Everything You've Been Told About AI Governance Slowing You Down Is Wrong

A new survey finds that companies with the most developed AI governance are more than three times as likely to have autonomous agents running in production.

Seventy-eight percent of the companies that rate their own AI governance as fully mature have autonomous agents running in production. Among the ones that call their governance developing, it's 22 percent.

The numbers come from Schellman's 2026 State of AI Governance, a survey of 525 U.S. professionals working in or around AI governance at organizations with more than 500 employees and at least $100 million in annual revenue. Researchscape fielded it between April 13 and May 11 of this year. The result gets one sentence on page six and no chart, which is roughly the attention it has received anywhere else.

Almost everyone assumes the relationship runs the other way. Oversight is overhead, the thing that adds three weeks to a launch, and agents put more strain on it than ordinary software does because an agent can act without being asked and can keep acting. On that assumption, the companies with the heaviest governance apparatus ought to be the slowest to put agents anywhere near production. The survey found the opposite, by a wide margin.

Where the deployments are stuck

Eighty-six percent of respondents have at least got agents into testing. Forty-six percent have them in production. That gap is where the interesting question lives, and very little of it looks like an engineering problem. The models work well enough. The integrations exist, and vendors are pushing agentic features into enterprise software whether or not customers asked for them.

What separates a pilot from a deployment is usually whether anybody can answer a short list of unglamorous operational questions: what happens when the agent touches billing, who signs off before it sends anything to a customer, whether a change it makes to an access control at two in the morning can be rolled back, and whether anyone will notice that it happened.

Companies that already have those answers tend to ship. Companies that don't usually attribute the delay to risk appetite or a cautious executive team.

A review model that can't grow

The same survey suggests the operating layer beneath most current deployments is thinner than it looks. Among organizations with agents in testing or production, the individual controls cluster between 52 and 57 percent. Fifty-seven percent have governance policies or guidelines specific to agents, and the same proportion say they have technical controls in place. Fifty-six percent conduct risk assessments, testing or formal audits. Fifty-four percent have assigned clear roles and accountability for agent decisions, and 52 percent have defined human oversight and escalation requirements. Those are per-control figures. The share of companies with all five is necessarily smaller, and Schellman doesn't report it.

The human review numbers are more revealing. Twenty-two percent have set defined thresholds that trigger a review. Thirty-eight percent say they review only high-risk or high-impact decisions, although nothing in the survey establishes that those firms have written down what high-risk means. The remaining 32 percent require review of every agent action regardless of risk.

That third arrangement is common and it doesn't hold up. If a person has to approve everything, the agent is expensive workflow software, and the deployment can't grow past the team that piloted it. These setups also rot quietly. Approvals become rubber stamps within a quarter or two, and the company ends up carrying the cost of review without the protection it was paying for.

Between the undefined and the unscalable, that's seven in ten companies with agents running on a review model that won't survive contact with scale.

What the ones who ship appear to do

Schellman reports that companies further along have converged on tiered internal risk frameworks, with the reasonable caveat that thresholds and definitions vary a lot by industry. The model in the report works roughly as follows.

Low-risk work, which covers internal process automation, expense reports, scheduling and the like, runs autonomously with a documented audit trail as the only control. Medium-risk work, including email drafting, customer communications and data analysis, executes immediately but gets reviewed within 24 hours and can be reversed if something is flagged. High-risk actions, meaning anything that touches financial systems, customer data, external communications or source code, require preapproval through a defined workflow. Critical actions aren't delegated at all: customer service agreements, legal commitments and data deletion stay with people, with the agent in an assisting role.

Within the high-risk band the framework splits further. An agent with repository access might push dependency updates and documentation changes under delayed review, while changes to encryption logic or access management need sign-off first. Policies that skip that distinction end up saying little more than that code is dangerous.

The reason to bother with tiers at all is that classifying actions in advance is what lets the low-risk majority of agent work run unattended, and unattended is where any return on the investment comes from. Companies without that structure tend to fail in one of two directions. They review everything and go nowhere, or they trust the agent and eat the incident.

Underneath the tiers sits a technical floor, which the report lays out as a checklist: timestamped logging of every agent action, including what data it touched and what came of it; rollback, so nothing an agent does is permanent by default; rate limiting; output validation before anything reaches a downstream system; integration controls restricting agents to approved systems and APIs; data residency enforcement; and traceable audit trails. None of this is new. Most of it is what a well-run shop already does for privileged service accounts. Agents need the same handling and frequently don't get it.

Reasons to hold the finding loosely

The 78/22 result is a correlation, and the report makes no attempt to establish direction. Causation could run the other way, or both numbers could be downstream of something else.

Organizations that have managed to operationalize AI governance probably also have platform engineering teams, working identity infrastructure, executive sponsorship and budget, each of which independently predicts getting agents into production. Governance maturity may be a marker of general institutional competence rather than the cause of anything.

Self-assessment is a weakness across the whole dataset. Maturity here is self-reported, and respondents are generous with themselves elsewhere. Seventy-four percent believe their organization could pass an AI compliance audit today, while only 27 percent describe their governance as fully mature and just 57 percent maintain a formal AI governance policy of any kind. If companies are inflating their maturity, the mature group in the 78/22 split includes a number of firms that are merely confident.

Then there's the sponsor. Schellman is an audit firm and an ISO 42001 certification body, so a report concluding that independent assessment is necessary and self-assessment insufficient isn't a neutral document. Its prediction that governance certifications will be contractually mandatory by 2027 is a forecast from an interested party, not a survey result.

None of which makes the number disappear. A three-and-a-half-fold gap in production deployment rates can absorb a lot of noise, and even if the mechanism stays contested, nothing here supports the story in which governance investment is the thing holding agent programs back.

For the next two quarters

If your agents are stuck in testing, the decision framework is a better place to look than the capability. Companies that got to production have generally done four things: classified agent actions by risk before deployment rather than adjudicating case by case, written down what triggers human review, built the logging and rollback layer that makes autonomous execution reversible, and spread accountability for agent decisions beyond whoever bought the tool.

Most organizations budget for that work under compliance, and it loses to feature work accordingly. On this evidence it belongs in the deployment plan instead.

Schellman's 2026 State of AI Governance surveyed 525 U.S.-based professionals involved in AI governance at organizations with more than 500 employees and $100 million or more in annual revenue. The survey was fielded by Researchscape from April 13 to May 11, 2026. Results are unweighted and self-reported.