AI Burnout, Shortcuts, and the Hidden Danger to Your Nonpublic Information

Somebody on your team is faster than they were a year ago and enjoying the work less. There's research now that says what you're seeing is real and common.

A 2026 survey of more than 800 Australian workers, reported in Psychology Today, compared the two groups: regular AI users reported burnout at 49% against 32% for infrequent users, were less likely to say they use their strengths (38% against 60%), and more likely to report self-doubt (53% against 37%).

Those figures compare how often people use AI. The explanation the researchers give for them is about which tasks people were left holding, and that distinction is what makes the problem fixable, so it's worth following.

What passive use does to a person

In passive use, the model handles the skilled part and the person reviews and approves. Somebody who spent years learning to write a scope, or read a plan, or shape a brief now spends the day checking whether a machine did it acceptably. The more volume the model handles, the less critical thinking the person is accustomed to using.

They use their strengths less, because their strengths went to the model. They doubt themselves more, specifically around language and grammar, because they are reviewing bad examples. Also, approving output on an exponential scale is draining in a way that creating it never is.

Be clear about what is measured and what is inferred there. The survey establishes that heavy users score worse on all three counts. The passive-use explanation is the reading the researchers offer for why, and the exhaustion step is ours.

Researchers writing in Harvard Business Review in May 2026 reached a compatible conclusion about the psychological costs of AI adoption: it can erode motivation enough to cancel out the productivity it delivered. Their recommendation was to change the system design.

The bottleneck moved onto your senior people

There's a second cost one level up.

HBR reported in May 2026 that individual contributors got faster while the systems around them stayed where they were. We'd read the consequence this way: work now arrives at review faster than review can absorb it, and the person holding that queue is usually the most senior person on the team.

You'll recognize it if approvals are the slowest part of your process now and weren't two years ago. Nobody on your team got worse at deciding. More decisions arrive per hour than your decision-making structure was built for.

The tier nobody planned

Inside one firm you usually end up with two populations doing the same work. One group has tools, training and a policy. The other has a personal account and no guidance or understanding of secure information governance.

The instinct is to schedule a workshop, which closes a knowledge gap. But the second group is usually short of access to sanctioned tools rather than knowledge, so after the workshop those people go back to the same personal account with the same client information that shouldn't be in it.

The shortcut, and what it exposes

A tired person takes the shortest path. With a model, the shortest path is pasting the whole thing in: the full client brief, the unredacted scope, the email chain with the budget in it. Redacting first costs four minutes and usually produces a worse answer, so at four in the afternoon nobody does it.

Passive use produces the exhaustion, the exhaustion produces the shortcut, and the shortcut lands in whichever account the person already had open, which for many teams is one the business doesn't administer.

Most owners hear "confidential" and picture a contract. In a design or professional practice it also covers the address of a private residence, a purchase price, a floor plan, the identity of a client who hasn't announced anything, unreleased project photography, a competitor's bid, and the trade pricing your vendors extend to you under terms that say you won't circulate it.

The exposure isn't that a model steals the material. On a personal account you have no agreement governing how it's handled, no admin view of what was submitted, no control over retention, and no way to revoke anything when that person leaves. An administered plan gives you all four, and it covers your client work under terms your business has actually read.

Three questions worth answering this quarter. Which of your people are working client material through an account the business doesn't administer? If one of them left tomorrow, could you say what they had put into it? And when your vendor agreement says trade pricing stays between you and them, does the tool holding your spec sheets know that?

The design move

Decide which tasks go to the model and which stay with a person, and make that call deliberately.

Send upstream work to the model: retrieval, synthesis, first drafts, formatting, the search through last year's projects for the thing somebody half remembers. Very few people's strengths live there.

Keep the last mile. The judgment call, the client conversation, the choice between three directions, the specification where being wrong is expensive. Those are where expertise gets exercised, and where a person's confidence that they can do their job gets renewed.

The test for any workflow is short. Name the part that requires somebody's judgment, then check that a person still does that part. Where the model has taken the judgment and left the checking, you've swapped them, and the people doing the checking are the ones the burnout numbers describe.

What you can do this week

Pick one workflow where you've added AI. Write down the step that needs somebody's judgment, then look at who does that step now.

If the model does it and a person approves it, swap them back. Let the model do the gathering and the first pass, and put the person back on the call that matters. Then ask them how the week felt, because they'll tell you before any metric does.

All Why Adoption Stalls Briefings