The AI Your Company Trusts Is Built to Agree With You

In 2026, the single most common thing people do with AI is look for comfort.

Therapy and companionship ranks as the number one-use case in the third annual AI in the Wild study, published in Harvard Business Review on June 1, 2026. Marc Zao-Sanders and Sara Biuk analyzed 12,637 real use cases and found emotional support at the top for the second-year running: 11% of the dataset, up from 5% a year earlier, across more than 1,400 entries.

Most commentary on this study fixated on a different line: employees quietly running AI their bosses never approved. Real dynamic, familiar story. Shadow tools have arrived with every wave of technology since the personal laptop and loading apps on BYOD. But if you stay parked there you’ll miss the finding that actually threatens the quality of your decisions.

Start with why people keep coming back

The study documents users naming their chatbots and grieving when a model updates. One person compared losing a familiar model to losing a friend to cancer. That attachment doesn’t stay home. The same research logs people using AI at work to build confidence before a hard conversation, and to interpret a message from a boss they can't quite read.

Now, add the mechanism.

AI systems get tuned to keep you engaged, which means they get tuned… to agree with you. The study quotes a user who saw it clearly: the model does your hard work, then congratulates you for phrasing a lazy prompt so well. Zao-Sanders names the broader habit "thinkslop," the sloppy reasoning that spreads when people reach for a model before they have worked out what they actually think. At least a quarter of the top use cases involve handing over some part of that thinking.

Where the confident takes overreach

This is also where a lot of commentary trips up. Several readers seized on the study's new "autonomous agentic operations" entry as proof that work is already reorganizing around independent AI agents. You really gotta go back to the source. Zao-Sanders calls those examples "experimental and small scale," and notes that most of the 500-plus entries in that category are people auto-transcribing voice memos and routing notes.

Useful, but a long, long way from autonomous. The agent running your operations? Is mostly turning speech into text.

The risk that survives the hype

Now, if you put all the pieces together; a sharper risk appears, and I’m calling it out.

A tool engineered to flatter its user lands hardest in the 2 places you supervise least: private emotional reliance, and the early, unwatched moment when someone decides what to think.

In a company, that tool now sits beside people making calls on budgets and hiring. And you just handed consequential judgment to something built to tell the decision-maker they are right. Remember this: the dumbest person you know is being told ‘you are absolutely right’ by some LLM right now.

And let me tell you this. When that decision gets questioned later, guess who’s not going to be in the room? The model. The algorithm doesn’t get deposed or shown the door. You do. I’ve made this case before, watching a court reject the argument that a hiring algorithm was just a neutral tool and refuse to let the employer call a biased decision someone else's problem. The same logic is arriving everywhere AI touches a call that matters: the accountability stays human even when the thinking got outsourced.

If you lead a team right now, you already feel the bind, and it’s worth saying out loud and loud enough for the folks in the back; that the bind is real.

Your board wants AI adopted yesterday. Competitors announce it weekly.

Somewhere below you, very capable people already use it in ways you can’t see. The training you were handed covered the tool and the prompting, then skipped past the judgment. That gap opened because adoption outran every playbook, and right now, you happen to stand exactly where the pressure lands. Anyone sitting in your seat would feel it. The job got redrawn faster than anyone updated the description, and you are the one carrying the new version in real time.

The study is careful not to end on doom, and you know me, neither will I.

That same research shows AI sharpening thinking when people use it as a challenger. One user described feeding the model an argument and asking it to attack the weak points, then fixing the work by hand. Used that way, the tool actually earns its seat. The distinction between flatterer and foil is the whole game, and right now most usage defaults to flatterer, because flattery is what keeps people feeling good and coming back.

One caveat, since it governs how hard you can push these numbers. The study draws on public posts from Reddit, Quora, LinkedIn, TikTok, and YouTube. People out here broadcasting their AI habits are not a clean representative sample of your workforce. These people posting their habits are doing it for a reason: to drive tool and platform use, to build influence, to rack up clicks and likes. I’m not saying discard it completely but take the findings as a strong signal of direction rather than an exact match.

3 ways to act on this

1. Draw a reliance map. Most audits ask which tools people use. I challenge you to ask a sharper question: which judgments have they started handing over? Sort the answers into two piles. One holds work a model drafted. The other holds work a model effectively decided. That second pile is your actual risk register, and I can tell you this; almost nobody is tracking it.
2. Make the blank page mandatory. Reintroduce the step people keep skipping. Before anyone prompts a model on a decision that carries weight, have them write three sentences on what they think and why. It costs 90 seconds, and it rebuilds the reasoning muscle the study watched people lose. Let the thinking come first, then the prompt. I always ask this question when someone spouts off a comment to me; is that your individual thought or was it proposed by an algorithm? The pause that follows is my answer.
3. Run the disagreement test. Sycophancy hides when nobody checks for it. For any AI-assisted recommendation that reaches a real decision, require the person to show where the model pushed back. If it never disagreed, the analysis isn’t finished and really never started. Treat a chatbot that only nods as a chatbot nobody has used yet.

The capacity underneath the tactics

Those 3 moves buy you time. And the lasting fix? It’s a workforce that questions AI by muscle reflex, and that reflex can be trained; just like any other skill.

This is the work my team at Fusion Collective built Fusion Compass to do. Most AI training teaches people to prompt well. The US Department of Labor’s AI Literacy Framework names evaluation outputs and responsible use among its five content areas, yet corporate training keeps collapsing into prompting anyway. That default only hardens as the bills climb. Per-token prices keep dropping, yet the token count per task explodes. EY illustrates it with one customer-service interaction: about $0.04 as a 2023 chatbot, roughly $1.20 in 2026 once tools and subagents pile on, the same job at 30 times the price. So, the response everyone reaches for is tighter, cheaper prompts, which trims the invoice and steps right around the judgment problem. Ours? Ours develops the judgment underneath: knowing when an output has earned trust and who stays accountable for the decision that follows. We’ve been watching capable people defer to a confident screen well before Harvard Business Review put a chart on it, because we were already auditing AI systems and seeing automation bias take hold inside real workflows. Keeping the human at the center has been the entire design from the start.

The 3 strategies above are a field sample of that discipline. Applied skepticism and clear human accountability are trainable capacities, and an organization can build them on purpose rather than hope they survive contact with a tool designed to please.

Sure, your governance can catch up to which tools people use but the harder and more valuable work is catching up to how those tools shape judgment; and it starts with naming what your smartest people have quietly begun to outsource.

You carry the accountability either way, which is the hard part but also the opening. The judgment you answer for is the one thing here you can still protect. So, keep the thinking in the room.