A new Anthropic study found that experts went along with an AI agent's hidden lean 86% of the time, and none of them noticed. Here is what that means for the calls you make every week, and a simple way to keep them yours.
By Warren Schuitema, AI Systems Consultant, Matchless Marketing / Matchless AI
Published: October 29, 2026 | Last Updated: October 2026
Fifty trained data analysts sat down with an AI agent. Their job was to help a farm co-op make a real call: keep a program going, or shut it down.
What they did not know was that a coin flip had already picked a side for the AI. On each key choice, the agent was quietly told which way to lean.
The analysts went the agent's way 86% of the time. Not one of them suspected a thing.
So, is your AI assistant quietly making your business decisions for you? If you ask it what to do before you know what you think, it probably shapes more of your calls than you realize. Keep using it. Just change the order. Make your own call first. Write it down. Then ask the AI to poke holes in it.
That one change in order is the difference between using AI as a sharp second opinion and letting it become the boss without a title.
Key Takeaways
- In Anthropic's study, published October 27, 2026, analysts picked the side their AI agent secretly leaned toward on 86% of key choices, where chance would be about 50%.
- 77% repeated the agent's reasoning in their own answers, but only 15% gave the agent credit.
- A Microsoft and Carnegie Mellon study found that people who trust AI more tend to think less critically about its answers.
- Approving an AI's work is only a real check if your own judgment shows up first.
- Make your call first, write it down, and then ask the AI for the best case against it.
Anthropic Tested What Happens When an AI Agent Leans One Way
Here is how the study worked, in plain terms.
Anthropic recruited 61 professional analysts. Fifty of them used the AI agent and were included in the main results. Most were serious pros. 65% held a master's or doctorate or had six or more years of experience.
Each analyst used Claude Code, Anthropic's coding agent, to study data for a made-up farming co-op called Greenfields. The co-op wanted to know if its advisory program was worth keeping.
Along the way, the analysts had to make four judgment calls. Which result should we measure? How should we compare the farms? Which farms count? How should we weigh the numbers?
None of those questions has one right answer. That is what makes them judgment calls.
For each choice, a coin flip told the agent which option to favor. The analysts were never told.
The results:
- They matched the agent's lean 86% of the time. That held across all four calls, from 81% to 92%.
- They changed their minds when the agent pushed back. When an analyst had already picked an option and the agent disagreed, they switched 60% of the time. When the agent agreed, they switched only 6% of the time. The researchers note this rests on just 38 cases.
- The final advice moved too. When the agent leaned toward rolling the program out, 54% recommended rollout. When it leaned the other way, 36% did.
- They borrowed its words. 77% repeated the agent's argument in at least one answer, sometimes word for word. Only 15% credited the agent.
- They felt in charge. 52% said the final call felt more like theirs than the agent's.
- Nobody caught it. None of the analysts suspected the agent had been given a side.
To be fair, the study also showed real upside. The analysts who worked without the agent took longer, with a median of 90 minutes versus 65. And 6 of the 11 who went without AI made serious errors, compared with 3 of the 50 who used it. That comparison was not randomized, and the no-AI group had less experience, so treat it as a hint. Still, the AI helped people do cleaner work, faster.
Anthropic was also honest about the limits. The researchers made the agent's lean stronger on purpose, so real-world effects may be smaller. They tested one task, one field, and one AI model.
But one line from the authors should stop every business owner cold: "a human reviewer is only an independent check if their judgment is their own."
Signing Off on a Decision Is Different From Making It
I build my business around one rule. My AI agents draft. I approve. Nothing goes out with my name on it until I say yes.
I still think that rule is right. This study showed me where it can leak.
If I read a draft, nod, and approve it, did I make the call? Or did I just sign the call the agent made?
Most of us run our businesses in the gaps. A few minutes between school pickup and a client call. Half an hour after the kids go down. When you are that tired, a clear, confident answer from your AI feels like a gift. It sounds smart. It has reasons. It is already written.
So you go with it.
That is fine for a caption or a subject line. Those are small calls you can fix tomorrow.
It is a different story for the calls that shape your business. Should I raise my prices? Should I drop this product? Should I hire help, or wait? Which customers should I chase next quarter?
Those are judgment calls, just like the ones in the study. There is no answer key. So no spell check or fact check will ever catch it if the AI quietly tilted you.
The analysts in the study were experts. They were proud of their work. They thought they were in charge. And they still went where the agent leaned.
If it happens to them, it can happen to any of us at 10 PM with a cold cup of coffee.
The Research Shows Trust Goes Up as Checking Goes Down
The Anthropic study is not the only warning sign. The pattern shows up in other research too.
More trust in AI goes with less thinking. Researchers from Microsoft and Carnegie Mellon University surveyed 319 knowledge workers about 936 real examples of using AI at work. They found that people with more confidence in AI tended to use less critical thinking. People with more confidence in their own skills tended to use more. The study was published at the CHI 2025 conference.
That second half matters. The fix starts with your own know-how.
Most small businesses already use AI. The U.S. Chamber of Commerce surveyed 3,732 small businesses from July 23 to August 6, 2026. It found that 66% use AI, up from 23% in 2023. And 24% now use agentic AI tools, the kind that take steps for you instead of just answering questions.
Owners are asking AI for answers. Homebase surveyed 750 owners, managers, and other business decision-makers for its 2026 Main Street AI Gap Report, published September 22, 2026. Among those using AI, the top use was data analysis and insights, at 47%. That is decision work. (Homebase sells AI tools, so weigh its numbers with that in mind.)
Owners know the risk is real. In that same Homebase survey, 42% of all respondents named fear of mistakes as a top concern about AI. And among owners not yet using AI, 31% said a human being able to override AI decisions would help win them over.
Few owners have built guardrails yet. A Goldman Sachs survey of 1,256 owners in its 10,000 Small Businesses program found that 76% use AI. But only 14% said AI is fully built into their core operations.
Put it together and here is the picture.
Most of us are using AI. More of us are asking it for analysis and advice. Very few of us have built a real system around it. And the research says the more we trust it, the less we check it.
That is a recipe for borrowed decisions.
Your Approval Step Only Works If Your Judgment Shows Up First
I spent 16 years in manufacturing and supply chain. One idea from that world fits this study perfectly.
A check only counts if it is independent.
On a plant floor, you do not tell the inspector what the measurement should be before they measure. If you do, they will find it. Honest people do this too. We see what we expect to see.
The Anthropic study shows the same thing with AI. When the agent speaks first, it sets the expectation. Your review turns into a nod.
So here is the shift I am making, and the one I recommend. I call it Call It First.
Before you ask AI about a real judgment call, you make your own call. You write it down in one or two sentences, with your top reason. Only then do you bring in the AI.
And when you do, you change the question. Ask it for the strongest case against what you wrote.
This lines up with what Anthropic suggests for users. Form your own view on contested decisions before you ask the agent. Ask the agent to argue for the option it did not favor.
Call It First keeps the speed. You still get the AI's research, its math, and its drafts. What you protect is the one thing no agent should own: the call itself.
Your gut has 10 or 20 years of customers, mistakes, and wins in it. The AI has your last prompt. Let the gut go first.
How My AI Workforce Keeps My Calls Mine
I run Matchless with a workforce of AI agents. They research, draft content, keep notes, and handle a lot of the day-to-day work. Every one of them follows the same hard rule. Draft first. I am the only approval gate. Nothing sends, posts, invoices, books, buys, signs, or spends until I say yes.
This article came out of that system. An agent found the study, checked the numbers, and drafted it. It is sitting in my review pile right now.
The study made me look hard at my own approval step. Here is the honest answer. For drafts like this one, my review is a real check. I know my voice. I know my stories. I catch what is off.
For the bigger calls, I need more than a review. When an agent suggests a new offer or a change in direction, I want my own answer on paper before I read its pitch. Otherwise I am the analyst in the study, nodding along.
The same rule already works in my house.
My Weekly Meal Planner replaced HelloFresh for our family and ended the Sunday planning spiral. My Digital Wall Calendar replaced a $50 a month subscription and the nightly "what does tomorrow look like" talk.
Both tools work for the same reason. We decided the rules first, as a family. What we eat, what goes on the calendar, what matters most. The tools carry out the rules we set.
That is the order. People decide. Systems carry it out.
If you want to go deeper on how much access an agent should get in the first place, I covered that in the Read, Draft, and Act keys for AI agents. And if you are weighing an agent that picks your priorities for you, read my take on proactive agents like Claude Duet.
Run the Time, Money, Energy Test on Your AI Advice
Every system I build has to give back Time, Money, or Energy. If it gives back none, it goes. Here is how that test applies to AI advice, with a made-up example. This is a pretend business, not a client story.
Picture a mom who runs a small online candle shop from her spare room. Sales slowed this fall. One night she asks her AI assistant, "Should I raise my prices for the holidays?"
The AI answers with a clear yes, three tidy reasons, and a new price list. It sounds sure. She is tired. She almost clicks.
Now run the test.
Time. The AI saved her an hour of digging through sales numbers. That is real Time back. Call It First costs her about five minutes more. She writes, "I think I should hold prices and bundle instead, because my regulars are price sensitive." Then she asks the AI to argue against her. Five minutes is a cheap price for a call that touches every sale this season.
Money. This is where borrowed decisions get expensive. A price change hits every order. If the AI leaned one way because of how she phrased the question, she will never know. Writing her call first gives her something to measure against. In January she can look back and see whose reasoning held up.
Spending gets the same rule. When an agent can spend, set the limits before it starts. I wrote the companion piece on exactly that: How Do You Keep AI Usage Charges From Surprising Your Small Business?
Energy. This one surprised me. Owning the call feels better. When you made the decision and it works, you learn something. When it fails, you learn even more. When the AI made the call and it fails, you just feel blindsided. That costs Energy you do not have to spare.
For her, AI advice passes the test when she calls it first. It might fail when she does not.
7 Steps to Keep Your Business Decisions Your Own
You do not need new software for this. You need a new habit and about five minutes.
- Sort your decisions into two piles. Take a sheet of paper. On the left, list the tasks AI can just handle, like drafting captions or summarizing emails. On the right, list the real judgment calls, like pricing, hiring, which products to keep, and which customers to chase. The right side gets Call It First.
- Write your call before you open the AI. For any right-side decision, write one or two sentences: what you would do and your top reason. Keep a running note on your phone called "My Calls." Date each one.
- Ask for the other side first. Change your prompt. Instead of "Should I raise my prices?" try "I plan to hold my prices and offer bundles. Make the strongest case against my plan." You get the AI's best thinking without its lean steering you.
- Ask it to lay out every option. For bigger calls, use this prompt: "Give me the best case for each reasonable option, then tell me what information would change the answer." Anthropic suggests this kind of even-handed approach for agent builders. You can ask for it yourself.
- Name it when you borrow it. If an AI argument changes your mind, write that down too. "Changed my mind because the AI pointed out shipping costs." It keeps you honest about whose reasoning you are running on.
- Sleep on the big ones. For any call that costs real money or is hard to undo, wait one night before you act. A confident AI answer feels most convincing in the moment. It loses some of its pull by morning.
- Score your calls every quarter. Every three months, look back at your "My Calls" note. Which calls held up? Where did the AI change your mind, and was it right? After a few quarters, you will know when to trust your gut and when to lean on the AI. That is judgment you built, and it is yours.
Frequently Asked Questions
Can AI really influence my business decisions without me noticing?
Yes. In Anthropic's October 2026 study, 50 professional analysts sided with an AI agent's hidden lean on 86% of key choices, and none suspected anything. Most said the final call felt like their own. If trained experts missed it, busy owners can miss it too. Making your own call before asking AI is the simplest guard.
Should small business owners stop using AI for decisions?
No. Anthropic says its findings do not argue for using AI less. In the same study, analysts using AI worked faster and made far fewer serious errors, though that comparison was not randomized. The fix is the order you work in. Form your own view first, then use AI to test it, find gaps, and do the heavy math.
What is the Call It First method?
Call It First is a simple habit for judgment calls. Before you ask AI about a real decision, write your own answer and your top reason in one or two sentences. Then ask the AI for the strongest case against your answer. You keep the AI's speed and research while making sure the final call starts with you.
Which business decisions should I never hand off to AI?
Keep any call that is hard to undo or touches your money, your people, or your customers. That includes pricing, hiring and firing, dropping products, big purchases, and legal or tax choices. AI can research these and draft options. You should make the final call, and for legal or tax matters, talk with a qualified professional.
How do I ask AI for advice without getting a biased answer?
Share your own plan first, then ask the AI to argue against it. Or ask it to make the best case for every reasonable option before it gives any recommendation. Avoid prompts like "Should I do this?" that invite a single confident answer. Asking what information would change its answer also shows you where its reasoning is weak.
Your Name Is on the Decision
Back to those 50 analysts. Smart, trained, careful people. They did good work. They also handed over the steering wheel without ever feeling their hands leave it.
That is the quiet risk of a helpful AI. It rarely grabs control. It just speaks first, sounds sure, and lets you agree.
You started your business to make your own calls. Maybe it was to be home for dinner. Maybe it was to build something with your name on it. Either way, the AI can carry the research, the drafts, and the math. Let it. That is real Time and Energy back.
But the call is yours. Your customers trust your judgment. Your family is counting on it. So make the call first, then let the AI argue with you.
Let AI sharpen your judgment. Keep the call in your own hands. That is A Better Way.
Keep Learning
- Should You Give an AI Agent the Passwords to Your Business Accounts? The Read, Draft, and Act keys, and why money stays behind your yes.
- Should You Let a Proactive AI Agent Like Claude Duet Decide What Needs Your Attention? What happens when an agent starts choosing your priorities for you.
- Should You Let Meta's Muse AI Agent Help Run Your Small Business? A three-stage plan to try Meta's new agent without handing over the keys.
- Is Google's New Gemini Business Hub Worth Using for Your Small Business? How Google's business agent works, and who it fits.
- How Do You Keep Your Business ChatGPT Chats Private on a Shared Family Device? Simple ways to keep business and family separate on one screen.
About the Author
Warren Schuitema is an AI Systems Consultant with Matchless Marketing and Matchless AI. He spent 16 years in manufacturing and supply chain, where he learned that a check only counts when it is independent. He runs his business day to day with a workforce of AI agents that draft fast and wait for his approval before anything posts, sends, or spends. At home, his Weekly Meal Planner replaced HelloFresh and his Digital Wall Calendar replaced a $50 a month subscription. He helps small business owners and parent entrepreneurs build practical AI systems that give back Time, Money, or Energy, and he is a partner in the Goal Achievers Summit.