A survey that confirms what the roadmap already suspected
Liferay's 2026 Agentic AI Adoption and Governance Report, fielded through the survey platform Pollfish, polled 500 U.S. professionals involved in AI decision-making, implementation, or day-to-day use at their organizations. The headline number is familiar: 54 percent of companies are running AI agents in production or piloting them. The more useful number sits two questions later. Only 24 percent of those same companies have a company-wide policy governing how agents get used, and only 25 percent measure the resulting impact with clear KPIs. More than half the market has agents doing real work with no enterprise-wide rulebook and no reliable way to say whether it is paying off.
We read that gap as the real story, not the adoption figure. A 54 percent adoption rate sounds mature until you notice that most of those deployments cannot answer a basic board question: what is this costing us and what is it returning. Liferay, a digital experience platform vendor, has a commercial interest in enterprises buying more governance tooling, which is worth keeping in mind. The underlying gap between adoption and measurement shows up consistently enough across other 2026 surveys that it reads as a real pattern rather than a vendor-manufactured one.
Adoption triples between education and technology
The industry breakdown is where the aggregate number stops being useful on its own. Technology companies report 72 percent adoption, three times the 24 percent figure in education. That spread tracks the obvious explanation: technology firms have in-house engineering capacity to pilot agents safely and the budget to absorb early failures, while education institutions typically have neither, plus procurement cycles built around annual budgets rather than quarterly experiments. Any sector-wide statistic on agentic AI adoption is really an average of very different starting positions, and a CIO benchmarking against a single number from a press release is benchmarking against the wrong reference class more often than not.
The practical implication for a laggard-sector CIO is to stop comparing adoption percentage against technology-sector peers and start comparing governance maturity against same-sector peers instead. A retailer or a healthcare system at 30 percent adoption with a documented policy and a measurement framework is in a materially stronger position than a technology company at 70 percent adoption with neither, even though the second number looks better in a vendor's case study slide.
Accuracy, not features, is what buyers are actually waiting on
Forty-two percent of respondents named better accuracy as the single change that would make agents more useful to their organization, ahead of any request for new integrations, faster response times, or additional model features. That is a telling priority ranking from people already running these systems day to day. It suggests the bottleneck most enterprises are hitting is not a missing capability so much as a reliability gap in the capabilities they already have, which is a harder problem to solve with a new feature flag and a much easier one to solve with narrower scoping and better evaluation before an agent goes live.
This also argues against the common procurement instinct to chase the vendor with the longest feature list. A platform that does fewer things correctly will outperform one that does more things unreliably, measured against what your own users actually say they need. Before the next agent pilot expands scope, it is worth asking the team running it whether accuracy or feature coverage is the binding constraint, because the survey suggests most teams would answer accuracy, and most roadmaps are optimized for the opposite question.
Data privacy sits at the top of every list that matters
Data privacy is the top concern cited by surveyed companies, and it also anchors the leading barrier to extracting more value from AI: security or privacy concerns, named by 30 percent of respondents, ahead of cost at 29 percent and lack of training at 27 percent. That ordering should inform where governance investment goes first. A company that spends its next governance dollar on training before it has settled what data an agent can access and under what review is solving the second problem before the first one, and the survey's own ranking suggests most respondents know it.
None of these three barriers is exotic or hard to name. What the survey suggests is harder is sequencing the response to them while the business keeps asking for more agent capability, not less. The 76 percent of companies without a company-wide policy are, in effect, running that sequencing question live in production rather than on a whiteboard, which is a more expensive way to learn the answer than the alternative.
What a defensible rollout looks like from here
Liferay co-founder and CMO Bryan Cheung frames the fix as three decisions every organization needs to make explicitly rather than by default: what data agents can access, which of their actions require human review before executing, and how the organization will measure whether the technology is actually helping. None of those three decisions requires new technology to implement. They require someone with the authority to make a call and write it down, which is precisely the step the survey's 24 percent policy-adoption number says most companies have skipped.
For a CIO or CTO reading this survey against their own agent rollout, the actionable test is simple: can you name, in one sentence each, what your agents cannot touch, who signs off before they touch something adjacent to that line, and which number on a dashboard tells you this quarter whether the deployment is worth its cost. If any of those three answers is missing, the survey says you are already in the majority. The fix costs a governance meeting, not a platform migration, and it is cheaper now than after the first incident forces the question.
