Nylas today released The State of Agentic AI: H2 2026, the second wave of its research into how organizations build, run, and govern AI agents. The survey of 1,030 U.S. professionals who build, buy, or influence AI systems found that 54% had an AI agent make a mistake in the past 90 days, most often by triggering a workflow the team had to reverse, and that only 53% can produce a full audit log of what their agents did last week and which agent did it.

When Nylas first surveyed this audience in December 2025, most respondents were still deciding whether to let an agent act on its own. Nine months later, more than half run six or more agents, and the incident rate rises with scale: 45% of organizations running one to five agents reported an incident, compared with 70% of those running more than 20.

“The category has moved from ‘will this work’ to ‘what do we do when it doesn’t,'” said Jeff Koets, CEO of Nylas. “More than half the companies in this survey have already answered that question at least once, by reversing a workflow, correcting a record, or apologizing to a customer. Teams extending autonomy safely can say who an agent acted as, what it did, and under which approval rule. Those answers depend on infrastructure decisions made before the agent ever sends an email.”

The report compares the two waves on the eight questions asked both times identically and finds a category getting more careful rather than more excited. The share of teams letting agents act without any human approval nearly doubled, from 4.2% to 7.5%, while the share requiring approval on every action fell eight points to 30%. Over the same period, fewer respondents said they fully trust an agent to run autonomously, and “meeting leadership expectations” nearly doubled as a stated reason for adopting agents, from 13.5% to 25.3%.

Key findings

  • 54% had an AI agent make a mistake in the past 90 days. 31% had an agent trigger a workflow the team had to reverse, 25% had one update the wrong record, and 18% had one send an incorrect or inappropriate email.

  • 53% can fully audit what their agents did last week. 37% can do so partially, and 10% either cannot or do not audit agent activity.

  • 26% say their agents authenticate with a dedicated identity of their own. At roughly 1 in 4 organizations, agents send email and manage calendars through an employee’s login or a shared inbox. Where agents have their own identity, 66% of teams report full audit capability, against 51% where agents act through an employee’s account.

  • Security approvals (42%) and integration complexity (40%) are the top blockers to implementation, ahead of engineering bandwidth (24%). Reliability (65%) and security controls (61%) lead what buyers want in an agentic AI platform.

  • 67% are considering replacing a vendor or tool with an AI agent. Separately, 81% have already replaced at least one vendor or tool with in-house tooling.

  • A third say agentic AI is already a baseline expectation in their market, and another 37% expect it within 12 months.

“How an agent gets into a mailbox decides whether anyone can later say what it did there,” said Troy Allen, CTO of Nylas. “When the agent runs on a person’s credentials, the log shows the person. The companies that gave their agents an identity of their own were the most able to audit them.”

The full report, including the methodology and the eight-question comparison with December 2025, is available without registration at nylas.com/agentic-ai-report-h2-2026. The next wave fields in early 2027. Nylas Agent Accounts, which give AI agents an email and calendar identity of their own, are generally available at nylas.com/products/agent-accounts.

About the research

Nylas surveyed 1,030 U.S.-based professionals between August 26 and September 12, 2026, through the Centiment research panel. Respondents work in engineering (33%), IT and internal systems (28%), product management (20%), data, AI, or ML (12%), and founder or executive roles (7%). Software and SaaS companies make up 48% of respondents. Only the eight questions asked identically in both waves are compared with the December 2025 survey of 1,026 respondents; all other figures describe September 2026 on their own.

About Nylas

Nylas is the communications data layer powering core product experiences and agentic AI workflows. Through a single integration, Nylas normalizes email, calendar, and contacts across Gmail, Microsoft, IMAP, and 250+ other providers, giving product teams reliable access to communications infrastructure without managing provider complexity.

As AI agents take on more work inside inboxes and calendars, Nylas provides the structured data and execution primitives they need to operate. Companies around the globe rely on Nylas to power productivity features, scheduling experiences, and workflow automation across more than 34.5 billion API transactions every month. Nylas has successfully completed a SOC 2 Type II examination, maintains ISO 27001 and ISO 27701 certifications, and undergoes independent HIPAA assessments. Our platform is designed with security and privacy controls that help customers support their obligations under applicable privacy laws, including GDPR and the CCPA/CPRA. Learn more at nylas.com.

Media gallery

About The Author
Business Wire