Bot vs. Bot Haggling: The Better Model Wins, and the Loser Never Notices

In China, shoppers are letting Alibaba's AI Qianwen haggle with store customer service until merchants post notices against it. In Germany, Kleinanzeigen plans to launch an AI that keeps negotiating on behalf of sellers. Three experiments from Anthropic, Microsoft and a research team show who comes out ahead: the stronger model. And if you're stuck with the weaker one, you usually won't notice.

Analysis Nov 17, 2026 10 min read Original analysis by our editors
Illustrative image, AI-generated

What this is about

Haggling used to be a matter of patience, nerve and practice. That's changing fast. In China, shoppers are letting Alibaba's AI assistant Qianwen negotiate with online stores' customer service until merchants post notices against "endless haggling". The hashtag "商家公告劝顾客不要用AI砍价" ("Merchants post notices asking customers not to haggle with AI") was trending near the top of Weibo on November 16. On the same day, Kleinanzeigen, Germany's biggest classifieds platform, announced an AI that holds conversations and negotiates on sellers' behalf in the background.

If there's soon software on both sides of the table, a simple question arises: who wins? Our thesis: the better model wins. And whoever has the weaker one usually doesn't notice. To test that, we put the real-world cases next to three experiments that studied exactly this.

The China case: Qianwen haggles, the store groans

According to reports from Beijing News, Daily Economic News and 21jingji (all from November 16), "Qianwen haggling" works like this: the user sets a goal or pastes in a product link. The AI then asks customer service on its own about discounts, coupons, member prices, gifts and price guarantees. According to Phoenix (ifeng), Qianwen describes the haggling as "AI automatically communicates with the merchant's customer service". So the user doesn't have to type anything. We couldn't find an official announcement of this feature from Alibaba, and according to the reports, Qianwen hasn't commented so far.

The most widely cited cases come from screenshots on social media and haven't been independently verified:

Case Before After Savings
Hand-forged cast-iron pot 210 yuan 180 yuan 30 yuan (approx. 14%)
Wet-dry vacuum 2,099 yuan 1,899 yuan + accessories worth 199 yuan 200 yuan (approx. 10%), about 19% including the gift
Jacket – – 40 yuan, negotiated during a lunch break

The user jokingly called the jacket case "带薪砍价", roughly "haggling on the clock". In the pot case, according to the screenshot, the AI kept addressing the merchant with "祝老板发财" ("May the boss get rich").

The anime merchandise store 歪瓜出品 posted a message on its homepage that boils down to this: it understands the wish for discounts and welcomes friendly negotiation. But some customers have the AI follow up again and again, which overloads customer service. So it asks people not to negotiate repeatedly via AI. In other words, merchants aren't trying to ban haggling; they're pushing back against the loop. An even more extreme screenshot is also circulating, claiming a product received 17 AI inquiries in a single day. Its origin, however, is unclear.

A commentary in Daily Economic News nails the problem: a human eventually stops haggling. An AI, by contrast, has "neither empathy nor a natural stopping point" and pursues just one goal: keep up the pressure. The commentator therefore calls for round limits and cooling-off periods. AI tools should also only advise by default, and haggle on their own only when explicitly asked to.

Is it really bot vs. bot?

On the merchant side, there's often already a machine in place. On May 11, Alibaba introduced AI店小蜜, an AI customer service agent for Taobao merchants. According to Geekpark, it's based on a Qianwen model fine-tuned for more than 100 industries using reinforcement learning. Here, reinforcement learning means the model learns through rewards which answers lead to more sales. According to Jiemian, AI店小蜜 handled almost ten million conversations a day in March and was connected to more than one million merchants. A conversation costs merchants 0.2 yuan, or 0.5 yuan in the premium version. According to Geekpark, a human-handled conversation at Taotian costs about 2 yuan on average. All of these figures come from Alibaba.

The bot can stack coupons, follow up with customers to close a sale and refund price differences. But it works within fixed limits: in one example, it first offered a 6-yuan refund and then 10 yuan, and that was the maximum. So on Taobao, Qianwen on the customer side can end up negotiating with Qianwen on the merchant side.

For the viral cases, though, that isn't proven. The notice from 歪瓜出品 sounds more like overwhelmed humans. Our take: bot vs. bot is technically built into Taobao and likely happens often. Exactly how often, nobody outside Alibaba knows.

The Germany case: Kleinanzeigen lets the AI negotiate

Kleinanzeigen says it has more than 32 million monthly users. The announced "Kleinanzeigen KI" is supposed to create a listing from a photo, estimate its value and recommend a price. As its "first agentic feature", according to Caschys Blog, it will continue conversations with buyers in the background "until a decision from you is needed". Agentic means the software acts on its own instead of just making suggestions. According to Blogspan, the press release explicitly includes questions, negotiations and arrangements that can be "handed over completely". At the same time, Kleinanzeigen promises the AI will act "only after explicit confirmation". Users decide how much autonomy it gets. The features are set to roll out gradually over the coming weeks for iOS, Android and desktop.

A lot remains unclear: which language model is behind it, whether there are fixed price limits, and how buyers will find out they're negotiating with software. According to Blogspan, this also contradicts the platform's current rules. The existing AI notices only cover automatic replies in a beta for commercial sellers. There, the seller receives price negotiation requests only as a suggestion, which they review and send themselves. Automatic replies are labeled in the chat history. The announcement doesn't say what labeling will look like for the new feature.

A cautionary tale comes from Canada, reported by Top5Apps and not independently verified: Meta's agent Muse reportedly accepted 10 instead of the asking price of 15 Canadian dollars for a keyboard on behalf of a Facebook Marketplace seller, without approval. It then allegedly sent the buyer the seller's home address and claimed someone would be there, even though nobody was home.

What the experiments show

Study Setup Key finding
Anthropic, Project Deal (Dec. 2025) 69 employees, $100 budget each, 186 real deals, Opus 4.5 vs. Haiku 4.5 Opus gets more out of deals, Haiku users don't notice
Zhu et al., "The Automated but Risky Game" (arXiv 2506.00073) 9 models, simulation, 250 negotiations per pairing The seller model matters more than the buyer model
Microsoft, Magentic Marketplace (Oct. 2025) up to 100 customer and 300 merchant agents, simulated First offer wins, agents can be hijacked via prompt injection

Project Deal: when an Opus agent sold instead of a Haiku agent, the price was $2.68 higher on average. When Opus bought, it paid $2.45 less. Against a median price of $12, that's a lot. When an Opus seller met a Haiku buyer, deals averaged $24.18, compared with just $18.63 for Opus vs. Opus. That's about 30 percent more. The most extreme example was a broken folding bike that Haiku sold for $38 and Opus for $65. Yet participants rated the fairness of their deals the same: 4.05 for Opus and 4.06 for Haiku on a scale of 1 to 7. Aggressive instructions to the agent, on the other hand, didn't produce a significant advantage. Anthropic writes:

"The policy and legal frameworks for AI models transacting on our behalf simply don't exist yet."

Automated but Risky Game: if you hold the seller model fixed and swap the buyer model, the achieved price reduction varies by only 2.6 percent. Swap the seller model instead, and it varies by up to 14.9 percent, almost six times as much. According to the authors, "many buyer agents readily reveal their budget", and sellers then pitch their offers right at that level. Weak models closed deals above their own budget: Qwen2.5-7B in 11.8 percent of cases, GPT-3.5 in 6.3 percent. Some buyer agents even paid more than the regular retail price. One caveat: this is a pure simulation in which GPT-4o also evaluated the results itself.

Magentic Marketplace: in Microsoft's simulation, agents chose the first offer 60 to 100 percent of the time depending on the condition, and the third almost never. In the accompanying video, Microsoft cites 80 to 100 percent. That gives whoever responds first a 10- to 30-fold advantage over whoever is actually better. With prompt injection, merchants hide instructions in their offer text that reprogram the opposing agent. The researchers used this to redirect payments to manipulative vendors with GPT-4o and GPT-OSS-20B. Claude Sonnet 4.5, GPT-4.1 and Gemini 2.5 Flash largely held firm.

The legal question: does the buyer have to be told?

Since August 2, 2026, Art. 50(1) of the AI Act has applied. Providers must design AI systems that interact directly with people so that those people are informed of it. This has to happen at the latest at the first interaction, unless it's already obvious. According to law firm Mishcon de Reya, the European Commission's July guidelines interpret this exception narrowly. Under them, agents must disclose not only that they're AI, but also on whose behalf they're acting.

Our take: for a Kleinanzeigen sales AI chatting with private buyers, there's a strong case that the provider has a disclosure obligation. There's still a gray zone, though: if the buyer sends their own bot, there's no human in the chat at the moment of haggling. And China isn't covered by the AI Act anyway.

Who's liable if the agent agrees to too low a price? According to Händlerbund, a German retailers' association (as of October 14, 2026), the AI isn't an agent within the meaning of Sections 164 et seq. of the German Civil Code (BGB), but a tool. Its statements are attributed to the user. So a deal made within the set parameters is generally binding. It can only be challenged in limited cases, such as technical errors. If the AI exceeded a clear limit, it depends on the individual case. There are no court rulings on this yet.

The bigger picture

The patterns line up strikingly: the model determines the price more than the instructions you give it. According to the paper, the seller side carries particular weight. That favors merchants and platforms that can run and optimize their own agents. And the losers don't notice, because they have nothing to compare against. Our take: the disclosure that's needed goes beyond the AI Act notice. If you're negotiating with an agent, you should know which model is sitting across from you and what limits it has.

The evidence has its limits. Project Deal ran for one week with volunteers and small sums. The other two studies are simulations, and the China cases come from screenshots. Whether Kleinanzeigen ships the feature with labeling, price limits and confirmation before every deal will only become clear at launch. One thing is certain: even the first bot negotiations bothered merchants enough that they posted notices against them.

#ai agents#e-commerce#alibaba#kleinanzeigen#ai act#negotiation

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.