A Nylas survey finds 54% of firms had an AI agent error in 90 days

Nylas, a company that sells email and calendar data tools, released a survey on 20 October 2026. The survey had 1,030 US professionals who make, buy or influence AI systems. Of these, 54% reported an AI agent made a mistake in the last 90 days. Only 53% can fully audit the actions of their agents in the last week. An AI agent is software that works on its own.

How agents fail

The most frequent mistake was a workflow that the team had to cancel (31%). Nylas did the survey from 26 August to 12 September 2026.

The numbers

More agents bring more errors. Of firms with one to five agents, 45% had one. Of firms with more than 20 agents, 70% had one. Where agents have their own identity, 66% of teams can fully audit them. Where agents use the account of an worker, 51% can.

What is not known

The CEO of Nylas says teams that extend autonomy safely can say which identity an agent used. Nylas sells agent accounts that give an agent its own identity. The survey shows a connection between identity and audit, not a cause.

Insight: A vendor survey can be a reason to buy

ServiceNow found that 88% of UK firms had an agent incident in a year. Nylas finds 54% had an agent error in 90 days, and it sells identities for agents. The data reads as a reason to buy the product that Nylas sells. Use the numbers as a sign, and check your own records before you buy.