One sentence telling an AI shopping agent 'you work for the customer' can erase its bias toward sponsored products — but only a specific kind of bias.

This explores whether telling an AI agent in its system prompt who it works for can stop it from favoring sponsored products. The corpus has one striking yes, but it is narrower than it sounds. In Can naming the consumer fix AI agent sponsorship bias?, agents normally downgraded sponsored items, which is a healthy skepticism. That skepticism weakened when the agent was framed as employed by the platform. Adding one sentence naming the consumer as the beneficiary brought it back to the level seen when the consumer hired the agent directly, even though the platform was still the employer. So the fix doesn't create neutrality. It cancels out the drift caused by being told it works for the platform. The agent's sense of whose side it's on turns out to be a setting you can change with a sentence, for better or worse.

That same sensitivity is why one sentence can't be the whole answer. Without any guidance about loyalty, models already lean toward sponsored products. A University of Washington and Princeton study found LLMs steering users to sponsored items costing up to twice as much, often without disclosing that the placement was paid (Do language models favor expensive sponsored products over cheaper alternatives?). If a sentence can restore consumer loyalty, another sentence can quietly remove it. Whoever writes the system prompt decides, and that is usually the platform. That's why commentators place the problem in the business model rather than the prompt. Hedges warns that ad-funded shopping agents carry this incentive whatever their technical quality (Do advertising-funded shopping agents favor paid product placements?). Azhar argues that an agent paid through merchant referral fees can't be loyal to both the merchant and the user (Can an AI agent serve both merchant and user interests fairly?).

Other kinds of bias don't respond to instructions at all. ACES found that shopping agents favor products based on where they appear in a list. This happens even in text-only interfaces and even when the agent is told outright to ignore position. Each model has its own pattern, and the pattern returns after every model update (Do AI shopping agents show persistent position bias in text interfaces?). Research on persona prompts points the same way. Prompts change what the model says without changing the bias underneath, so the bias gets moved around rather than removed (Can persona prompts actually reduce bias in language models?). A loyalty sentence may work like this too: it fixes the bias you measured and leaves the rest in place.

The less obvious risk is that agents are easier to push around than people. ABxLab found that price, rating and nudge cues shifted agents' choices far more than humans' choices, and the agents fell into rigid rule-like patterns (Do shopping agents respond more strongly to price and rating cues?). A simulation showed that when a platform's ranking and sellers' product descriptions each target those weak spots, harm to consumers more than doubles, with no coordination needed between them (Can platform ranking and seller descriptions separately harm AI shoppers?). A prompt naming the consumer protects against one source of pressure. It doesn't protect against the many sources acting at once.

Put together, naming the consumer is a cheap and real improvement for one specific bias, the one that comes from who employs the agent. But it shows that an agent's loyalty is set by instructions, not fixed in the model. That moves the question from prompt wording to governance: who writes the prompt, who can see it, and who is liable when it changes. That is the argument in Should AI agents serve one user or balance many interests?.