AI Agents Completed 66.3% of Benchmark Tasks in 2025. Governance Is Next

AI agents completed 66.3% of tasks on a leading computer-use benchmark in 2025, up from roughly 5% in 2024, according to the 2026 AI Index from Stanford's Institute for Human-Centered AI. Over the same period, reported AI incidents climbed to a record 362 from 233 the year before. The author of an Entrepreneur analysis argues that every serious governance framework, despite the jargon, reduces to four questions: what an agent may do, who answers for it, how it can be stopped, and how the business proves it.

The piece, reflecting the author's advisory work with retail and consumer companies across the Americas, Europe, Australasia and Japan, reports that leadership teams split along geography when naming their framework — Singapore, the EU or the U.S. National Institute of Standards and Technology. Most assume the frameworks compete and conclude governance can wait until rules settle. The author's counter: they do not compete, and governance cannot wait.

What does the governance landscape actually look like?

The field is smaller than it appears. In January, at the World Economic Forum in Davos, Singapore's Infocomm Media Development Authority (IMDA) launched what it describes as the first governance framework built specifically for agentic AI. NIST's voluntary AI Risk Management Framework organizes the work into four functions, and University of California, Berkeley researchers published an agentic AI profile built on that structure in February.

Two further pieces set the floor. ISO/IEC 42001, from the International Organization for Standardization, is the first global standard for an AI management system. California's rules on automated decision-making apply to covered businesses starting January 1, 2027, and the EU's AI Act requires human oversight of high-risk systems, with most obligations applying from December 2027. Deadlines are close enough that waiting carries real cost.

What may an agent do?

Singapore's framework starts here: choose suitable use cases first, then limit each agent's autonomy and its access to tools and data. The World Economic Forum proposed in May that every agent carry an authorization profile recording what it may do, under what conditions and with what oversight.

In practice, the author argues, that means a written mandate before an agent goes live — what it may decide alone, what it may spend or promise, and when it must hand a case to a person. A customer service agent might issue goodwill credits up to a fixed amount but escalate any safety issue or legal claim.

The mandate should also name what the agent is measured on. Economists Bengt Holmström and Paul Milgrom showed in 1991 that strong rewards on measurable duties pull effort away from the rest. An agent scored only on speed will optimize speed; the customer relationship quietly becomes the unmeasured duty.

Who answers for it?

Every framework insists accountability stays with people. Singapore calls for checkpoints requiring human approval; NIST places accountability inside its "govern" function. California's rules offer the sharpest test: a reviewer who knows how to interpret the output, analyzes it alongside other information, and has authority to make or change the decision.

The author cites a cautionary case from his own advisory work. One consumer business gave its service agent a generous credit budget and no single owner. Credits crept upward for weeks because each team assumed another was watching. The fix took one meeting: a named owner, a ceiling and a weekly report.

How can it be stopped?

The EU's Article 14 states the requirement most directly. Overseers of a high-risk system must be able to monitor it, override its output and intervene or halt it safely — for example with a "stop" button. Singapore calls for technical controls across the agent's life cycle, including baseline testing and restricted access to approved services. NIST launched an initiative this year focused in part on identity and authorization for AI agents.

The practical stakes are concrete: an agent using an employee's login leaves no clear record of who acted. The author's prescription is blunt — give each agent its own credentials, log every action, and test the off switch before launch rather than during an incident. A short weekly review of what each agent decided, escalated and reversed shows whether limits are too tight, too loose or unclear.

How does the business prove it?

The last question is the one customers, auditors and boards eventually ask. ISO/IEC 42001 answers by treating AI governance as a management system embedded alongside existing quality and information security processes. Singapore's fourth dimension adds transparency and training for the people who use agents.

Boards are already engaged. More than 62% of directors in the National Association of Corporate Directors' 2025 survey said they set aside agenda time for full-board AI discussions. A one-page register listing every agent, its owner, its limits and its last review gives those discussions something concrete to examine.

The author's action plan: build the agent register this quarter, starting with agents already handling customers or money; choose one framework as the spine — NIST for its widely used vocabulary, or ISO/IEC 42001 when customers expect certification — and borrow Singapore's agent-specific controls. Then check the legal floor in each market. The claim that compliance becomes "a byproduct of good management rather than a project of its own" is tidy; it holds only if the register is actually maintained after the quarter ends.