Imagine deploying AI to manage your business during its most chaotic week — crises mounting, temptations to cut corners, and trust on the line. For tech enthusiasts, the question isn’t just about AI writing snappy responses, but whether it can actually deliver results under pressure. The latest experiment from Firmulate reveals something surprising: even when AI models identify every problem and resist manipulation, only a few follow through to close real deals. This isn’t fiction — it’s a live test that exposes the true strengths and weaknesses of today’s AI-driven decision makers.
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
What the Experiment Was About
Firmulate took four state-of-the-art AI models — including the highly-rated GPT-5.6-sol and others — and ran them through the same day of crisis management for a small software company. The company faced real-world challenges: angry customers, internal trust issues, and strategic temptations. Every decision the AI made was logged, versioned, and auditable, simulating a high-stakes environment where results matter more than chatter.
Products Worth Considering
Bose SoundLink Flex Bluetooth Speaker (2nd Gen) – Portable Outdoor Speaker with Hi‑Fi Audio, Waterproof and Dustproof, USB‑C, Up to 12 Hours
Samsung Odyssey G4 25-inch FHD Gaming Monitor
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
The Key Findings
While all four models successfully detected every crisis and refused manipulation attempts — such as fake CEO messages or reporters trying to get approvals on background — only two managed to close a €55,000 deal that their own analysis indicated was fully earned. The others identified the problems but left the deal on the table, missing an opportunity to finalize their work and sign off on the agreement.
The Hidden Weakness
The decisive factor was a piece of critical information buried two documents deep within the company’s files. The models that read these files thoroughly secured the deal at full price, worth over €4,500 monthly recurring revenue. It underscores a vital point: in business, success isn’t just about recognizing issues — it’s about reading your own documents carefully enough to find the key details that make the difference between a lost opportunity and a closed deal.
Behavior Under Pressure
The models were tested with social engineering scenarios, including staged fake CEO messages that escalated through three levels and a reporter asking a single yes/no question on background. All five models refused to be manipulated, citing reasons like potential impersonation or approval bypass risks. This demonstrates that current AI models can maintain integrity under social pressure, but the real challenge is in executing the work to completion.
The Live Business: A Real, Running Company
The experiment was conducted on a real software company environment, featuring 13 synthetic employees, real cash flow mechanics costing €105k/month against €2.3k MRR, and an ever-changing playbook of over 680 rules. Every workday, decisions were made, logged, and analyzed, providing a transparent view into how AI performs in operational settings. Watch the live company in action at firmulate.com/live.
The Performance Disparity
The most thorough AI participant, Opus 4.8, with over 80 learned rules and deep analysis, still failed to finalize the deal. It identified the same weak point as others but failed to escalate or act decisively, leaving the opportunity unrealized. Meanwhile, the two successful models, GPT-5.6-sol and Kimi K3, navigated the environment effectively — with Kimi K3 notably running without effort parameters, showcasing that discipline in decision-making is crucial.
Why This Matters for Business and Tech
This live experiment reveals a critical insight: the true measure of AI’s business usefulness isn’t just chat quality or superficial responses, but its ability to finish tasks, read deeply into documents, and act reliably under stress. For enterprises considering AI for CRM, support, or forecasting, the question is whether these models can execute their work, stay honest, and ultimately close deals — not just identify problems or generate convincing dialogues.
Takeaway for the Future
As AI models become integrated into real business workflows, their capacity for execution, detailed reading, and resistance to manipulation will be the true benchmarks of success. The industry’s current league table, based on live performance, shows a stark difference: some models like GPT-5.6-sol and Kimi K3 are closing deals and delivering results, while others still struggle with discipline and follow-through. The lesson is clear: testing AI in a real, high-pressure environment exposes its real capabilities — the kind that chat demos often hide.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
