enterprise Transformation: From Automation to Autonomy,” 2026. [Online]. Available: https://www.youtube.com/watch?v=acJy2AqN5Zg&list=PLBgGU0w8a0c4xZ6HdA0NW4r8tP4p8we4q&index=18. [Accessed 21 May 2026].
By: Julen Del Arco
Avasant’s research and other publications are based on information from the best available sources and Avasant’s independent assessment and analysis at the time of publication. Avasant takes no responsibility and assumes no liability for any error/omission or the accuracy of information contained in its research publications. Avasant does not endorse any provider, product or service described in its RadarView™ publications or any other research publications that it makes available to its users, and does not advise users to select only those providers recognized in these publications. Avasant disclaims all warranties, expressed or implied, including any warranties of merchantability or fitness for a particular purpose. None of the graphics, descriptions, research, excerpts, samples or any other content provided in the report(s) or any of its research publications may be reprinted, reproduced, redistributed or used for any external commercial purpose without prior permission from Avasant, LLC. All rights are reserved by Avasant, LLC.
Login to get free content each month and build your personal library at Avasant.com
