When Your Work AI Fails, You Might Get the Blame
Somebody at your company is going to tell you to “just use the AI agent” for a task this year, and nobody is going to tell you who’s on the hook when it gets the task wrong. That’s not a guess. It’s the actual finding in a survey released October 13, 2026 by Wharton Human-AI Research and GBK Collective, the fourth annual study of its kind, covering 1,416 U.S. enterprise leaders and employees. When an AI agent makes a mistake on the job, the companies in that survey can’t agree on whose fault it is. And a quarter of the time, the answer lands on the employee — not the tool, not the company that told them to use it.
Sit with that for a second, because if you’re early in your career and getting pushed to use AI tools constantly, this is the first hard data that tells you what happens when the thing you were told to trust gets it wrong.
The short version
| What the survey found | What it means for you |
|---|---|
| When an AI agent makes an error, blame splits between governance/risk teams (28%), the employee’s manager (26%), and the employee themselves (25%) | There’s no agreed-upon answer. You could be the one left holding it. |
| 1 in 5 managers say they don’t know who’s accountable for AI agent errors at all | Your own manager may genuinely have no plan if an agent-driven mistake lands on their desk. |
| Manager daily AI use rose to 62%, up from 46% last year | Leadership isn’t pulling back on AI. They’re leaning in while accountability stays undefined. |
| Managers describing themselves as “excited” about AI fell 8 points to 51%; about a quarter now call themselves skeptical | The people above you are souring on the tools faster than they’re clarifying who owns the risk. |
| Just 17% of departments let an AI agent finish a task without a human approving every step | Most companies know the agent needs a human watching it. Few have said in writing which human eats the consequences. |
Who’s actually accountable when AI makes a mistake at work?
Ask three different people at the same company this question and you’ll likely get three different answers, because the survey found the responsibility for an AI agent’s mistake splits almost evenly three ways:
- Governance or risk teams — 28%. The people who set AI policy get named most often, but barely more than the other two groups.
- The employee’s manager — 26%. Almost a dead heat with governance. The person signing off on your work gets blamed nearly as often as the compliance department.
- The employee who used the tool — 25%. One in four times, the person closest to the keyboard takes the fall, even though they didn’t build the model and usually didn’t write the policy either.
Add it up and you get a company where no single group owns more than 28% of the blame for an AI agent’s mistake. That’s not consensus. That’s a coin flip with a third option bolted on. And on top of that three-way split, the same survey found that 1 in 5 managers say they’re not sure who’s accountable at all — meaning a fifth of the people who’d normally decide this for you have genuinely never thought it through.
Why nobody’s agreed on this yet
Here’s the part that explains the confusion: most companies haven’t finished deciding how much independence to give these tools in the first place. The same Wharton-GBK study found two-thirds of departments use AI agents in some form, but only 17% let an agent finish a task without a person checking each step along the way. The largest group — 48% — sits in what researchers call the “supervised middle”: the AI does the work, a human reviews it, then it goes out the door.
That middle zone is exactly where accountability gets murky. If an agent drafts something and a human signs off on it, and it’s wrong, who screwed up — the tool that drafted it, or the human who was supposed to catch it? Companies are still writing that rulebook in real time, which is why governance teams, managers, and employees are all getting roughly equal shares of the blame. Nobody built the permissions and the consequences together. They bolted the second one on after the fact, and it shows.
Managers are souring on AI at the exact moment it matters most
Here’s a number that should make you pay attention, not less. Managers and above reported their daily AI use climbed to 62%, up sharply from 46% the year before. They’re not stepping back from the tools. They’re using them more than ever.
But their enthusiasm is dropping at the same time. The share of managers calling themselves “excited” about AI fell eight points, to 51%, while roughly a quarter now describe themselves as skeptical. Wharton’s Sonny Tambe, who co-directs the research, put it simply: “The amazement is kind of wearing off a little bit.” He noted something I think you should underline — managers are now saying out loud what Gen Z workers told Wharton researchers back in January: that leaning on AI too hard is making colleagues “dumb and lazy” and quietly eroding the critical thinking that used to come with the job.
Read those two findings together and you get a strange picture. The people running your company are using AI constantly, trusting it less by the month, and still haven’t written down who’s responsible when it breaks. That’s not a stable arrangement. It’s a company running faster than its own rulebook, and someone eventually pays for the gap.
Why the newest person in the room carries the most risk
I’m not going to pretend this part surprises me. It’s the same pattern that shows up almost everywhere in a career, and the first stretch of your career absorbs more risk than any other stretch — not because you’re less capable, but because you have the least power to push back when something above your pay grade goes sideways.
Think about who’s actually running the AI agent day to day. It’s rarely the governance team. It’s rarely a senior executive. It’s the newest hire, the one told on day one to “use the AI tools, that’s how we work now,” with no training on where the tool lies, no seat at the table when the policy got written, and no political capital saved up if the output turns out wrong. When a quarter of the blame for an agent’s mistake lands on “the employee,” that employee is disproportionately going to be someone a few years into their career, not twenty.
That’s the sentence I want you to actually hear: using the tool your company handed you is not the same as being protected by the company that handed it to you. Those are two different promises, and right now, a lot of companies have only made the first one.
What this means if you’re using AI at your first job
None of this means stop using the tools. AI isn’t ending the career ladder — it’s rearranging which rungs exist, and refusing to touch the AI agent at your desk isn’t a protest, it’s just a slower way to fall behind. But “use the tool” and “use the tool with your eyes open” are different instructions, and this survey is exactly the data that tells you to pick the second one.
A few things worth doing differently starting this week:
- Ask the direct question before you need the answer. “If this agent’s output turns out wrong after I’ve used it, who’s accountable — me, you, or risk and compliance?” Ask your manager that in a calm moment, not a crisis. If they don’t know, that’s useful information too — it tells you nobody’s decided yet, which means the default answer might end up being you.
- Keep your own paper trail. Save the prompt, the output, and what you changed or flagged before it went out. Not to cover for a mistake — to show your judgment was in the loop, which is the exact thing that separates “the employee who blindly trusted a tool” from “the employee who used a tool and verified it.”
- Treat “approved by a human” as your job, not a formality. The survey found 48% of companies put a person in the approval seat before an agent’s work goes out the door. If that person is you, that step is not busywork. It’s the actual job you’re being paid to do, and skipping it is where the real risk lives.
- Don’t let speed talk you out of the check. The whole appeal of an agent is that it’s fast. The whole reason accountability research like this exists is that fast and wrong is still wrong, and it still has your name closer to it than you’d like.
I know it feels strange to be the one doing the extra verifying when the entire pitch of the tool was “it saves you time.” But the raise, the promotion, the next job — the conversation where you ask for more money is a conversation about trust, and trust is built or lost in exactly these unglamorous moments nobody’s watching. Being the person who caught the agent’s mistake before it shipped is worth more to your career than being the person who shipped fast and got blamed for it later.
The takeaway
A tool with no agreed-upon owner for its mistakes isn’t actually safe to use blindly, no matter how confidently it was handed to you. Use the AI. Use it daily, even. Just don’t hand it your judgment along with your task list — because right now, if it gets something wrong, the company backing you up is one-in-four odds of also being the company pointing at you.
Your job isn’t a life sentence and it isn’t a trap either — but it is a place where the fine print matters more than the pitch. Read the fine print on this one before you need to.
This article is part of the Career & Work collection.
Browse all Career & Work lessons →