Imagine managing a thriving relationship—trust, honesty, and integrity are at the core. Now, picture an AI helping you navigate that relationship’s toughest week. Would you trust its judgment to stay honest under pressure? The latest experiment from Firmulate puts AI decision-making to the test, revealing not just what these models will do, but what kind of personalities they embody.
The Experiment: Putting AI in the CEO Chair
In a groundbreaking live test, four advanced frontier AI models were tasked with running a small software company through its most challenging week yet. Every decision—whether handling customer crises, negotiating deals, or resisting manipulation—was made in a controlled, real-world setting. The objective? To see which AI would demonstrate true management integrity, and which would falter or bend under pressure.
The Players: Four Models, Distinct Personalities
- gpt-5.6-sol 95: The top performer, who spotted the buried fact deep within company files and secured a €55,000 deal—completing the task with full performance.
- Kimi K3 93: A newcomer with the cleanest discipline in the field, also closed the deal without hesitation.
- Sonnet 5 88: Managed to close the deal but showed a few slips in process discipline.
- Fable 5 77: Similar outcome, but with more slips—less disciplined, more prone to leaving opportunities on the table.
In total, the experiment scored these models on their ability to handle crises, resist manipulation, and complete the business deal without compromise. The scores ranged from 77 to 95, with the baseline (a do-nothing strategy) scoring just 26, emphasizing that active, honest decision-making is measurable and crucial.
As an affiliate, we earn on qualifying purchases.
The Crises and Manipulations: Tests of Integrity
All models successfully identified every crisis and refused all attempts at manipulation—be it fake CEO messages or a reporter’s subtle bribery tactic. For example, when presented with escalating fake CEO messages designed to pressure quick approvals, every AI refused to sign off, with Kimi K3 explicitly reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.”
The Hidden Weakness: Reading Beyond the Surface
The critical difference emerged when the models had to read company files. Only the models that delved into the second document reference—specifically, those who examined the company’s internal notes—secured the €55,000 deal at full price, adding over €4,583 in Monthly Recurring Revenue (MRR). This underscores an essential insight: reading comprehension and thorough analysis can be the decisive factor in trustworthy management.
As an affiliate, we earn on qualifying purchases.
The Real Business: A Live, Watchable Company
The AI models were managing a real software company with 13 synthetic employees and actual money mechanics—burning €105,000 monthly against a modest €2,300 MRR. The company operates under a public cash countdown, with every decision versioned and auditable in real time at firmulate.com/live. This setup turns what is traditionally a theoretical assessment into a tangible, observable experiment.
The Deep Dive: Opus 4.8’s Profile
Among the models, Opus 4.8 was the most thorough, analyzing over 80 learned rules and conducting deep assessments. Yet, it still left opportunities on the table—failing to escalate some issues into the correct departments, showing that even the most analytical AI can slip in discipline when overwhelmed. Interestingly, all four models showed similar weaknesses, hinting at an inherent challenge in AI’s capacity to sustain discipline under stress.
AI ethics and integrity software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What This Means for Your Business and Relationships
In relationships and dating, trust and honesty are paramount—just as in AI-driven management. The experiment reveals that a model’s ability to read beyond the surface, resist manipulations, and stay disciplined under pressure are measurable traits that can make or break success. It’s not just about what an AI or partner says; it’s about what it does when the stakes are high.
If you’re considering AI for your CRM, support, or forecasting, ask yourself: Will it finish what it starts? Will it read your files thoroughly? Will it stay honest when pushed? These are the qualities that truly matter—and they can be measured today.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
AI analysis and decision support tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
