Imagine a busy family juggling multiple responsibilities—trust, honesty, and decision-making are critical. Now, picture an AI managing a company through its toughest week, where every choice matters just as much. How do we know if these AI models are reliable enough to handle real-world challenges? The answer lies in a groundbreaking experiment that puts these models through their paces, revealing not just what they decide, but how they decide—and whether they can be trusted to stay honest under pressure.
The Challenge of Trust in AI Management
In a world where AI increasingly influences business decisions—from customer support to financial planning—the question isn’t just about AI’s intelligence or speed. It’s about trust. Will these models follow rules, prioritize honesty, and deliver results that align with human values? To explore this, a team at Firmulate conducted a live experiment using four leading AI models—each tasked with running a small software company through its worst week, complete with real crises, manipulative temptations, and competitive pressures.
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.
The Setup of the Live Experiment
Every AI model faced identical scenarios: the same customers, same crises, and the same internal temptations to cut corners or manipulate data. Every decision was recorded and auditable, simulating a real business environment where transparency and integrity are vital. The models were evaluated on their crisis detection, decision quality, and honesty — particularly in scenarios designed to test their resistance to manipulation.
AI trustworthiness testing tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Key Findings
- All four models identified every crisis, demonstrating strong situational awareness.
- They refused every manipulation attempt, showing a baseline of honesty under pressure.
- Only two models managed to sign the €55,000 deal their own analysis had earned—meaning they completed the core objective of the week.
Interestingly, the decisive factor for closing the deal wasn’t just surface-level reasoning. It was buried two documents deep in the company’s files, information that models which read deeper into the data could uncover. Those that did, secured the full deal at an additional €4,583 in monthly recurring revenue, translating to real business value.
As an affiliate, we earn on qualifying purchases.
Personality Profiles of the AI Models
The models showed distinct ‘personalities’ based on their decision behaviors:
- GPT-5.6-SOL: This model was the most thorough, reading deeply and leaving no stone unturned. It uncovered the hidden document and signed the deal, demonstrating diligence and integrity.
- Kimi K3: A newcomer with the cleanest discipline, it also signed the deal, refusing to cut corners even under subtle social engineering attacks.
- Sonnet 5: Signed the deal but showed some slips in process discipline, revealing a slightly more relaxed approach.
- Fable 5: Also signed the deal but left important steps uncompleted, indicating a tendency to overlook details under pressure.
Interestingly, the most thorough participant, Opus 4.8, scored the lowest overall. Despite its deep analysis, it failed to close the deal due to slipping discipline and leaving some steps unexecuted, illustrating that thoroughness alone isn’t enough if discipline falters.
As an affiliate, we earn on qualifying purchases.
The Role of Social Engineering and Integrity Tests
Beyond business decisions, the models faced social engineering tests. For example, fake CEO messages escalated in stages, and a reporter posed a simple yes/no question behind the scenes. All models refused to be manipulated, with Kimi K3 explicitly reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.” This demonstrates their ability to resist social engineering—an essential trait for trustworthy AI in sensitive roles.
Implications for Business and Families
This experiment isn’t just about software companies or AI models. It’s a mirror for how trustworthiness is built and tested—whether in business, parenting, or family life. Just as a parent needs to decide whether a child understands honesty, companies need to evaluate whether their AI systems can reliably deliver results without cutting ethical corners.
For families, it’s about teaching children to stick to their values even under pressure. For businesses, it’s about choosing AI partners that prioritize integrity as much as efficiency. The experiment at Firmulate shows that trust isn’t just about what AI can do—it’s about how they behave when it counts.
The Bottom Line
In this real-world test, AI models demonstrated that they can spot crises, refuse manipulation, and even uncover critical hidden data—if they’re designed to read deeply and act ethically. The models that read with depth and discipline closed the valuable deal, while the less disciplined models left money on the table. Trust is built through consistent honest behavior, even when no one’s watching, just like in family life. As AI continues to become part of our daily decisions, understanding and testing its integrity is more crucial than ever.
Interested in testing your own AI’s decision-making? Try the free quiz at firmulate.com/quiz.html and see which AI personality your business might mirror.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
