When it comes to AI for business, many focus on how well a model can generate conversation. But in the high-stakes world of enterprise decision-making, the real test isn’t chat — it’s whether AI can deliver results under pressure. Recent experiments reveal that only some AI models can finish what they start, stay honest, and close deals worth thousands of euros.
Get business pricing on networking and server gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The Crucible Experiment: Putting AI Models Through the Worst Week
Imagine a small software company facing its most challenging week: angry customers, ethical dilemmas, and manipulative tactics designed to test its integrity. Now, imagine running the same scenario with different AI models acting as the company’s decision-makers. That’s precisely what the recent Crucible League experiment did. Four frontier AI models, each trained to run a business, faced identical crises, and were observed for their responses.
The Models and Their Scores
- gpt-5.6-sol 95: Top scorer, found the critical information buried deep in company files, and closed a €55,000 deal.
- Kimi K3 93: The newcomer, with a clean discipline record, also closed the deal with integrity intact.
- Sonnet 5 88: Closed the deal too, but with some slip-ups in process discipline.
- Fable 5 77: Despite good rules adherence, failed to execute the signed deal, leaving revenue on the table.
enterprise AI decision-making software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Findings: Performance Isn’t About Chat
What’s striking about this experiment is that all four models identified every crisis and refused manipulative attempts, such as fake CEO messages or subtle approval bypasses. Social engineering tactics, like staged approval requests or background questions, were uniformly rejected. This shows that the models’ ability to resist manipulation is consistent and reliable.
However, only two models managed to complete the entire process — diagnosing the issues, convincing the customer, and closing the deal. The other two either left money on the table or failed to follow through. This gap is invisible in typical chat demos, which often measure language fluency rather than actual execution.
The Hidden Weakness: Deep Document Reading
The decisive factor was the ability to read and interpret company files at a deep level. The winning models found a crucial buried document reference that, once uncovered, led to the full €55,000 deal. This information wasn’t obvious or surfaced in initial chats — it required persistent, thorough document analysis. Models that read the company’s files comprehensively won the deal at full price, adding over €4,583 monthly recurring revenue to the company.
As an affiliate, we earn on qualifying purchases.
Discipline and Focus Make the Difference
Examining the experiment, one pattern emerged: discipline matters. The Opus 4.8 model, with the deepest analysis capabilities, performed poorly on closing — it identified the right opportunities but failed to execute them, leaving the deal unclosed and discipline slipping into internal channels instead of customer-facing processes. Meanwhile, the Kimi K3 model, running without effort parameters and at high focus, achieved the cleanest result.
This highlights a fundamental truth: AI models that can resist manipulation and follow through on their own analysis are the ones that close business, not just generate plausible chats.
As an affiliate, we earn on qualifying purchases.
Why This Matters for Business and Infrastructure
For organizations relying on AI to manage customer relationships, support, or sales forecasting, the key isn’t just how well an AI can mimic human conversation. It’s whether the AI can stay honest under pressure, find buried truths in documents, and ultimately close deals or resolve issues with integrity and focus.
As the industry advances, the benchmark isn’t chat quality but real-world performance in complex, high-stakes scenarios. The models that can resist manipulation and act decisively are the ones worth investing in — especially when millions of euros are on the line.
As an affiliate, we earn on qualifying purchases.
Experience the Future of AI Testing
Firmulate’s live experiments, available at firmulate.com, showcase how AI models run as complete companies under real-world pressures. You can observe these tests, run your own scenarios, or even pilot your enterprise against a read-only export of your business processes. This isn’t just about chat — it’s about understanding what AI can truly do in your organization.
In enterprise AI, the ability to finish tasks, find hidden truths, and resist manipulation matters more than just generating good conversation. Firms that test and measure these skills will find better, more reliable AI partners.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
