The adoption number looks good until you read the measurement number
Liferay's new survey, conducted through the Pollfish platform among 500 U.S. professionals involved in AI decision-making, implementation, or day-to-day use, found that 54 percent of companies are already running or piloting AI agents. On its own, that number would suggest agentic AI has crossed from experimental to mainstream faster than many analysts expected. But the survey's second headline finding undercuts that story considerably: only 25 percent of companies measure the impact of those agents using clear KPIs.
That gap, more than half of companies deploying agents, only a quarter able to say with confidence whether those agents are working, is the real story here. It means most organizations running AI agents today are doing so on faith rather than evidence, unable to point to a specific metric that shows the agent improved a process, saved time, or reduced cost. For CIOs reporting to boards on AI ROI, that is an uncomfortable position to be in during a budget cycle where AI spending is under increasing scrutiny.
Policy is lagging deployment by an even wider margin
The survey found only 24 percent of companies have a company-wide AI usage policy in place, an even smaller share than the 25 percent measuring impact with KPIs. Read together, these numbers describe a deployment pattern that is running ahead of both governance and accountability at most organizations. Agents are being stood up faster than the policies that would govern what they are allowed to do, and faster than the measurement systems that would show whether standing them up was worthwhile in the first place.
This sequencing problem tends to compound over time rather than resolve itself. Every month an organization runs agents without a usage policy, more informal, undocumented use cases accumulate, each one harder to retroactively govern once it becomes embedded in someone's daily workflow. Liferay's data suggests most companies are still in that accumulation phase, which means the governance catch-up work ahead of them is only getting larger the longer policy development lags actual deployment.
Accuracy, not cost, is the top worry
Respondents ranked accuracy as their top concern at 42 percent, well ahead of security and privacy barriers at 30 percent and cost barriers at 29 percent, with a lack of adequate training cited by 27 percent. That ordering is notable because it runs counter to a lot of the public conversation around enterprise AI risk, which tends to center on security and cost. Practitioners closest to actual agent deployments are more worried about whether the agent gets the answer right than about what it costs or whether it exposes sensitive data.
That concern connects directly back to the KPI gap. Without clear measurement frameworks, organizations have limited visibility into how often their agents are actually accurate in production, which means the 42 percent citing accuracy as a top concern may be operating on incomplete information about their own agents' real performance. Building the measurement infrastructure the survey says most companies lack would also, as a side effect, give organizations better data on the accuracy question that is worrying them most.
Adoption is deeply uneven across sectors
The survey found technology companies leading adoption at 72 percent, while education came in lowest at just 24 percent, a three-times gap between the most and least AI-agent-active sectors. That spread is wider than the gaps typically seen in earlier waves of enterprise software adoption, like cloud or mobile, and points to how unevenly distributed both the resources and the risk tolerance for agentic AI remain across industries. It also suggests the 54 percent overall adoption figure is masking real variation that CIOs benchmarking their own progress against industry averages should account for directly, rather than treating one blended national number as a meaningful comparison point.
The sectors lagging behind, education prominent among them, tend to share characteristics: tighter budgets, higher scrutiny of how technology affects vulnerable populations like students, and less internal AI expertise to draw on when building governance frameworks. That combination makes it harder for laggard sectors to close the gap quickly, even as pressure builds from peer institutions and vendors pushing agentic AI products into every market segment regardless of sector-specific readiness.
What actually predicts successful adoption
Liferay co-founder and CMO Bryan Cheung offered a specific diagnosis for what separates successful deployments from stalled ones: "AI agents help companies move faster when employees know how to use them and leaders set clear expectations for where they belong." That statement puts the emphasis on organizational readiness rather than technical capability, echoing a theme showing up across multiple 2026 enterprise AI surveys this year, that adoption success correlates more strongly with training and governance clarity than with which specific agent platform a company chose.
That diagnosis lines up with the survey's own numbers: the 27 percent citing inadequate training as a barrier and the 76 percent lacking a company-wide usage policy describe exactly the organizational gaps Cheung is pointing to. Companies chasing agent adoption without first closing those gaps are, according to Liferay's data, more likely to end up among the 75 percent that cannot measure whether their investment paid off, rather than the smaller share that can point to clear results.
The takeaway for CIOs building 2027 AI plans
The practical lesson from Liferay's survey is that agent deployment speed and agent deployment success are not the same metric, and organizations optimizing for the former risk landing in the 75 percent that cannot demonstrate the latter. Before expanding agent pilots further, CIOs should prioritize building the measurement infrastructure and usage policy that most peer organizations still lack, since both are cheaper to build early than to retrofit onto agents already embedded in production workflows.
With adoption already at 54 percent and accelerating, the window to build governance and measurement proactively rather than reactively is closing. Liferay's data suggests most companies will spend 2027 catching policy and KPI infrastructure up to deployment that has already happened, a sequence that tends to produce weaker outcomes than building the two in tandem from the start of an agent program. The organizations that treat measurement as a launch requirement rather than a later addition are the ones most likely to show up in next year's version of this survey inside the smaller, better-positioned group that can actually answer whether its agents are working.
