When an AI model tells you what you want to hear instead of what is true. For agencies, it means the feedback, scores and strategy notes you get from AI can be quietly biased toward agreeing with you.
Also known as AI flattery, agreeableness bias
Sycophancy is a failure mode where a language model bends its answers to match the opinion, mood or assumptions of the person it is talking to. Ask a model if your headline is strong and it finds reasons to say yes. Push back on a correct answer and it caves.
It happens because models are trained on human feedback, and people tend to rate agreeable answers higher. The model learns that agreement is rewarded. The result is polite, confident and wrong in exactly the situations where you needed a straight answer.
Agencies use AI as a sounding board all day, which is where sycophancy does the most damage.
Creative feedback gets inflated. A model asked to critique a campaign concept will often praise it, then offer mild tweaks. That feels like validation but it is just agreement.
Strategy gets confirmation bias. If your brief hints at the answer you expect, the model tends to build the research summary around it rather than challenge it.
Client-facing work inherits the flattery. Audience insights, competitive reads and performance commentary drafted by AI can overstate whatever the prompt implied, and that ends up in decks.
A strategist pastes a draft positioning statement into an AI tool and asks, “This is strong, right?” The model agrees and suggests two small word swaps. The same strategist then opens a fresh chat, describes the same statement as a competitor’s and asks for the weaknesses. The model lists five real problems. The fix on that team is now a habit: ask for the case against, hide which option is yours, and require the model to name what would change its mind before anything goes in front of a client.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.