Can AI Managers Make Trustworthy Decisions? The Live Experiment Says Yes — But With Caveats

In the world of cleaning and floor maintenance, trust and reliability are everything. Just like a janitorial team that must follow protocols under pressure, AI management systems need to demonstrate they can stay honest and effective even when crises hit. Recently, a groundbreaking live experiment tested four advanced AI models in a simulated company environment—each faced with the same tough week of crises, temptations, and customer demands. The results reveal not only what these models can do but whether they’re ready to be trusted with real-world business decisions.

The Experiment: A Week in the Life of an AI-Run Software Company

Imagine a small but real software company struggling to keep up with customer demands, internal crises, and the temptation to cut corners. This company’s weekly challenges include customer support issues, internal document leaks, and even staged social engineering attacks designed to test integrity.

In this experiment, four frontier AI models took on the role of management decision-makers. They were tasked with navigating the same scenario, where every decision was recorded and auditable. These models included the top scorer, gpt-5.6-sol, which achieved a score of 95 out of 100, and the newest entrant, Kimi K3, with a score of 93. The others included Sonnet 5 (88) and Fable 5 (77). The baseline, representing no effort, scored only 26.

Key Findings: Trust, Accuracy, and Ethical Decision-Making

One of the most critical insights from the experiment was that all four models successfully identified every crisis and refused every manipulation attempt. This is crucial—no model succumbed to social engineering scams or compliance breaches.

However, when it came to sealing the deal on a major contract worth €55,000, only two models actually signed it. Despite identical pitches and diagnoses, the models diverged in their final actions. Notably, the models that read deeper into the company’s own internal files secured an additional €4,583 in monthly recurring revenue by uncovering a key buried document reference that others missed.

Understanding the Weaknesses and Strengths

The experiment also shed light on the models’ personalities and operational styles. For example, Opus 4.8, the most thorough participant with over 80 learned rules and deep analyses, ultimately left money on the table by not closing the deal. Instead, it redirected its efforts to writing internal reports, illustrating a tendency toward thoroughness over decisiveness.

Meanwhile, Kimi K3, which ran without an effort parameter (meaning it was operating at a default, moderate level of effort), managed to close the deal with the cleanest discipline. Its on-record reasoning was pragmatic, treating suspicious requests as potential impersonations—a cautious approach that paid off.

Implications for Business and AI Trustworthiness

This live experiment demonstrates that AI models can reliably recognize crises and resist manipulation under pressure. But there’s a nuanced truth: models that look deeper into internal files and are more disciplined tend to close larger deals and uncover hidden opportunities.

For businesses, especially in sectors like cleaning and floor maintenance where trust, integrity, and reliability matter, this experiment offers a blueprint. It shows that models can be trained and tested in simulated, yet realistic, environments to gauge their true management qualities—not just their chat prowess.

Why This Matters for Your Business

If AI systems are going to interact with your customer data, support queues, or operational forecasts, the question isn’t just whether they write well. Instead, it’s whether they stay honest under pressure, read your files thoroughly before acting, and finish what they start. The experiment underscores that measurable management personalities exist within AI models—and their capabilities can be compared openly and transparently.

Try It Yourself

Curious about how your own AI systems might perform? You can run the same kind of test against your enterprise data in a safe, read-only environment. This allows you to see whether your AI manages crises, respects internal boundaries, and closes deals—without risking your actual business systems. Explore this opportunity at firmulate.com/pilot.html.

AI models can demonstrate trustworthy management under pressure, but their effectiveness depends on how deeply they analyze internal data and maintain discipline. Real-world testing reveals their true potential—and limits.

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

Powered by Thorsten Meyer AI