Half of companies saw an AI agent make a mistake. Could you tell what yours did?

If your AI agent sent the wrong email last Tuesday, could you find out exactly what happened? New research says many teams cannot.

Nylas, a software company, released its State of Agentic AI report for the second half of 2026 on October 20. It surveyed 1,030 U.S. professionals who build, buy, or influence AI systems, from August 26 to September 12. 54% said an AI agent made a mistake in the past 90 days. The most common problem was a workflow the team had to reverse, reported by 31%. A quarter saw an agent update the wrong record, and 18% saw an incorrect or inappropriate email go out.

Here is the part that stayed with me. Only 53% can fully audit what their agents did last week. And roughly 1 in 4 let agents work through a shared inbox or an employee's own account. Teams that gave agents a dedicated identity were more likely to have full audit capability, 66% compared with 51% for agents acting through an employee's account. This is vendor research and self-reported, but the lesson is clear.

NIST makes a similar point. In an August blog post, two of its experts wrote that agents should have their own unique identifiers, tied to the person or system running them, and that sharing credentials creates accountability gaps.

Why this matters for a small business

Mistakes will happen, with people and with agents. What protects your client relationships is being able to see what went wrong, fix it fast, and explain it honestly. When an agent uses your personal login, its actions and yours look the same.

What I would do this week

• List every agent or automation that can send email, update records, or touch client data.
• Where your tools allow it, give each agent its own account instead of your login or a shared inbox.
• Pick one person to review the agent's activity log every week.

My takeaway: You cannot fix what you cannot see. Automate the process, not the relationship, and keep a clear record of who did what.