I used to trust a quiet incident queue.

A clean dashboard felt like evidence the rollout was holding. Few reports, low severity. These days I treat a quiet queue as the first thing to be suspicious of.

The dashboard the COO loved

Last month, on a coaching call with a chief operating officer at a mid-market insurance firm in Zurich. She pulled up the quarterly review for her claims-triage agent. 14 incidents filed. 12 of them low severity, 2 medium, 0 high. "We're calibrated now," she said. "The team knows what to flag."

I asked which failure categories the 14 reports came from. She didn't know. She pulled the raw list up on screen. The reports clustered around polite, obvious slips. The agent misquoted a policy line. It summarised a claim badly. A customer got sent the wrong contact person.

Six months earlier the team had drawn up a list of the categories that worried them most. She remembered two of them. The agent silently expanding its own scope. The agent exposing customer data in a reply. "We haven't had any of those," she said. "That's good, right?"

I told her I thought that was the problem.

What the paper actually measured

Mittal, Kim, Counts, Neville, Das and Laban published a user-centered audit of agentic failures this week. The team is at Microsoft Research and Georgia Tech. 163,487 Reddit posts across eight AI-agent subreddits. 82,951 unique users. 18 months of posts, plus a 73-participant evaluation study. They derived a taxonomy of 23 failure categories and asked users to rate each one for severity and noticeability.

Across the 23 categories, severity and noticeability are strongly negatively related. Spearman ρ=−0.61, p=0.002.

The failures users rated most severe are the ones they say they are least likely to notice. Security circumvention scored 4.24. Information exposure 4.14. Both are near-silent in the incident data.

Two more findings compounded for me. Users who had personally experienced a given failure rated it less severe than people reading a scenario about it. The gap was b=0.60 and b=0.84, both p<0.001. And the three most frequent failure categories in the Reddit corpus had the lowest research maturity (ρ=−0.46, p=0.03). Those three: unbounded execution at 15.4%, instruction violation at 13.0%, execution deviation at 11.5%.

Workers get used to the system. Their alarm fades. The queue goes quiet. Governance reads the quiet as calibration.

The COO's dashboard reads clean. It is answering a different question than she thinks. Her team is reporting what felt bad enough to flag, not what the agent did. Mittal's data says those two sets barely overlap. The move I asked her to make this quarter: stop collecting severity in the field. Collect what happened and let a reviewer outside the operation grade the severity against a fixed rubric. The quiet queue is what team comfort looks like from the outside. Comfort was supposed to be the last metric to watch.