AI is now making decisions that used to belong to people: who gets a loan, who gets an interview, and which batch passes quality control. But here is the part that rarely makes the headlines: someone still has to stand in front of a customer, a hiring manager, or a buyer and explain that decision. Increasingly, that person did not make the decision and may not fully understand it.
New research from Anne-Sophie Mayer, Elmira van den Broek and Tomislav Karačić, published in Harvard Business Review, followed employees across banking, recruitment and biotechnology for several years to see what actually happens when workers become the human face of machine decisions. Their findings should matter to anyone hiring for, or applying to, roles where AI is part of the workflow. In industrial and qualified sectors, that is fast becoming most roles.
Three ways people cope, and only one works well
The researchers found employees respond in very different ways when they are accountable for AI outputs they did not produce.
Some cover for the machine. Bank staff in the study invented plausible-sounding expert reasons for automated loan rejections rather than admit they could not interpret the system's logic. It protected their credibility in the short term, but quietly eroded customer trust when the explanations did not add up.
Some lean on the machine. Recruiters at a consumer goods firm used AI scores as ammunition to win over sceptical hiring managers. It worked, but with a sting in the tail: the more polished and self-explanatory they made the AI's outputs, the more invisible their own expertise became. The credit went to the software.
And some build a new expertise alongside the machine. Seed-quality specialists at a biotech firm were given access to the AI's underlying data, time to investigate its classifications, and regular contact with its developers. Over time they became fluent translators, able to connect the model's judgements to real operational decisions, spot its blind spots, and make it better. Their role did not shrink. It changed, and arguably grew.
What this means if you are hiring
The difference between those outcomes was not the technology. It was the organisation around it. If you are deploying AI in quality control, logistics, engineering or recruitment, the research points to a clear pattern.
Do not make people defend decisions they cannot question. If your operators, inspectors or advisors have no way to interrogate the system's outputs, they will either fake understanding or disengage. Build in access to the underlying data and a channel to raise concerns.
Redefine what "expert" means in the role, and say so in the job specification. The most valuable people in an AI-heavy workflow are often those who can interpret outputs and connect them to business outcomes, not those who compete with the algorithm. If that is the job, advertise it that way, and reward it in performance reviews.
Create feedback loops. Employees get better at explaining AI decisions when they regularly hear how those explanations land with customers, buyers or colleagues. One-off transactions teach nobody anything.
What this means if you are job hunting
If AI is part of the role you are applying for, ask about it in the interview. It tells you a lot about the employer.
- Can staff see the data behind the system's decisions, or just the verdict?
- Is questioning the system encouraged, or is the expectation simply to relay its outputs?
- How is interpretation work recognised? Is it in the job description, or treated as invisible glue work?
A company that expects you to be accountable for a black box is handing you the risk without the tools. A company that invests in helping you understand its systems is offering you something better: a skill set that gets more valuable as AI spreads, not less.
The machines may be making more of the decisions. But the people who can make those decisions make sense to customers, colleagues and regulators are becoming some of the most employable people in industry.
Based on research by Mayer, van den Broek and Karačić, "When Employees Are Held Accountable for AI-Generated Decisions", Harvard Business Review, July 2026.