Nylas: 54% of companies saw an AI agent error in the last 90 days

A Nylas survey of 1,030 US professionals found 54% had an AI agent make a mistake in the past 90 days, rising to 70% among teams running more than 20 agents.

Nylas published its State of Agentic AI report for the second half of 2026 today, based on a survey of 1,030 US professionals who build, buy or influence AI systems¹. Nylas says 54% had an AI agent make a mistake in the past 90 days, most often triggering a workflow that had to be reversed (31%) or updating the wrong record (25%)¹. The rate climbs with fleet size: 45% for organizations running one to five agents and 70% for those running more than 20¹.

The identity findings are the more useful part. Nylas reports that only 26% give agents their own identity, and about one in four let agents act through an employee's login or a shared inbox¹. Full audit capability was 66% where agents had their own identity and 51% where they borrowed an employee's account¹. A separate ServiceNow study of more than 3,700 senior security and technology leaders found 29% had discovered agents in production that were never registered, and 66% cannot act automatically when an agent behaves anomalously².

Vendors are reading the same signal. Microsoft announced Guardian Agents, which monitor the agents a company runs, and its security chief Hayete Gallot wrote that frontier models "cannot be trusted" because they are unpredictable, not malicious³. The tension is plain: adoption keeps growing while many teams still cannot say which agent changed which record.

For data teams this is a lineage problem before it is a model problem. If an agent writes to a warehouse table under a person's credentials, the audit log and the lineage graph both name the wrong actor. We should treat agents as first class principals: a dedicated role per agent, scoped grants, and query and write history we can filter by that role. Pair that with data quality checks on the tables agents touch, so a wrong update shows up as a failed test rather than a downstream complaint.

One step for this week: list every service account and user credential that an agent uses to reach your warehouse or pipelines. Any agent running on a human login gets its own role, with grants limited to the tables it needs.