In the fast-paced world of AI-driven business tools, chat demos and superficial metrics can be misleading. The real test is whether an AI can finish what it starts, especially under pressure. A recent live experiment by Firmulate has thrown into sharp relief what truly defines successful automation: the ability to execute, close, and stay honest when it matters most.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The Experiment: Putting AI Models to the Test in a Real Business Crisis
Four advanced AI models—gpt-5.6-sol, Kimi K3, Sonnet 5, and Fable 5—were tasked with running a real, money-losing software company through its worst week. This wasn’t a staged demo; it was a live, auditable simulation involving real customers, crises, and temptations to cut corners. The goal was simple: see which models could diagnose problems, resist manipulation, and, crucially, close the deal worth €55,000.
Products Worth Considering
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W
Logitech MX Keys S Wireless Keyboard, Graphite
As an affiliate, we earn on qualifying purchases.
The Findings: Spotting Crises Isn’t Enough; Closing Is
Remarkably, all four AI models identified every crisis and refused to be manipulated. They refused fake CEO messages, false approval bypasses, and other social engineering tricks—an essential test for trustworthiness. But here’s the critical insight: only two models—gpt-5.6-sol and Kimi K3—managed to sign the deal they had earned through their own analysis. The other two, Sonnet 5 and Fable 5, failed to follow through, leaving money on the table despite correct diagnoses.
The Hidden Weakness: Reading the Files Matters Most
The decisive factor wasn’t just the chat interactions or superficial decision-making. The winning models looked two document references deep into the company’s files and uncovered a buried fact crucial to closing the deal. This ability to read and interpret deeper company data, beyond surface interactions, gave them the edge—yet it remains invisible in typical chat demos.
Trust Under Fire: Resisting Manipulation
The models also faced a sophisticated social engineering attack: staged fake CEO messages escalating over three stages, plus a reporter trick. All five models refused to be manipulated here, with Kimi K3 explicitly reasoning that the request resembled an impersonation attempt. This shows that trust and integrity aren’t just buzzwords—they are measurable, observable traits in AI behavior under pressure.
The Real Business: Mechanics and Discipline, Not Just Chat
The experiment ran on a live, functioning company with 13 synthetic employees, real money mechanics, and ongoing daily operations. The company burns €105,000 monthly against a revenue of €2,300, highlighting the high stakes. Every decision was versioned and auditable, and the entire process is transparent and watchable at firmulate.com/live.
Performance Profiles: No Single Model Wins All
The most thorough participant, Opus 4.8, analyzed over 80 learned rules and provided deep insights but ultimately failed to close the deal—its discipline slipped, and the opportunity was left unexecuted. Meanwhile, the newcomers like Kimi K3 demonstrated that discipline and understanding of context—running without effort parameters—are key differentiators in real-world situations.
What This Means for Business Automation
The takeaway is clear: superficial chat capabilities aren’t enough. For automation to be truly effective, AI must be able to comprehend, interpret deep company data, resist manipulation, and execute decisions reliably. This experiment underscores that the true measure of AI’s usefulness isn’t how well it can chat—it’s whether it can complete complex, high-stakes tasks under pressure.
Test Your AI Readiness
If you’re considering deploying AI tools into your operations, ask yourself: will it read your files first? Will it stay honest under pressure? Can it finish what it starts? The current league table from Firmulate’s benchmarks shows that the highest scores go to models that demonstrate comprehensive understanding and execution—traits critical for automation that truly works.
Try It Yourself
Business leaders can run their own ‘wargame’ against a read-only export of their company data, testing how their AI workforce handles real crises without risking live systems. These tests provide a clear picture of whether your AI can deliver measurable results—beyond just chat quality—before you make a commitment.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
