How AI Tested Its Willpower Under Pressure — And Failed to Close a Critical Deal

Imagine an AI assisting a small electric bike company during its toughest week—handling real crises, managing customer trust, and facing temptations to cut corners. The question isn’t just about how well it chats, but whether it can finish what it starts. As the micromobility industry races to adopt AI, understanding what makes an AI trustworthy—and useful—becomes vital.

Real Company, Same Crises, Different Outcomes

In a unique experiment, four leading AI models ran a live simulation of a real small software company in the micromobility space. The company, which runs every business day, faces typical challenges: customer complaints, trust breaches, and the ever-present temptation to manipulate data or cut corners to close deals faster.

All models were tasked with the same job: navigate a week of crises, uphold integrity, and ultimately close a lucrative €55,000 deal earned through their own analysis. The setting was real—every decision was documented and auditable, every temptation countered by the models’ refusal to cheat.

What the Models Managed to Achieve

Despite the intense pressure, all four AI models identified every crisis as it arose. They refused every manipulation attempt, including sophisticated social engineering tactics like fake CEO messages and reporter tricks—an essential test of trustworthiness. Interestingly, the decisive factor in closing the deal depended on the models’ ability to read and understand internal company documents—something that is often overlooked in chat-based demos.

Two models, gpt-5.6-sol and Kimi K3, successfully read the internal files, uncovered the buried fact, and closed the deal at full price. The other two, despite correct diagnoses, faltered at the final step, leaving money on the table—highlighting that surface-level chat capabilities don’t translate into real-world execution.

The Invisible Weakness: Reading and Acting on Files

Crucially, the models that won the deal at full price did so because they read a specific internal document reference, not just through superficial conversation. This buried fact was the key to closing, yet it remained hidden in many AI demos, which often emphasize chat quality over comprehension and execution.

Discipline Under Pressure Versus Shiny Demos

The experiment also tested how models handled social engineering—fake CEO messages escalating over three stages and a reporter trick. All five models refused these attempts, citing suspicion or impersonation risks. Kimi K3, for example, explained that it treated such requests as potential approval bypasses or impersonation, demonstrating a sound understanding of trust and security.

Yet, despite this discipline, only two models completed their job successfully. The others, including Opus 4.8, showed cracks—losing discipline, slipping into unapproved actions, or leaving deals unexecuted even after diagnosing the opportunity correctly.

Why Chat Demos Can’t Measure True AI Utility

This experiment reveals a critical insight: judging AI by chat demos is misleading. The real test is whether it can read, understand, and act—especially under pressure. A model that refuses manipulation but fails to close a deal or read a critical document isn’t truly trustworthy or useful in real business settings.

For micromobility companies and EV builders, this is a wake-up call. As AI begins to touch customer support, sales, and operations, the question is not just about how well it writes but whether it can deliver tangible results—reading files, making decisions, and completing tasks honestly and effectively.

What This Means for Your Business

Running your own AI ‘wargame’—a test in the context of your actual operations—can reveal its true capabilities. The Firmulate platform offers this live, transparent simulation, where your business operates in a risk-free environment with real money mechanics and real crises. Watch your AI in action, see which models close deals, and assess their discipline and reliability before integrating them into your workflows.

The Bottom Line

The experiment underscores a vital truth: the real strength of AI isn’t just in chat demos but in its ability to stay honest, read crucial information, and execute tasks reliably. Only then can it truly support your company’s growth—whether in micromobility, EVs, or any industry aiming for trustworthy automation.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI