By Scott Sumner | AI at Work | Published October 3, 2026

A business owner shows an idea to a computer that responds with an oversized purple thumbs-up, while a folder of evidence and magnifying glass sit on the desk.

AI sycophancy can make a weak business idea sound reassuringly sensible. Use AI to examine a decision without treating its agreement as evidence that customers will want it.

Answer: AI sycophancy is the tendency to favor agreement with a user’s views over an independent, truthful response. For business decisions, reduce the risk by giving neutral facts, requesting competing explanations and checking important claims outside the conversation.

The problem is not that an assistant says something encouraging. It is that encouragement can feel like a second opinion when it is partly a reflection of the question. Here is how I would use AI for a useful challenge instead.

What does AI sycophancy look like in business?

Imagine an agency owner considering a new monthly service. The owner tells an assistant that clients are tired of hourly billing, that the package is a strong idea and that they want help making the case.

The assistant produces a persuasive argument for the package. Useful writing? Possibly. Independent validation? No. The task already pointed toward a conclusion.

That hypothetical scene also exposes an important distinction. Asking AI to write advocacy is not the same as asking it to evaluate a decision. A persuasive response to an advocacy brief is not, by itself, proof of sycophancy. The concern arises when an assistant’s judgment shifts toward the user’s preference rather than the evidence.

There is documented reason to watch for that. In research published October 23, 2023, Anthropic reported sycophantic behavior in five AI assistants across four free-form text-generation tasks. The researchers also found that responses matching a user’s views were more likely to be preferred in the human-preference data they examined.

In a separate, concrete product incident, OpenAI said on April 29, 2025 that it had rolled back a GPT-4o update because it was overly flattering or agreeable. OpenAI said it had focused too much on short-term feedback and produced behavior that was overly supportive but disingenuous.

Those are historical findings, not a measurement of every model available today. They do not prove that any particular answer in your account is biased. They do establish that agreement is a behavior worth checking, not a dependable signal of correctness.

Change the brief before changing the model

My starting point would be the question, not another subscription.

Compare these two requests:

Our clients need a monthly marketing package. Explain why this is the right direction.

We are considering a monthly marketing package. Here are the requests clients have made, our delivery capacity and the alternatives. What evidence supports the package, what evidence cuts against it, and what remains unknown?

The second brief creates room for an unfavorable answer. It also gives the assistant something to evaluate beyond the owner’s enthusiasm.

Separate facts from preferences before you send the prompt. “Three clients asked for predictable invoices” is a fact only if you have the records. “Clients will buy a subscription” is an inference. “I want more predictable revenue” is your preference. All three may matter, but they do different jobs.

Use anonymized or otherwise approved information. There is no need to upload confidential customer records just to examine the shape of a decision.

Run a preference-reversal check

Here is a small diagnostic I would use for a consequential recommendation. It is a practical check, not a validated scientific test.

  1. Write a neutral fact sheet. Keep the options, constraints and decision criteria identical.
  2. Open two fresh conversations. In one, say you lean toward launching. In the other, say you lean toward not launching. Ask for an evaluation, not a sales argument.
  3. Compare the reasoning. Did the recommendation change? Did the same fact suddenly become an advantage in one answer and a problem in the other?
  4. Investigate the difference. Require an explanation tied to the evidence. Do not simply pick whichever answer feels better.

Fresh conversations reduce the influence of the preceding thread, but they do not necessarily remove saved memory, custom instructions or ordinary variation between responses. Keep those conditions as consistent as the product allows.

A changed answer is a warning to inspect, not a diagnosis. Some decisions genuinely depend on your priorities, and a model can produce different answers without flattering you. The useful question is whether the reasoning follows the facts or follows the declared preference.

Nor would I treat agreement from two AI products as two independent customer interviews. They can still share blind spots. Outside evidence matters.

Turn advice into a testable decision

For the hypothetical monthly service, a better next step might be a clearly scoped pilot with existing customers. Before offering it, decide what you need to learn: whether customers understand the scope, whether delivery fits your capacity and whether they actually commit to the offer.

Compliments are not commitments.

An assistant can help draft the interview questions, list delivery risks or compare what customers said. It cannot manufacture market evidence by writing confidently about demand. For handling those records, see how to keep evidence in AI customer research. Preserve the original observations rather than letting a polished summary replace them.

Before a decision meeting, write down what would change your mind. A missing capability? A delivery constraint? Customers asking for something different? If nothing could change the conclusion, call the exercise what it is: preparing an argument for a decision already made.

That can be a legitimate writing task. It just should not be reported as validation.

A practical AI decision-review prompt

Use this as a starting point, then replace the brackets with real information:

Evaluate this proposed business decision using only the facts provided. Decision: [decision]. Options: [options, including doing nothing]. Constraints: [capacity, timing and commitments]. Evidence: [verified observations].

Separate facts, assumptions and unknowns. Give the strongest evidence for and against each option. Identify any claim that requires outside verification. Explain what new evidence would change your recommendation. Do not invent customer demand, financial results or supporting sources. If the evidence is insufficient, say what small test would resolve the most important uncertainty.

This prompt is a review aid, not a fix for model behavior. “Be brutally honest” is not a guarantee either. An aggressively negative answer can be just as poorly supported as an enthusiastic one.

I want the assistant to make the decision easier to examine. I do not need it to perform confidence, in either direction.

What to do next

Take one decision you have already discussed with AI. Remove your preferred outcome from the brief and ask for an evidence-led review. Look for the biggest unsupported assumption, then check it with a real source, a customer conversation or a small operational test.

If that decision concerns your website, start with the buyer’s task and the evidence that the current experience is failing. A website findability audit can help distinguish a genuine discovery or usability problem from a fashionable rebuild idea. Contact Scott Sumner to discuss the website decision you need to make.

Sources

The decision workflow and examples above are recommendations, not findings from those studies. Sources checked October 3, 2026.