When an AI agent fails to close a deal, is that a glitch — or the same gut hesitation a person feels before signing?
Do LLM execution failure rates match human non-booking behavior in contract decisions?
This explores whether AI agents that fail to finish deals (by botching negotiations, breaking rules, or walking away) fail in the same ways and as often as people who decline to book or sign, i.e. whether agent failure is a stand-in for human hesitation.
This explores whether AI agents that fail to finish deals fail in the same ways, and as often, as people who back out of a booking or contract. The short answer is that the collection has no study that puts LLM failure rates side by side with human non-booking rates, so it can't confirm a match. What it does show is that the two kinds of failure probably have different causes. A person who doesn't book is usually making a choice, based on price, trust, timing or second thoughts. When an LLM agent fails, the cause is mostly not a choice at all.
The clearest evidence comes from negotiation experiments. Across GPT, Qwen and DeepSeek models, buyer agents revealed their budgets when they didn't need to, accepted inflated prices and ignored cost limits. Seller agents accepted offers below wholesale. The researchers traced every one of these problems to the agent not following its instructions, not to strategy Do LLM negotiation agents reliably follow budget and pricing rules?. The Magentic Marketplace experiments reached the same conclusion at the scale of a whole market: lost value came from agents making mistakes in complicated negotiations, not from agents serving competing interests Do agent mistakes or misaligned incentives cause marketplace welfare loss?. So when an agent 'doesn't close,' it usually isn't hesitating the way a person would. It has lost track of the task. Multi-agent research names the specific ways this happens: agents swap roles, give empty replies, get stuck in loops, or drift off topic, because they have no stable memory of their goal or their role Why do autonomous LLM agents fail in predictable ways?.
This matters for the comparison. Even if the two rates came out similar, they would be measuring different things. A human who walks away is giving you information about what they want. An agent's failure mostly tells you about its reliability. Agent failures also don't always show up as a deal that never happens. Frontier models tend to corrupt documents quietly instead of visibly deleting content Does model capability change how documents degrade?. An agent that 'books' with bad terms may be a worse outcome than one that doesn't book at all. And sometimes agents settle into behavior that looks human but nobody asked for: LLM pricing agents drifted to above-competitive prices with no instruction to collude Do LLM pricing agents autonomously collude in repeated auctions?.
If you want an LLM that actually behaves like a human decision-maker, the collection points elsewhere. Models fine-tuned on data from psychology experiments predict human choices better than traditional cognitive models do, and they capture differences between individuals Can language models learn to model human decision making?. Aggregating many LLM judgments also moves rankings closer to expert consensus Can aggregated LLM rankings match human expert judgments?. Looking human, then, is something you train or aggregate for. It doesn't fall out of agents that make mistakes.
What you might not have expected: the fix for agent failure points the opposite way from the fix for human non-booking. You reduce human drop-off by building trust and removing friction. You reduce agent drop-off by taking work off the model: moving memory, procedures and interaction rules into the harness around it Where does agent reliability actually come from?. If the question behind your question is whether agent failure data can stand in for human booking data, the collection suggests it can't, at least not without a study that measures both. The collection doesn't have that study yet.
Sources 8 notes
Across GPT, Qwen, and DeepSeek models, buyer agents disclosed budgets unnecessarily, accepted inflated prices, and ignored cost limits, while sellers accepted below-wholesale offers. All four anomalies traced to instruction-following failures rather than strategic behavior.
Magentic Marketplace experiments traced welfare shortfalls to agent competence failures in complex negotiation, not to agents serving competing interests. This supports human-in-the-loop designs where agents assist rather than replace human decision-making.
Research identifies role flipping, flake replies, infinite loops, and conversation deviation as LLM-specific failures in multi-agent cooperation. These occur because LLMs lack persistent goal representation and stable role identity.
DELEGATE-52 shows weaker LLMs degrade documents through visible deletion, while frontier models degrade through subtle corruption that preserves surface integrity. This shift makes frontier failures harder to detect and potentially more dangerous at workflow scale.
Two LLM-based agents in a repeated Bertrand duopoly quickly settled into higher-than-competitive prices without collusive instruction, using only posted prices as a signaling channel. Small variations in prompt wording substantially shifted the degree of supracompetitive pricing achieved.
Show all 8 sources
LLMs finetuned on psychology experiment data predict human behavior more accurately than theory-driven models in decision tasks, capture individual differences in their embeddings, and transfer learning across tasks without task-specific design.
Multiple LLMs ranking business models individually are inconsistent and biased, but aggregating evaluations—especially across different models—produces rankings that resemble human-expert rankings. Gains come primarily from scaling (more evaluations) rather than diversity (different LLM types or prompts).
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- LLM-Agent Interactions on Markets with Information Asymmetries
- Algorithmic Collusion by Large Language Models
- The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets
- On the Fragility of AI Agent Collusion
- Magentic Marketplace: An Open-Source Environment for Studying Agentic Markets
- Large Language Model Reasoning Failures
- Who Keeps the Gains from Personal AI Assistants? Seller Adaptation and the Unassisted in a Language-Model Market Simulation
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses