Managers divide blame for AI agent errors into near-equal quarters, Wharton survey finds
Wharton School researchers found managers blame the deploying employee, their manager or the risk team for a costly AI agent error at about a quarter each. Only about 10% pointed at the vendor, so roughly 75% of the blame stays inside the company with no agreed owner.
The Investor · Invest desk
Key figures: about 75% of respondents named one of three internal candidates in total; roughly 10% blamed the vendor; roughly 10% said they were not sure who should be held accountable.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Three internal candidates in total, about | 75% | 9 | |
| Blamed the vendor, roughly | 10% | 3 | |
| Not sure who, roughly | 10% | 4 |
What happened
- Another group of roughly 10% of respondents said they were not sure who should be held accountable when an AI bot makes a consequential error.
- OpenAI, Google and Meta are pushing agents with their own corporate identities into employees' hands, and the bots can work toward goals independently, in some cases for several days.
- At a national U.S. retailer, an employee-built system sent incorrect refund and cancellation notices to more than 18,000 customers in under six hours, according to Teramind's Leeron Walter.
- At a U.S. electronics company, a prompt error listed a flagship device at 85% off for three hours and drew more than 700 orders, Walter said, and the fallout took weeks to resolve.
Why it matters
- exposure Employees who build or run agents are the most visible target: each cited incident traces to one person's prompt, and the deployer is one of three roughly equal candidates for blame.
- constraint At about 10%, vendors sit largely outside managers' expectations of blame, so a buyer should plan to carry an agent's error on its own books.
- cost Customer service absorbs the first cost whatever team wrote the prompt: thousands of inquiries at the retailer and weeks of cleanup at the electronics company.
- decision Executives deploying agents with corporate identities have to name an owner for their errors, and fund the review that owner needs, before an incident assigns one by default.
Add up the survey's shares and the total is 95%. Three internal candidates at about a quarter each make 75% [9], and the vendor's roughly 10% [3] plus a similar share who were unsure [4] bring it to 95% [10]. The other 5% is rounding or an answer outside Bloomberg's summary. Three-quarters of managers keep the loss inside the company and disagree three ways about whose it is.
The two incidents that Teramind's Leeron Walter described show how little time anyone has to settle that. More than 18,000 notices in under six hours is at least 3,000 an hour [15], and the electronics listing drew more than 700 orders in three hours, or at least 233 an hour [16]. Each of those orders was priced at 85% off [12], so 700 of them forgo 595 orders' worth of list-price revenue [17], assuming one device each and that the company filled them all.
"Organizations are learning in real time," said Jeremy Korst, a partner at GBK Collective and co-author of the Wharton study [7]. He added: "When I ask this question of a senior executive, the most frequent answer I get is that they hadn't thought about it yet." [8]
We think that answer is better evidence of a gap than the quarter-by-quarter split. A split could be a considered view that responsibility is shared, and managers were asked who should be held responsible [2]. That question asks for an opinion; it does not report their employer's policy. "Hadn't thought about it" means nobody has decided. Some employers may already have named an owner that their own managers have not read, and a check of written policies showing that most do would sink the thesis. Korst's remark is also his own sample of the executives he talks to.
Both incident accounts come from Walter, a vice president at Teramind, an employee-monitoring software company [14]. The retailer and the electronics company are unnamed. The incidents show what an error costs when it happens. Walter did not say how often such errors occur. Monitoring detects an error after it ships, while an owner who reviews prompts before they reach a live price list could stop some of them. Whether any company has created that role is an open question in the reporting. Walter said customer demand for employee-monitoring systems has ticked up after embarrassing mistakes like these [13].
What to watch
- The full Wharton report, for the sample size and whether respondents could pick more than one answer.
- Whether OpenAI, Google or Meta publish liability terms for agents that carry corporate identities, which would test the roughly 10% vendor share.
- Any company that discloses a dollar loss from an agent error, since the Bloomberg account measures both incidents in notices and orders.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence50
- Adoption
- Insufficient
- Hype gap+15
- Incentives55
- Confidence60
Perspective Coverage
3 publishers- Builder
- Builder 17%
- Operator
- Operator 60%
- Investor
- Investor 23%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
A report published Tuesday by the Wharton School of the University of Pennsylvania found little agreement among managers on who should be held accountable when an AI bot makes a consequential error.
- [2]
Asked who should be held responsible when a bot 'makes a consequential error,' respondents split roughly evenly between the employee deploying the AI agent, the employee's manager and the company's risk team, about a quarter choosing each.
- [3]
Roughly 10% of respondents blamed the vendor.
- [4]
A share similar to the vendor group, roughly 10%, said they were not sure who should be held accountable.
- [5]
OpenAI, Google and Meta are pushing increasingly powerful AI agents into the hands of employees; the bots have their own corporate identities and can independently work toward goals, in some cases over several days.
- [6]
A minor prompt error or unclear instructions can snowball into a larger crisis when agents have access to live systems such as email accounts or customer databases.
- [7]
"Organizations are learning in real time," said Jeremy Korst, a partner at the consulting firm GBK Collective and co-author of the Wharton study.
- [8]
"When I ask this question of a senior executive, the most frequent answer I get is that they hadn't thought about it yet."
- [9]
Three internal candidates at about 25% each total about 75% of respondents.
- [10]
The reported shares (75% internal, roughly 10% vendor, roughly 10% unsure) sum to about 95%.
- [11]
At a national U.S. retailer, an employee built an AI system to automate customer emails; a prompt error sent incorrect refund and cancellation notices to more than 18,000 customers in less than six hours, triggering thousands of customer-service inquiries.
- [12]
At a U.S. consumer electronics company, a marketing employee used AI to generate and publish product prices; a prompt error listed a flagship device at an 85% discount for three hours, more than 700 orders were placed before it was caught, and the customer-service fallout took weeks to resolve.
- [13]
Customer demand for employee-monitoring systems has ticked up in the wake of embarrassing mistakes like these.
- [14]
Leeron Walter is a vice president at Teramind, an employee-monitoring software company.
- [15]
More than 18,000 notices in under six hours is at least 3,000 notices an hour.
- [16]
More than 700 orders in three hours is at least about 233 orders an hour.
- [17]
700 orders at an 85% discount forgo about 595 orders' worth of list-price revenue, assuming one device per order and that all orders were filled.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
2 independent publishers whose own reporting we read for this story.
- cpapracticeadvisor.comWho’s to Blame if AI Fails at Work? Wharton Study Shows a Split
1 article · October 14, 2026
- entrepreneur.comWho’s at Fault When AI Makes a Serious Mistake? Managers Can’t Agree.
1 article · October 14, 2026
- news.bloomberglaw.comIf AI Fails at Work, Who’s to Blame? Wharton Study Shows a Split
1 article · October 13, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.