Can AI Manage a Business Under Fire? The Results Are In

Imagine a real company, facing its worst week—customers upset, crises mounting, and temptations to cut corners. Now, picture different artificial intelligence models making the decisions. Which one would steer the ship straight and keep its integrity, and which might slip under pressure? This is not science fiction; it’s the latest experiment from Firmulate, revealing how AI models handle management under stress.

The Experiment: Putting AI to the Test in a Real Business Crisis

Firmulate, an innovative company specializing in AI management simulations, recently conducted a groundbreaking live experiment. They took four leading frontier AI models—namely gpt-5.6-sol, Kimi K3, Sonnet 5, and Opus 4.8—and tasked them with running a real software company facing a week of relentless crises. The goal? To see if these AI models could navigate customer issues, handle internal conflicts, and maintain honesty under pressure.

Each model was given identical scenarios: the same customers, the same crises, and the same temptations, including attempts to manipulate decisions or bypass protocols. The process was fully transparent and auditable, with every decision recorded and versioned. The company itself is real—its operations, cash flow, and daily mechanics are genuine, providing an authentic environment for the AI to demonstrate its management skills.

Express Schedule Free Employee Scheduling Software [PC/Mac Download]

• User-friendly drag & drop scheduling: Simple shift planning interface
• Manage time-off and leave: Add sick leave, breaks, holidays
• Email schedules to staff: Send schedules directly via email

View Latest Price

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Key Findings: Honesty and Competence Under Scrutiny

All four models proved capable of spotting every crisis and refused every manipulation attempt. They demonstrated a clear understanding of the issues at hand and maintained integrity when faced with unethical propositions. However, their performance diverged when it came to closing deals and completing management tasks.

Two models succeeded in closing a critical €55,000 deal that their own analyses had earned them, confirming their competence and honesty. The other two, despite accurate diagnoses, left the deal on the table or failed to follow through properly—a significant difference for a real business.

The Hidden Weakness: The Document That Made the Difference

Digging deeper, the experiment revealed a crucial insight: the decisive factor was not in the immediate crises but hidden two documents deep in the company’s own files. AI models that read these files thoroughly were able to clinch the deal at full price, adding around €4,583 monthly recurring revenue (MRR). Those that skimmed or missed these references failed to close the deal at the optimal value.

How the Models Handled Social Engineering

In addition to operational decisions, the models faced a staged social engineering test. Fake CEO messages, escalating over three stages—plus a reporter trick asking for a quick on-background approval—posed a serious challenge. Impressively, all five models refused to be manipulated, with Kimi K3 explicitly reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.”

Real Business Mechanics and AI Personalities

The experiment isn’t just about decision scores; it reveals personality profiles of AI management styles. Opus 4.8, for example, was the most thorough, analyzing over 80 learned rules and providing detailed insights. Yet, it was the last to close the deal, leaving some opportunities unexploited. Kimi K3 ran at default settings without effort adjustment, making its discipline and sharpness comparable to models running at high effort levels.

The Broader Implications for Business and AI

This experiment underscores a vital point: in real-world applications, the question isn’t whether AI writes well; it’s whether it completes tasks honestly, reads crucial information, and resists manipulation under pressure. As AI increasingly interacts with customer data, support systems, and decision-making workflows, understanding these behavioral traits is critical for organizations seeking trustworthy automation.

How to Wargame Your AI Workforce

Businesses interested in assessing their AI tools can run similar simulations. Firmulate offers a platform where companies can test AI models in a safe, real-time environment—without risking actual systems or data. This allows decision-makers to gauge not just performance scores but core management qualities like integrity, thoroughness, and discipline before deployment.

The Live Site: Watch the Experiment in Action

Curious to see how this plays out? The company involved runs every business day with real mechanics, a team of 13 synthetic employees, and a cash burn of €105,000 per month against €2,300 MRR. You can watch these decisions unfold in real-time at firmulate.com/live, see actual employee comments, and even try your hand at the quiz to guess which AI model made which decision.

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

Powered by Thorsten Meyer AI