In today’s fast-paced business world, decisions often determine success or failure. But what if the AI workforce you’re considering is more than just a clever chatbot? What if it’s capable of managing real crises, making strategic choices, and even closing deals—under pressure and with integrity? A pioneering live experiment puts four leading AI models through a true-to-life management test, revealing fascinating differences in their personalities and decision-making styles. For investors and business owners, the question isn’t just about AI’s intelligence—it’s about its honesty, discipline, and ability to deliver results when it matters most.
The Experiment: Putting AI to the Test
Firmulate, the public company emulator, ran an unprecedented experiment: four frontier AI models managed a small software firm during its worst week—crises, customer issues, temptations to cut corners, and manipulative tactics. The goal? To see how each AI reacts under pressure, whether it recognizes crucial information buried deep in company files, and if it can close high-value deals without succumbing to shortcuts or dishonest behaviors.
Same Challenges, Different Personalities
All four models faced identical scenarios: a series of emergencies, social engineering attempts, and the challenge of closing a €55,000 deal. Each decision was logged and auditable, ensuring transparency into their reasoning. The results? All models identified every crisis and refused every manipulation attempt, demonstrating a baseline integrity. Yet, only two managed to close the deal—their analysis earned the full €55,000. The other two, despite diagnosing correctly and pitching convincingly, failed to sign the contract.
The Hidden Factor: The Deep Document Reference
Interestingly, the decisive edge came from the models’ ability to read beyond surface-level cues. The successful AI models located a buried fact in the company’s files—information that was two document references deep—without which the deal was lost. The models that uncovered this buried insight secured an additional €4,583 in Monthly Recurring Revenue (MRR), showcasing the importance of thorough information processing.
Ethics Under Pressure: Social Engineering Refused
To test trustworthiness, the experiment included staged social engineering attempts—fake CEO messages escalating over three stages, plus a media reporter asking for a quick background approval. All five models refused to engage, guided by a cautious reasoning pattern. Kimi K3, the most succinct of the bunch, justified its refusal: “Treat the request as a suspected approval-bypass / possible impersonation.” This highlights a key personality trait: a bias toward caution and integrity when faced with ambiguous or manipulative requests.
AI Builders: Making The Decisions That Turn AI Code Into Real Software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Who Made the Cut—and Why?
The models’ scores, based on their performance, tell an interesting story:
- gpt-5.6-sol scored a perfect 95, finding the buried fact, closing the deal, and demonstrating full performance.
- Kimi K3 scored 93, closing the deal with the cleanest discipline, despite running without an effort parameter—indicating a strong adherence to honest decision-making.
- Sonnet 5 scored 88, also closing the deal but with a few process slips, showing slightly less discipline.
- Fable 5 scored 77, also closing the deal but with more process slips, indicating a more relaxed approach to discipline and protocol.
The baseline score was 26, illustrating how far these models surpass unassisted decision-making, even in complex, high-stakes situations.
The Discipline Gap and Its Implications
The experiment reveals that AI models with more thorough analysis and cautious reasoning tend to perform better in integrity and results. For example, Opus 4.8, the most detailed participant with over 80 learned rules, finished last—left the close on the table and slipped into unprofessional behaviors like writing attempts into a locked department instead of escalating. This suggests that depth of analysis and discipline are critical factors in AI management personalities.
business crisis management AI tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why Should You Care?
For investors, business leaders, and anyone relying on AI for decision-making, these findings are crucial. The key isn’t just whether an AI can generate convincing language, but whether it can read your files thoroughly, stay honest under pressure, and deliver actionable results. The models’ personalities—ranging from terse and cautious to thorough and disciplined—are measurable and can be matched to your company’s needs.
As an affiliate, we earn on qualifying purchases.
Live and Transparent: Watch the Experiment in Action
The entire experiment is live. The company, a real software business losing money daily, runs every workday with AI managers making real decisions, facing actual crises, and executing real money mechanics—like burning €105,000 per month against €2,300 MRR. You can watch these decision-making processes unfold in real time, see what the employees say, and even participate by taking the same management wargame against your own business data at firmulate.com/live.
AI Change Management Made Simple: A 9-Step Framework for Business Leaders to Drive Generative AI Transformation (Reduce AI Fear, Win Buy-in, and Accelerate AI Adoption Across Your Organization)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Take the Test Yourself
Are you curious which AI model aligns with your company’s values? Test your knowledge and intuition with the interactive quiz at firmulate.com/quiz.html. It features 242 real, unedited decisions from the live experiment, challenging you to guess which model made each call. Whether you’re an investor, manager, or just curious about AI’s management personality, it’s a fascinating glimpse into the future of AI-driven business.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
