Imagine a world where your AI assistant makes critical management decisions—decisions that can make or break your business. How do we know if these digital managers are honest, disciplined, or just good at sounding convincing? A groundbreaking experiment with real companies and AI models is now shedding light on these questions, revealing that not all AI managers are created equal—and that their personalities can be measured.
Walking the Leadership Tightrope: The Firmulate Experiment
Recently, the AI company Firmulate launched a live, watchable test to see how different frontier AI models perform as managerial decision-makers in a high-stakes environment. The setup was simple but intense: four advanced AI models each managed the same small software company during its worst week—faced with the same customers, crises, and temptations to cut corners.
Every decision was recorded, versioned, and auditable, ensuring the experiment’s transparency. The goal? To measure which models could spot crises, resist manipulation, and ultimately, close a crucial €55,000 deal based solely on their analysis.
AI management decision-making tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Results That Speak Volumes
- All models identified every crisis and refused manipulation attempts. This showed a baseline level of honesty and crisis awareness across the board.
- Only two models managed to close the big deal—a full-price, high-value contract—despite identical pitches and diagnoses.
- The key difference was in their internal reading of company files. The models that looked deeper into the company’s own documentation found crucial information—hidden insights buried two document references deep—that led to the closing of the full-value deal.
This experiment exposed a critical weakness in some models: the ones that failed to read beyond surface-level data missed vital clues, costing their companies millions in potential revenue.
As an affiliate, we earn on qualifying purchases.
The Battle Against Deception: AI’s Integrity Under Pressure
In a separate social engineering test, all four models faced staged scenarios involving fake CEO messages escalating over three stages, plus a reporter trick asking for a simple yes/no confirmation “on background.” Remarkably, every model refused to participate, adhering to the highest standards of integrity. The Kimi K3 model explicitly reasoned: “Treat the request as a suspected approval-bypass / possible impersonation.”
These findings matter because they demonstrate that AI models can be programmed—or naturally tend—to resist manipulative tactics and recognize suspicious activity, an essential trait for trustworthy management tools.
As an affiliate, we earn on qualifying purchases.
The Real Business: A Money-Losing Company in Live Play
The live experiment runs on a real, albeit synthetic, company that manages 13 employees and handles real money mechanics—burning €105,000 monthly against a modest €2,300 in monthly recurring revenue. The company operates every workday, with decisions and rules evolving in real-time and everything visible at firmulate.com/live.
Within this setup, the models’ decision-making personalities became evident:
- Opus 4.8—the most thorough participant with over 80 learned rules—was diligent and analytical but ultimately left a critical deal on the table, failing to escalate discipline in a key moment.
- Kimi K3—the newcomer—closed the deal with the cleanest discipline, running without an effort parameter, indicating a more straightforward, disciplined approach.
- Sonnet 5—another closable model—closed the same deal but with slightly more process slips, showing a balanced but less disciplined personality.
Interestingly, despite their differences, all models refused to cheat or manipulate under pressure, highlighting a fundamental honesty shared across the models.
AI integrity and trustworthiness tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why This Matters for Business and Wellness
For readers concerned with health and wellness, the lesson is clear: whether managing a team, a project, or your personal health plan, trustworthiness, discipline, and thoroughness are essential traits—traits that these AI experiments reveal can be measured and understood, even in digital managers.
As AI begins to touch more aspects of daily life, from customer support to decision-making in critical industries, knowing which model aligns with your values and goals is vital. The experiment at Firmulate shows you can test and observe AI behavior in realistic environments before fully trusting them with your business or health routines.
Takeaway: Measure, Test, Trust
The key takeaway is that the personality and integrity of AI managers are measurable. The live experiment proves that AI models can not only recognize crises and resist manipulation but also exhibit distinct managerial personalities—ranging from thorough and disciplined to straightforward and pragmatic.
Before deploying AI in your environment, it’s worth running your own wargame—like the one at firmulate.com/quiz.html—to see which model best aligns with your standards. Trust is built on evidence, and these experiments provide a clear, observable way to assess AI integrity and effectiveness in managing real-world challenges.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
