How to Preserve Professional Judgment When Employees Use AI: Five Practices That Keep People Accountable

Last updated Nov 8th, 2026

Among employees who use AI at work, about 2 in 3 say they have relied on its output without checking whether it was accurate, and more than half say they have made mistakes in their work because of AI. The figures come from a 2025 global study of more than 48,000 people in 47 countries, led by the University of Melbourne with KPMG.

In our AI adoption work at The Decision Lab, we find that organizations keep professional judgment intact by deciding which calls a named person must own, asking people to form their own view before they see the AI's answer, matching review effort to the stakes, and measuring how people perform when the AI is wrong or switched off. These practices sit inside the workflow, so they protect judgment without asking employees to use AI less.

Our view comes from the adoption side. When we mapped how employees at a Fortune 500 HR technology company moved from first exposure to everyday AI use across a 20,000-person workforce, two of the five barriers we found were low trust in AI accuracy and unclear usage guidelines. We see both as judgment problems. Employees hold back when they cannot tell how far to rely on the tool, and employees who rely on it without limits produce the errors the Melbourne and KPMG figures describe. This guide sets out five practices for keeping people accountable for AI-assisted work, covering decision rights, review, escalation, the manager's role and the feedback loops that show whether judgment is holding up.

1. Set decision rights. Decide when AI can act alone or only advise, and name the person accountable for each outcome.
2. Form a human view first. Record an independent assessment before seeing the AI's recommendation.
3. Match review to risk. Increase review and escalation requirements as the consequences of an error rise.
4. Hold managers accountable. Reward checking and challenge, and protect time for junior staff to build skills.
5. Measure judgment over time. Track overrides and caught errors, and test performance with the AI switched off.

Automation bias is the tendency to accept an automated recommendation in place of checking the evidence yourself. Kathleen Mosier, Linda Skitka and colleagues, studying pilots working with automated cockpit aids, described two forms of automation bias: commission errors, where people follow a recommendation that is wrong, and omission errors, where they miss a problem because the system did not flag it.

Expertise reduces the effect without removing it. In a 2023 study of automation bias in mammography led by Thomas Dratsch and published in Radiology, 27 radiologists rated 50 mammograms alongside suggestions they were told came from an AI system. When the suggestion was correct, the least experienced readers rated about 4 in 5 mammograms correctly. When it was wrong, they rated about 1 in 5 correctly. Radiologists with around 15 years of experience held up better, but their accuracy still fell from about 4 in 5 to under half.

AI also helps on some tasks and hurts on others that look similar. In a field experiment with 758 Boston Consulting Group consultants by Fabrizio Dell'Acqua and colleagues, consultants using GPT-4 produced higher-quality work faster on tasks within the model's capabilities. On one business problem chosen because it sat outside those capabilities, about 85 in 100 consultants working without AI reached the correct answer, compared with 60 to 71 in 100 of those working with GPT-4.

Habitual reliance can then change how much effort people put into thinking at all. A survey of 319 knowledge workers on generative AI and critical thinking, run by researchers at Microsoft Research and Carnegie Mellon University and covering 936 real examples of AI use at work, found that people who were more confident in AI reported less critical thinking, while people more confident in their own skills reported more.

Skills can also fade with time. At four endoscopy centres in Poland, doctors who had started using an AI system to spot precancerous growths during colonoscopy later found growths in about 22 of every 100 procedures they performed without the AI, down from about 28 of every 100 before it was introduced. A higher detection rate is better, because each growth found and removed lowers the patient's cancer risk. The Lancet Gastroenterology and Hepatology study on endoscopist deskilling was observational, so it cannot prove the AI caused the drop.

The costs reach colleagues too: in a survey of 1,150 full-time US desk workers by BetterUp Labs and Stanford's Social Media Lab, reported in Harvard Business Review's article on AI-generated workslop, 4 in 10 said they had received AI-generated work in the past month that looked finished but did not do the job, and each case took them nearly two hours to deal with.

Taken together, these findings point to a workflow design problem. The conditions around AI use affect how much judgment people apply, and organizations can deliberately shape many of those conditions.

Decision rights answer two questions for each type of work: how much the AI is allowed to do, and which person owns the result. When nobody answers them, employees answer them individually, and the answers drift toward whatever is fastest.

We suggest sorting decisions by how much harm an error could cause and how easily it could be reversed, into four tiers:

1. Automate. AI acts on its own. This tier covers low-stakes, easily reversed work, such as reformatting a document or summarizing internal notes.
2. Approve. AI drafts and a person approves. This tier covers work that leaves the team, such as client emails or reports, where a named person signs off before it goes out.
3. Advise. A person decides and AI advises. This tier covers consequential decisions, such as pricing a contract or rating an employee's performance, where the AI's output is one input among several and the decision owner records the reasons for the call.
4. Human-only. A person decides without AI. This tier covers decisions where regulators or the people affected expect the reasoning itself to be human.

Put the tier into the workflow itself, at the point where people meet the decision, so nobody has to remember a policy document. In every tier above Automate, "the AI suggested it" is never an acceptable explanation for an outcome.

Clear decision rights help in both directions: they tell cautious employees where AI is welcome and tell enthusiastic ones where it has to stop. They are also the most concrete form of AI governance in the workplace, because each one says who answers for an AI-assisted outcome.

Once people see an AI recommendation, it becomes the starting point they adjust from. The mammography study shows how strong that pull can be, even for specialists. For decisions in the Advise tier, ask the decision owner to write down their own assessment first, whether that is an estimate or a recommendation, then consult the AI and note where the two differ and why.

The difference itself is useful information. Agreement on an easy case tells you little, while a disagreement points to the part of the problem that deserves a closer look.

How the AI presents its output matters as well. The Dratsch team suggested showing radiologists how confident the system was in each suggestion. We worked on a related problem with Mila on the COVI contact tracing app, whose AI model calculated a graded risk score. We replaced blunt exposed-or-not alerts with graded risk guidance, and the work contributed to Apple and Google changing their exposure notification system from binary alerts to multiple risk levels. A graded output gives people something to weigh, where a single verdict invites them to accept or reject it whole.

Keep this step for decisions that need it. Adding a written step to routine work costs time and adoption without protecting much judgment.

Reviewing everything at the same depth wastes time on low-risk work and spreads attention too thin on the work that matters. Match the review to the tier instead, and let both review and escalation requirements rise as the consequences of an error rise.

For Approve-tier work, reviewers should go straight to the claims most likely to be wrong: figures, quotes and citations, names and dates, and any statement the author cannot explain. Ask the author to say how they checked the output, in a sentence or two, when they hand it over. AI-assisted work tends to look polished regardless of its quality, so a review that rewards finish will miss the errors that matter.

For Advise-tier decisions, add an independent review by someone who forms their own view before reading the AI's output or the author's conclusion.

Keep a shared, plain-language list of the task types where your tools have failed before. The consulting experiment shows that capable people cannot reliably tell from the outside which tasks an AI handles well, so a team's own record of past failures is the most practical guide to where review should be heaviest.

An escalation path tells employees what to do when something about an AI-assisted decision feels wrong. Without one, the easiest option is to go along with the tool.

Ask people to escalate in four situations: when the AI's output conflicts with their own expertise, when a case falls outside what the tool was set up for, when the AI cannot show where a claim came from, and when an error would be hard to reverse. Each escalation should go to a named person who answers quickly, and raising one should carry no penalty, even when the AI turns out to have been right.

Escalation works because it makes responsibility concrete. In the cockpit studies by Mosier, Skitka and colleagues, pilots who felt personally accountable for how they worked with the automation were more likely to double-check it against other cues and made fewer errors. Telling pilots they were accountable did not significantly change performance on its own. We take that as a warning against relying on policy statements, and as a reason to give people a concrete action they own.

Most AI rollouts give managers an adoption target. Few give them a judgment target, so the incentives point one way. A judgment target might include documented checks on high-risk work, recorded overrides, escalations raised and accuracy on periodic AI-off tests.

Managers protect judgment through what they ask about and what they reward:

1. Ask about the process. "What did you check, and what did the AI get wrong?" belongs in reviews of AI-assisted work alongside "Is it done?"
2. Reward caught errors and escalations visibly, so that finding a mistake in the tool counts as good work.
3. Use AI in front of the team the way they want the team to use it, including saying when they rejected its output.
4. Protect learning for junior staff. In the mammography study, the least experienced readers were the most affected by wrong suggestions, and the Microsoft and Carnegie Mellon survey links self-confidence to more critical thinking. Junior employees need some work they do unassisted, then compare with the AI's answer, so the expertise that makes review possible keeps developing.

Managers also need different conversations with different people. Our Fortune 500 diagnostic found two groups with opposite failure modes: enthusiasts who were already experimenting but hit friction, and skeptics who distrusted AI's value and stalled at first contact. Enthusiasts need to hear where the limits are. Skeptics need to hear that checking the tool is part of using it well, and that their caution has a place in the workflow.

Judgment that nobody measures will erode without anyone noticing. A feedback loop turns the practices above into data.

Log overrides of AI suggestions and errors caught in review, and look at the log every month with the people who own each workflow. Patterns in the log show where a task belongs in a different tier and where the guidance or the tool itself needs changing.

Every few months, test a sample of work with the AI switched off. The colonoscopy study only found the decline because doctors kept doing some procedures without the system, and most organizations have no equivalent check. Organizations can run this check without pausing AI use: test a small, representative set of tasks, or run periodic exercises where people work from the same inputs without AI assistance.

Test a new review step or manager routine against a comparison group before rolling it out. We ran the programs in our Fortune 500 adoption work this way, piloting them against a control group for a month, and we would use the same design for a judgment program: measure how often people catch planted errors alongside how confident they feel with AI. The Microsoft and Carnegie Mellon findings suggest the two can move apart, so rising confidence on its own says little about whether judgment is holding up.

The five practices do not need to launch at once. A low-friction first step is to apply them to a single workflow:

1. Choose one AI-assisted workflow where errors matter but are manageable.
2. Place it in one of the four decision tiers, and name both the decision owner and the escalation contact.
3. Define the checks required before work is approved.
4. Log overrides and caught errors for 30 days.
5. Review the log with the team and adjust the workflow before extending it to others.

We treat responsible AI adoption as one piece of work: increasing AI use and protecting the judgment that makes it safe. That usually starts with a behavioral diagnostic of how employees currently use AI and where judgment is slipping. We then design decision rights and review steps that fit real workflows, and pilot them against a comparison group before they scale. Our AI adoption case study with a 20,000-person Fortune 500 workforce shows how we run that process, and our guides to behavioral barriers to enterprise AI adoption and AI adoption training cover the adoption side in more depth.

What is automation bias in the workplace?

Automation bias is the tendency to accept a recommendation from an automated system, such as an AI tool, without checking it against other evidence. It shows up as following a wrong suggestion or as missing a problem the system failed to flag. Research on pilots and radiologists shows experienced professionals are affected too, though usually less than beginners.

Does using AI make employees worse at their jobs?

It can, depending on the task and how the work is set up. A field experiment with Boston Consulting Group consultants found GPT-4 improved work within its capabilities and reduced accuracy on a task outside them. Organizations limit the downside by setting clear decision rights and by periodically testing how people perform without AI.

Who is accountable when an AI-assisted decision goes wrong?

The person or role the organization named as the decision owner. AI tools cannot hold responsibility, so every workflow above the Automate tier should have a named person who approves or decides. Writing that owner into the workflow also makes people more likely to check the AI's output carefully.

What is human-in-the-loop AI?

Human-in-the-loop AI means a person stays responsible for reviewing, approving or making decisions that use AI output. How involved that person needs to be depends on how much harm an error could cause and how easily it could be reversed. In our four-tier model, the Approve and Advise tiers are human-in-the-loop by design.

How can you tell whether employees are over-relying on AI?

Look at behavior. Signs include work the author cannot explain, few or no recorded overrides of AI suggestions, errors reaching clients that a basic check would have caught, and weaker performance when the tool is unavailable. Logging overrides and running periodic AI-off tests turns these signs into figures a team can track.

Should junior employees use AI tools at work?

Yes, with some work deliberately kept unassisted. Junior staff build the expertise they will later need to review AI output by doing tasks themselves and comparing their answers with the AI's. Research on radiologists found the least experienced were the most likely to follow wrong AI suggestions, so this practice matters most early in a career.