Securing and Advancing Production Agentic Workloads in the Enterprise
Why Agentic AI Ambitions Are Hitting a Wall — and What Separates the 12% Who Break Through
Executive Summary
Enterprise leaders weighing agentic AI investments face a compounding challenge: spending is accelerating while evidence of return remains thin — and a landmark Wharton study now warns that if expected productivity gains do not materialize, the current buildout could represent the largest misallocation of capital in history.
Four signals define the moment:
- The ROI gap is widening. 95% of AI pilots fail to deliver measurable ROI (MIT). Only 14% of CFOs report clear, measurable impact from AI investments. 42% of companies scrapped most AI initiatives in 2025 — more than double the prior year.
- Agentic AI amplifies the risk. 88% of agentic AI pilots never reach production at scale. The governance, security, and integration gaps that kill standard AI pilots are even more consequential when autonomous agents are in play.
- AI builder business models must change. The economics of frontier model consumption are structurally unsustainable at enterprise scale. As token costs rise and ROI scrutiny intensifies, AI workloads will naturally migrate toward the edge — running on purpose-built, locally-trained models optimized for specific enterprise domains rather than general-purpose frontier models built for breadth.
- The winners are already moving. The 12% successfully scaling agentic AI are not moving faster — they are moving with greater discipline: bounded use cases, P&L-tied metrics, and governance built before the deployment, not after.
The enterprises that treat this moment as an accountability inflection point — and build toward secure, efficient, fit-for-purpose agentic workloads — will be positioned to lead the next era. Those that do not will be managing the fallout.
The Belt-Tightening Begins
Within a matter of weeks, the belt-tightening began. Companies started putting AI budgets on a diet — setting token limits, auditing spend, and asking a question they should have asked at the outset: what, exactly, are we getting for this?
Coinbase capped usage. Walmart restricted AI coding tools to limit duplicative requests. Amazon shut down its internal token leaderboard — the one that had gamified AI consumption and rewarded the heaviest users. The message from finance and operations leaders was unmistakable: burning tokens is not a business strategy.
And yet AI spending continues to soar. Technology and media companies spent an average of $66.29 per employee on AI in May 2026, up from $58.84 just a month prior. Investment is accelerating even as the evidence of return remains stubbornly thin.
This is the ROI reckoning — and for enterprise leaders navigating agentic AI, it may be the most important inflection point of this decade.
The Wharton Warning: A Business Model Reckoning
A new study by Wharton economists Jessica and Jonathan Wachter delivers the starkest assessment yet of where the industry stands. Technology companies, they find, are spending as if a productivity boom is inevitable. But if it does not materialize, "the current buildout will be the largest misallocation of capital in history" — with some major technology companies at risk of bankruptcy if productivity does not scale rapidly.
This is not a warning about AI capability. The models are capable. It is a warning about the business model — and it points toward a structural shift that every enterprise technology leader should be watching closely.
The economics of frontier model consumption are not built for enterprise scale. As token costs rise and ROI scrutiny from boards and finance committees intensifies, the pressure to find a more efficient architecture will only grow. The natural consequence: AI workloads will migrate toward the edge. Purpose-built, locally-trained models — highly optimized for specific enterprise domains, workflows, and security requirements — will increasingly outperform and outcompete general-purpose frontier models for the tasks that matter most inside the enterprise.
The shift from general frontier models to custom edge models is not a distant scenario. It is already being driven by the same economic forces now showing up in quarterly budget reviews. Enterprises that understand this trajectory will make very different infrastructure decisions today than those that assume frontier model access is the permanent architecture of enterprise AI.
The model router trend is an early signal: enterprises are now deploying routing layers to direct workloads to cheaper, more efficient models when full capability is not required. This is the first sign of mature AI portfolio management — and a preview of the edge AI architecture that will define the next phase of enterprise agentic deployment.
The Productivity Paradox at the Heart of Enterprise AI
At Uber, COO Andrew Macdonald surfaced what many executives are quietly confronting: there is no direct correlation between increased AI use and useful consumer features. More AI activity, in many cases, means more cost — not more value. His comments ignited what became known as the "tokenmaxxing reckoning," exposing a fundamental flaw in how enterprises have been measuring AI adoption.
"When a metric turns into a goal, it stops being a good metric. It's not about measuring people's productivity according to how many tokens they burn; that's absurd. The metric should be, what have you achieved? What have you been able to accomplish?"
— Enrique Dans, IE University
McKinsey has identified a "gen AI paradox" persisting across enterprises: companies cannot figure out how to scale AI across their operations even after heavy pilot investment. A National Bureau of Economic Research working paper based on nearly 6,000 executives found that roughly 90% of firms actively using AI reported no impact on productivity over the prior three years.
The Agentic AI Stakes Are Higher — and the Failure Rate More Severe
If the ROI gap is sobering for general AI deployments, it is even more pronounced for agentic AI, where autonomous systems are being tasked with consequential, multi-step workflows across enterprise environments.
MIT research found that 95% of generative AI pilots fail to deliver demonstrable ROI. The root issue is not model performance — it is that surrounding infrastructure, governance, and operational readiness were never scoped into the pilot. For agentic AI, this gap is wider still: 88% of agentic AI pilots never reach production at scale.
Only 14% of CFOs report a clear, measurable impact from AI investments. S&P Global data shows 42% of companies scrapped most AI initiatives in 2025 — more than double the 17% who did so the year prior. Finance leaders are losing patience, and boards are starting to ask the questions that should have been asked at the beginning.
Why Security Leaders Have the Most at Stake
For CISOs, the ROI reckoning lands differently. The pressure to enable agentic AI is real — but so is the liability when autonomous agents operate outside defined boundaries, access sensitive data without appropriate controls, or trigger compliance exposure that no one anticipated at the pilot stage.
The same failure patterns that produce zero ROI also produce security incidents. Rushed agentic deployments share three structural weaknesses: fragmented data environments that agents cannot navigate securely, integration complexity that creates unanticipated attack surfaces, and governance gaps that leave no one accountable for what the agent does once it is running.
Security leaders who are not part of the ROI conversation from day one will be left managing the fallout when the agentic deployment that skipped the governance step creates the breach that ends careers. The edge AI shift makes this even more consequential: as workloads move to locally-deployed, domain-specific models, the security perimeter, data governance requirements, and accountability structures all change fundamentally.
What Separates the 12% Who Break Through
Enterprises successfully scaling agentic AI share a common discipline: they identify repeatable, bounded use cases before they build. They define success metrics tied to P&L — not token consumption or time saved on isolated tasks, but measurable business outcomes. They treat the pilot as a governance audit, not just a technology demonstration.
The 12% who get agentic AI to production do not move faster than the rest. They move more deliberately. They close the gap between experimentation and accountability before they scale — because they understand that an agentic system reaching production without governance is not a success story. It is a risk event in motion.
The Path Forward
The next phase of enterprise agentic AI will be quieter than the last. Fewer sweeping announcements. More paused initiatives. More scrutiny from boards that have grown skeptical of narratives untethered from results.
This is not a reason for pessimism. It is a necessary correction — one that will separate organizations building durable agentic capabilities from those running expensive experiments that never compound into enterprise value. And it will accelerate the architectural shift toward edge AI: leaner, more secure, domain-trained models that deliver the efficiency and control that frontier model consumption has so far failed to provide.
Spreadsheets did not become the backbone of global finance overnight. AI will not either. But the enterprises that treat this moment as an accountability reckoning — not a temporary headwind — will be the ones positioned to lead when agentic AI matures into the foundational capability it has the potential to become.
The ROI scrutiny is not the enemy of agentic AI ambition. It is the prerequisite for it.