Adoption Doubled, but the Results Gap Only Widened
On the surface, 2026 looks like the year enterprise AI finally broke through. PwC's 2026 Digital Trends in Operations survey of 767 operations and supply chain leaders at U.S. companies found that enterprise-wide AI adoption doubled year over year, with 24% now running AI at full scale versus just 12% a year ago. Eighty-five percent of respondents even claim they are ahead of most competitors in digital transformation. Yet the same study reports that 89% say their technology investments have not fully delivered the results they expected. We read that pairing as the defining tension of the year: activity is accelerating, but value capture is not keeping pace.
That gap is the story every CTO and CIO should be studying. Doubling adoption while the majority of investment still underdelivers tells us the constraint is no longer the availability of capable technology. Models work. Copilots ship. Agents can be stood up in weeks. The problem sits downstream of the tools, in the operating structures that are supposed to turn capability into repeatable, accountable business outcomes. As PwC's data makes clear, most enterprises have proven AI can generate value in isolated settings. What they have not solved is how to make that value systemic, and that is a governance question long before it is a technology one.
The Pilot Trap: Disconnected Projects Masquerading as Strategy
Jenny Colapietro, PwC partner and digital core modernization platform leader, puts the diagnosis plainly. Technology, she argues, is not the biggest obstacle to AI success: the absence of governance, accountability, and operational readiness is. "The biggest issue is that companies still treat AI as a series of disconnected pilots instead of an enterprise-wide transformation," she told BankInfoSecurity in a July 6, 2026 interview. Too many organizations, she notes, never establish ownership, decision rights, or workforce alignment before they try to move past proof of concept, so their pilots succeed in isolation and then stall the moment they meet the real enterprise.
The survey data corroborates the pattern. Only 27% of organizations have fully embedded an AI strategy across their business units, which means nearly three-quarters are still operating AI as a portfolio of experiments rather than a coordinated transformation. We have seen this failure mode repeatedly: a promising pilot delivers a compelling demo, leadership funds a dozen more, and the organization mistakes a growing count of projects for a strategy. Colapietro's fix is structural, not technical. She recommends identifying a small set of critical decisions and redesigning them to operate in real time, with clear ownership, guardrails, and embedded agents. The unit of scale, in other words, is the decision, not the model.
Data Debt Is the Silent Tax on Every Initiative
If governance is the missing discipline, poor data is its most visible symptom. In the PwC study, 87% of leaders say weak data quality has undercut the value of their digital initiatives, and only 30% report significant improvement in their data foundations. This is the unglamorous work that pilots are designed to skip. A proof of concept can be hand-fed clean, curated inputs; a production process operating across business units cannot. The moment an agent is asked to act on live enterprise data, inconsistent definitions, stale records, and fragmented systems turn a working prototype into an unreliable one.
We would go further than treating data quality as a technical hygiene problem: it is an accountability problem. Someone has to own the definition of a customer, an order, or a supplier, and someone has to be answerable when that definition drifts. Notably, only 51% of respondents establish clean data foundations before scaling their initiatives, which means roughly half are scaling on top of known debt. The organizations that will pull ahead are the ones that treat data governance, master definitions, lineage, and ownership as a precondition for autonomy, not an afterthought to be reconciled once the agents are already running.
Agents Reach the Handoff and Stop
Nowhere is the readiness gap sharper than in agentic AI. Just 37% of leaders say they are comfortable assigning AI agents to execute full, end-to-end processes in operations. That number is the honest measure of trust in the enterprise today, and it is low for good reason. Full autonomy demands exactly what most organizations have not built: clear guardrails, defined escalation paths, auditable decision logs, and a human accountable for outcomes. Without that scaffolding, a CTO is right to keep an agent on a short leash, because an unsupervised agent acting on poor data through undefined decision rights is not efficiency, it is unmanaged risk.
This is why the agent conversation cannot be separated from the governance conversation. The 37% figure will not climb because models improve; it will climb because organizations earn the confidence to hand over control. PwC's own AI performance research reinforces the point: leaders that scale successfully are markedly more likely to run responsible-AI frameworks and cross-functional governance boards, and their employees show roughly twice the confidence in AI outputs. Trust, in short, is manufactured through operational design. The comfort to let agents run end to end is downstream of the discipline to define what they are allowed to do and who answers for it.
Governance as an Operating Discipline, Not a Compliance Checkbox
The word governance still triggers the wrong reflex in many boardrooms, conjuring policy documents and risk committees rather than operating capability. Colapietro's argument reframes it: boards should view AI governance as an operational leadership issue, not simply a technology challenge. That distinction matters. Compliance governance asks whether a system is permitted; operational governance asks who owns a decision, how it is measured, when a human intervenes, and how the process improves. The first protects the enterprise from AI. The second is what actually lets AI scale inside it. The two are related, but only the second closes the results gap.
PwC's leaders frame the same reality in blunt commercial terms. Global Chief AI Officer Joe Atkinson notes that "many companies are busy rolling out AI pilots, but only a minority are converting that activity into measurable financial returns." Encouragingly, 56% of operations leaders now foresee AI reshaping their operating models around capability-based roles like orchestration and governance, a sign that the discipline is starting to be recognized as a function rather than a formality. That recognition is the prerequisite for progress. Governance stops being a brake on AI and becomes the mechanism that lets an enterprise safely take its hands off the wheel.
The CTO Mandate: Redesign Decisions, Assign Ownership
For technology leaders, the takeaway is not to slow AI adoption but to change what they are scaling. "The biggest shift isn't adopting AI, it's redesigning how the enterprise makes decisions," Colapietro argues. "Most companies are still using AI as a tool to drive efficiency on top of existing processes: AI leaders are already transforming their operating model and driving growth aligned to business strategy." Bolting AI onto broken workflows produces impressive demos and disappointing returns, which is precisely the 89% results gap the survey captured. The work ahead is to pick the decisions that matter most and rebuild them around clear ownership and real-time execution.
We believe the 2026 numbers mark an inflection, not a disappointment. The doubling of full-scale adoption proves the appetite and the capability are real; the persistent value gap proves the discipline is not yet in place. The CTOs and CIOs who close that gap will be the ones who stop measuring AI progress by the number of pilots launched and start measuring it by decisions redesigned, data foundations owned, and processes an agent can be trusted to run end to end. Governance is not the tax on that ambition. On this evidence, it is the only path to it.
