The number that should worry deployment teams
Sixty percent is the headline figure, and it is large. But the number that matters for anyone rolling AI out at work is the second one. More than one in five workers say their organisation discourages employees from challenging AI recommendations or decisions. A manageable error rate becomes a systemic risk the moment nobody is allowed to flag it.
The pattern behind that number is familiar. Workers catch mistakes regularly: a customer service issue was the worst consequence for one in five respondents, operational disruption for 13 percent, and nearly half said the most serious error they saw went beyond minor inconvenience into financial, compliance, or reputational territory. The workers are the last line of defence before a bad output reaches a customer or a business decision. In one workplace in five, that line has been told to stand down.
The caveat, up front
Read the methodology before quoting the numbers. This is vendor research. Appian sells AI automation for enterprise processes, and the study's conclusions point conveniently toward buying governance tooling. The sample is 500 workers in a single country, surveyed in Q3 2026. Treat it as directional rather than universal.
That said, the direction matches other signals. California banned AI-only firings earlier this month. Workers here echo the same instinct: 94 percent say AI should not make decisions affecting employees or customers without limits or human oversight, and only 14 percent fully trust the AI-generated outputs in use at their organisation.
The gap between agreement and action
The survey's companion finding comes from HBR Analytic Services: 92 percent agree that AI agents need rules-based guardrails, but fewer than 48 percent of organisations have defined them. Deployment is running ahead of control, and the people paying for the gap are the workers who are asked to trust outputs they keep catching in error.
If you run a team with AI in the workflow, the practical takeaway is unglamorous. Make it safe and normal to override the model. Log the overrides. Every worker who corrects the AI is free quality assurance; a culture that punishes corrections loses both the correction and the feedback the model team never sees. The cheapest governance mechanism in any AI deployment is an employee who is not afraid to say the machine is wrong.
Sources
- [1] Appian press release, "Six in 10 Australian Workers Have Caught AI Getting It Wrong at Work" (PR Newswire, Oct 8, 2026)Read source
- [2] Original release with research materialsRead source