Findings from the 2026 Agent Readiness Survey. 67 senior leaders. 8 industries. Fieldwork May 2026.
The conversation about AI is shifting. A year ago, organizations asked how their people would use AI. Today, the question is what work AI agents will do on their own, and what happens when they do it badly.
This is the inaugural ZAI Institute Agent Readiness Survey, fielded in May 2026. The 67 respondents represent the working senior leadership of the enterprise: 21 percent C-Suite, 27 percent VP or EVP/SVP, 18 percent Director, 19 percent Manager, across technology, professional services, healthcare, financial services, manufacturing, consumer goods, government and nonprofit. They are deploying agents at scale.
The headline finding is the same pattern this institute has documented across the AI Readiness Gap and the AI Skills Gap reports: the strategy gap is wider than the skills gap. Organizations can deploy agents. Most cannot tell you who is accountable when an agent acts in their name, what the policy says about agent behavior, or whether the agent is doing better work than the human it replaced. The work of deploying agents is now ahead of the work of governing them.
Nearly seven in ten respondents are past the exploration phase. Yet fewer than four in ten have a formal policy governing how agents should behave, who reviews their work, and what to do when they make mistakes.
The gap is concrete. Of the 46 organizations piloting or running agents in production, only 24 have a formal oversight policy. That leaves 22 organizations running real agent activity on informal practice, work-in-progress policy, or nothing at all. The pattern is consistent across industries. The deployment curve is not waiting for governance to catch up.
Must have guidelines, process to trust and verify in place. — Director, Technology
The survey asked who is accountable when an agent acts on behalf of the organization. More than a third of respondents could not give an answer.
The 37 percent who answered "undecided" or did not answer is the headline number here. When an agent makes a bad call, that share of organizations does not have a designated owner of the consequence. Of those who do have an answer, the most common is a human reviewer (27 percent) followed by the department that deployed the agent (21 percent). Only 2 percent name the vendor. Enterprises are taking the accountability on themselves. Most just have not finished deciding who, internally, that means.
More than a third of organizations are running agent programs without a designated answer to the question "who is responsible if this agent does something wrong." Accountability is not a policy problem you can patch after a bad incident. It has to be decided before the incident, or the incident decides it for you.
The survey asked respondents to compare how much they trust an AI agent versus a human to make a high-stakes decision in their domain. Even among organizations running agents in production, humans still win the trust comparison.
Fifty-five percent trust humans more (40 percent slightly, 15 percent much more). Twenty-four percent trust them equally. Twenty-one percent trust agents more. The split is meaningful because it cuts across deployment status. Among the 23 respondents running agents in production today, only 5 (22 percent) trust agents more than humans for high-stakes decisions. The remaining 78 percent trust humans at least as much as the agents they have already deployed. Trust calibration is the unfinished work that happens after deployment, not before it.
Agents need clear feedback and evaluations. They are like the most capable interns ever. — C-Suite, Professional Services
Respondents were asked what share of work in their function they expect to be performed by AI agents within two years.
Sixty-three percent of respondents expect agents to do at least 15 percent of the work in their function within two years. Twenty-eight percent expect 30 percent or more. The forecast is bullish on volume. Layer that forecast against the oversight numbers from Finding 1 and the accountability numbers from Finding 2, and the picture sharpens. Most enterprises are planning to put substantial agent volume on top of a governance infrastructure that is not yet built.
Business leaders preparing for an agent-driven workforce should also understand: work design must be rebuilt, not just updated. Roles, workflows, and operating models need restructuring so humans and agents collaborate by default. — Director, Technology
When asked to identify the single biggest barrier to scaling AI agents in their organization, respondents pointed clearly at trust-and-governance concerns. Budget ranked fourth.
Security and accuracy together account for 55 percent of biggest-barrier responses. Internal expertise adds another 16 percent. The three barriers leading the table are all trust-and-capability concerns. Budget, at 10 percent, is not the primary blocker. Workforce resistance, at 5 percent, is almost nowhere. The workforce is not refusing to adopt agents. The leadership is not refusing to fund them. What is missing is confidence that the agents can be trusted with the work, and the operational scaffolding to verify that trust at scale.
It is still garbage in, garbage out. Train on curated data, not bulk data. — C-Suite, Technology
Respondents identified the functional areas where they see the most opportunity for AI agents. Customer service led decisively.
Customer service is the consensus first landing zone. Forty-six percent of respondents named it as an opportunity area, well ahead of any other function. IT and engineering, marketing, research, and knowledge management form a tight second cluster in the low thirties. Finance, sales, HR, and compliance trail. Executive support is almost an afterthought.
The functions leading the list are not coincidental. They share three properties: high volume of repetitive interactions, clear ground truth for whether the work was done correctly, and existing systems that can absorb new automation without restructuring the operating model. The functions trailing the list are the ones where ground truth is harder, the work is more judgment-bound, or the consequences of a bad decision are higher. The deployment curve is following the contours of where agents can be supervised.
There are compliance roles where agents fit well and are augmenting humans. There are excellent research tools providing more insights than we could have had prior, so we can improve faster. — Vice President, Technology
The 2026 Agent Readiness Survey paints a picture of an enterprise leadership class that is moving forward on agent deployment with confidence about the destination and uncertainty about the route. Two-thirds of organizations are past the exploration phase. Nearly two-thirds expect agents to do at least 15 percent of their function's work within two years. But the structural prerequisites for safe agent scale (formal oversight policy, designated accountability, calibrated trust) are present in only a minority of organizations. The barriers leaders themselves name (security, accuracy, expertise) all point at the same root: confidence that an agent can be trusted with the work in production.
This is the same pattern documented across the institute's prior research. The strategy gap is wider than the skills gap. Organizations can deploy AI. Most still cannot answer who is accountable for it, what the policy says about it, and whether it is doing the work better than the human it replaced. That gap is the work of the next eighteen months.
Thirty-seven percent of organizations cannot say who is accountable when an agent acts. That number has to come down before the deployment volume climbs further. The institutions that resolve it early will not face the question reactively when an incident forces it. Designate the role (human reviewer, deploying department, or named individual) and put it in policy before the next agent ships.
Sixty-nine percent are piloting or in production. Thirty-nine percent have a formal policy. The thirty-point gap is the operational risk surface. Oversight does not have to be perfect to be useful. A short policy that names the supervisor, the review cadence, and the escalation path is better than a long policy that takes a year to write.
Fifty-five percent of respondents still trust humans more than agents for high-stakes decisions, including respondents whose organizations are running agents in production today. That gap will not close on its own. The organizations that close it will do it through structured comparison work: evidence on where agents outperform humans, where they underperform, and where parity is the right outcome.
The skills the workforce now needs are the skills of oversight: policy design, accountability frameworks, audit and review practice, and the judgment to know when an agent's output should be trusted and when it should be challenged. University AI programs that treat this as the next phase, not a footnote to applied AI, will lead the next cohort of partnerships.
The functions leading agent deployment (customer service, IT, marketing, research, knowledge management) each have distinct oversight needs. A generic AI governance course will not meet the buyer where they are. Function-specific governance modules, taught by faculty with operational credibility, are the offering the data points to.
The 2026 Agent Readiness Survey was distributed through the ZAI Institute network to senior leaders in management roles across multiple industries. The data presented reflects 67 completed responses collected between May 5 and May 25, 2026.
The respondent pool represents C-Suite (21%), EVP/SVP (6%), Vice President (21%), Director (18%), Manager (19%), and other roles (15%) across Technology (27%), Professional Services (21%), Healthcare (12%), Financial Services (9%), Manufacturing (9%), Government and Nonprofit (9%), Consumer Goods (6%), and other industries (7%).
Responses are aggregated and de-identified. No finding is attributed to a named individual or named company without separate written consent. Selected free-text quotations have been lightly edited for clarity and length.