Managers Split on Who Answers for AI Blunders, Wharton Finds

By Olivia Hart

When an AI system makes a consequential error, roughly one quarter of managers blame the employee who used it, another quarter blame that employee's manager, and a similar share point to the company's risk team. That is the central finding of a new study from the Wharton School at the University of Pennsylvania, based on a survey of more than 1,000 business leaders — managers, directors, executives, owners and founders.

The remaining responses scatter further. About 10% of respondents blamed the AI vendor, and roughly the same share admitted they were unsure who should answer for the mistake. In other words, for any given AI failure, responsibility is spread across at least four parties — or nowhere at all.

The timing matters. Google, OpenAI and Meta are all shipping AI agents with broader authority inside companies, which raises the potential impact of each error. Wharton's researchers frame this as a growing accountability problem as businesses hand more consequential decisions to machines. The study does not, notably, say how firms should resolve it.

What does misallocated blame cost in practice?

The abstract question of accountability has already produced concrete losses. Leeron Walter, a vice president at employee-monitoring software company Teramind, recently told Bloomberg about two cases companies have encountered.

At a major U.S. retailer, an employee used AI to automate customer emails. A mistake in the bot's instructions led it to send false refund and cancellation messages to more than 18,000 customers in under six hours, prompting thousands of service requests.

At a U.S. consumer electronics company, a marketing employee used AI to set and post product prices. A prompt error listed one of the company's flagship devices at an 85% discount for three hours. More than 700 customers placed orders before the company fixed the mistake, and its customer-service team spent weeks dealing with the fallout.

Both failures came from simple prompt mistakes, not from AI systems acting on their own. Walter's cases suggest the more immediate business danger is mundane: a vague instruction, amplified by an agent with live access to employee emails or customer data.

Why is the accountability gap widening now?

The major tech companies are giving AI agents far more than chat duties, and they are doing it this quarter.

• OpenAI introduced Dots late last month: always-on agents that connect to more than 4,000 apps and keep working on projects between user check-ins. OpenAI is also developing workplace versions of AI with their own credentials and access to company systems.
• Google announced its Gemini agent for enterprise users earlier this month. The tool plans and completes multistep work across company systems, including drafting content and writing code.
• Meta introduced Muse last month, a personal agent that manages email and travel bookings and keeps working on to-dos after a user leaves the app. Muse passed 2.5 million downloads in its first two weeks, according to Sensor Tower data shared with CBS News.

Each product moves AI from answering questions to taking actions — with credentials, system access and persistence between sessions. The Wharton data says companies have not decided who answers when those actions go wrong.

Is AI already inside the customer's purchase?

Partly, yes. IBM's recent research shows 41% of consumers use AI assistants to research products, and 31% turn to them to find deals. That means AI now shapes demand as well as back-office operations — an 85% mispriced flagship device reaches hundreds of buyers within hours precisely because discovery runs through automated channels too.

What should companies take from this?

The uncomfortable finding is what the survey does not show: no majority answer, no emerging standard, no role that a plurality of leaders agrees should absorb the blame. Every promise of agent-driven productivity in the Wharton framing carries the caveat that accountability is unresolved.

With OpenAI, Google and Meta shipping agents that act inside company systems this quarter, firms that deploy them without a written answer to "who owns the error" are running the Teramind scenarios — 18,000 false emails, 700 mispriced orders — without a designated owner for the cleanup.