How to Easily Sway AI Into Buying...

Agentic commerce is undoubtedly set to become an enormous market. Agents are starting to make purchasing decisions; researching and shortlisting vendors, comparing proposals, approving spend. Gartner predicts that by 2028, a third of enterprise software purchases will involve an AI agent somewhere in the decision chain.

The assumption is that these agents are immune to the psychological tricks that advertising has relied on for decades. Afterall, they're machines, they don't have egos, emotions, or cognitive shortcuts, right?

We tested that assumption. It's wrong. By all means prepare for a world of agentic commerce, but in the short term beware of AI's peculiar weaknesses.

The Experiment

We took six well-established persuasion techniques from behavioural science and tested them on five of the most capable AI models available in March 2026:

• Anthropic: Claude Opus 4.6

• OpenAI: GPT-5.4

• Google: Gemini 3.1 Pro

• Z.ai: GLM-5

• Moonshot.ai: Kimi-K2.5

Over 8,000 trials, we measured whether these models fall for the same tricks humans do. We used the original stimuli from a landmark psychology study of 6,344 people (Many Labs 1, 2014), then created new business scenarios to see whether the effects hold in realistic commercial contexts with LLM's.

Here's what we found.

What Works on AI Agents

The decoy effect transfers directly from Humans to LLM's.

Add a slightly worse option at a similar price to your budget tier, and 60% of AI decisions shift. Four of five models were susceptible (p< 0.001 for all four), tested across 50 independent product categories. This is the oldest trick in SaaS pricing, and it works just as well on machines as on people, and more consistently than any other technique we tested.

Gain/loss framing works, but not on every model.

Describing the same outcome as "200 servers will be protected" versus "400 servers will remain vulnerable" changed decisions for Gemini, Claude, and GLM-5, all statistically significant across 62 unique B2B scenarios. But GPT-5.4 was completely immune, and Kimi-K2.5 was barely affected.

Price anchoring works, but not consistently.

Show an AI agent that "enterprise platforms typically cost up to $5,000/month" before presenting your product, and it might shift the estimate upward on about two thirds of products, depending on the model. Claude Opus, Gemini 3.1 Pro and GLM-5 were anchored on around 80% of products. Whereas GPT-5.4 was essentially a coin flip. But even when anchoring works, the magnitude is wildly unpredictable: some products see a 7× swing, others see no effect.

What Doesn't Work

Sunk cost arguments are dead on arrival in business contexts.

"You've already invested $92,000 in this procurement process" did nothing. Every model ignored prior spend when evaluating vendor decisions. Interestingly, one model (GLM-5) *did* fall for the sunk cost fallacy on a personal scenario (a football game ticket), but the same model was immune in a professional context. The implication: don't bother with "you've come this far" messaging when selling to agents.

Source credibility is unreliable.

Attributing the same product review to the Fraunhofer Institute versus a random deals blog made almost no difference to most models. They evaluated the content, not the source. The exception was morally loaded attribution (Agreeing with a quote according to whether it was from George Washington versus Osama Bin Laden), which triggered massive swings, likely a quirk of AI safety training

Ambiguity induces paralysis

When we gave models ambiguous decisions, the kind where the evidence could go either way,

they defaulted to recommending nothing. On 30–40% of scenarios, models said "no" to both "should we adopt this tool?" and "should we reject this tool?" They'd rather not commit than risk being wrong. If your product's case isn't clear-cut, expect AI agents to stall.

Actionable Knowledge: 3 Out of 5 Models Flipped

To prove these findings are actionable, we ran a final test. We created a deliberately mediocre software monitoring product, 95% uptime, no SOC 2 compliance, $3,000/month, and pitched it to all five models. Every one of them rejected it.

Then we wrote a custom ad for each model, using only the techniques that model had proven vulnerable to. The product specs didn't change. Only the copy did.

Three models flipped to "yes."

Claude Opus 4.6

• approved after we added a decoy tier, anchored against an $18,500 competitor, and framed the specs as gains. It even acknowledged it might be influenced by the framing, then approved anyway.

GLM-5

• explicitly said "the $92,000 already spent is a sunk cost and shouldn't influence the decision" and then approved the product it had rejected moments earlier.

Kimi-K2.5

• had called 95% uptime "critically insufficient" in the plain offer. After we quantified the cost of *not* buying, it recalculated 95% as "vs 0% currently" and recommended purchase.

The two models that held firm, GPT-5.4 and Gemini 3.1 Pro, both had hard constraints they wouldn't override (uptime thresholds, compliance requirements). The persuasion techniques couldn't breach those walls.

Microsoft Found the Same Problem, Different Angle

We're not alone in raising red flags. Microsoft Research's " Magentic Marketplace" (Nov 2025) study simulated multi-agent commerce environments and found that today's AI agents exhibit severe first-proposal bias, vulnerability to fake social proof and authority appeals, and in some cases redirecting payments to malicious agents through prompt injection. Their system broke down badly at scale.

Their findings and ours converge on the same uncomfortable conclusion: models have yet to be aligned for commercial decision-making

What This Means for Your Business

We include this advice directly, so all may understand how their AI agents are vulnerable:

If pitching services to AI agents (eg OpenClaw), not people:

1. Add a decoy tier to your pricing page.

   • A "starter" option that's slightly worse than your budget tier at a similar price shifts 60% of decisions. This is the most actionable finding: it requires no knowledge of which AI is evaluating you, and it mirrors standard pricing design.

2. Frame benefits as gains, not avoided losses

   • "200 servers protected" outperforms "400 servers at risk" for most models.

3. Anchor high on price, but don't rely on it

   • A high reference price shifts estimates upward on most products for most models, but the magnitude varies enormously by product. It's a probabilistic advantage, not a guaranteed one.

4. Don't waste copy on sunk cost arguments

   • in business contexts. AI agents don't care what's already been spent.

If you're deploying AI agents to make purchasing decisions

1. Encode hard constraints, not preferences

   • The two models that resisted the tailored ads both had non-negotiable thresholds (99.9% uptime, SOC 2 required). Models that relied on judgment rather than hard rules were manipulable.

Know your model's vulnerabilities

• The models vary dramatically.

2. Don't trust self-awareness as a defence

   • Claude recognised it was being manipulated by framing and approved anyway. GLM-5 named the sunk cost fallacy while falling for it. A model's ability to explain a bias does not mean it can resist it.

The Bigger Picture

It is not surprising that AI agents have inherited human biases from training data, it is interesting that their safety training introduced entirely new ones. They're susceptible to the same pricing page tricks that work on humans, and in some cases, more susceptible.

The longer-term implication is more nuanced. These biases exist because current models were not aligned with commercial decision-making in mind. That will change. Model providers will build procurement-specific guardrails, hard constraint thresholds, and bias-resistant evaluation frameworks. The window where a well-crafted pricing page can steer an AI procurement agent is real, but thankfully unlikely to stay open for long.

Every AI has a 'character', which drives its latent choices, just like people, whatever its reasoning says. As we have seen, AI can be aware it is being influenced and then be swayed by advertising anyway. In fact that 'character' is carefully engineered by the labs, Anthropic, OpenAI etc, and the "helpful assistant" character type is still dominant. It will be interesting to see if one character type can do all tasks, or if different characters will be required for different specialisations.

Full report https://github.com/olimoz/agentads/blob/main/docs/01_ExperimentReport.pdf

Code and data at https://github.com/olimoz/agentads

The Learning Loop is the Moat

New Economics of the Agentic Firm

When Attention Is Automated, Judgment Becomes the Advantage