REPORTED DEVELOPMENT

A study published on arXiv on 1 October 2026 presents 240 trials of AI agents performing economic-value attribution tasks. Of 175 executions that passed all applicable final numerical checks, 162 (92.6%) contained at least one additional deviation detected by process checks.

The research distinguishes numerical outcomes from correct tool calls, operating sequences and data persistence. It covers one system and two related skills: 92.6% is not a general enterprise-agent error rate, nor necessarily the frequency of harmful errors.

WHY IT MATTERS TO THE MARKET

Reliable evaluation must separately verify the result, the authorised process and the integrity of records.

MARKET EVIDENCE

SOURCE-LINKED INDICATORS

Peer-accessible experimental evidence separates answer correctness from the integrity of the process producing that answer.

240AI-agent experimental trialsRESEARCH SAMPLEOCT 2026 / One study / one system and two related skillsPreprint experiment: 240 runs across two skills, specifications, harnesses and three models. The 92.6% denominator is 175 numerically passing runs, not all enterprise deployments.IBM RESEARCH / ARXIV / VIEW SOURCE
175Executions passing applicable final numerical checksRESEARCH RESULTOCT 2026 / Subset of the 240 research trialsAdditional process deviations occurred in 162 of 175 numerically passing runs. Narrow experimental population; not a universal agent failure rate.IBM RESEARCH / ARXIV / VIEW SOURCE
92.6%Numerically passing trials with at least one additional process deviationRESEARCH RESULTOCT 2026 / 162 / 175 numerically passing executions; not an enterprise-wide agent failure rateAdditional process deviations occurred in 162 of 175 numerically passing runs. Narrow experimental population; not a universal agent failure rate.IBM RESEARCH / ARXIV / VIEW SOURCE
WHAT THE DATA SUGGESTS

Evaluate economic attribution and executive dashboards using separate tests for numerical output, authorised operations, tool-call sequence and persisted records.

INTERPRETATION BOUNDARY

The 92.6% denominator is only the 175 numerically passing trials in one experimental setting. A process deviation is not necessarily a harmful operational outcome.

ORIGINAL REFERENCES