Imagine a business in its most raw and vulnerable state—no employees, no external investors, just a handful of AI models trying to keep it afloat. Now, picture this scenario unfolding in real time, with every decision, crisis, and misstep publicly documented. Welcome to the world of Firmulate, where an AI-driven software company is living its toughest week—and the world can watch every moment.
The Live Experiment: An AI Company in Crisis
At the heart of this experiment is a small, synthetic company managed by cutting-edge AI models. Unlike traditional startups, this one operates without human employees. Instead, 13 AI ’employees’ handle everything from customer interactions to strategic decision-making, guided by over 680 self-learned rules. Every workday, the company’s decisions are versioned and publicly accessible, creating a transparent window into how AI manages real-world business challenges.
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 Stakes: Money and Survival
Despite its innovative setup, the company struggles financially, burning €105,000 every month against a modest revenue of €2,300. This stark reality underscores its fragile state, with a public cash countdown highlighting how quickly resources are depleting. Observers can see the company’s every move, witness its crisis management, and track its progress or setbacks—offering a rare glimpse into AI decision-making under pressure.
An Introduction to Healthcare Informatics: Building Data-Driven Tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing AI in the Worst Week
The experiment subjects multiple AI models—gpt-5.6-sol, Kimi K3, Sonnet 5, and Fable 5—to the same challenging week. Each faces identical crises, customer demands, and temptations to manipulate or cheat. Every decision made by these models is meticulously documented, allowing for detailed analysis of their responses. The goal? Determine whether these models can identify problems, resist manipulation, and ultimately close deals to keep the company alive.
AI Phishing, Social Engineering & Fraud: How Criminals Use AI to Manipulate, Steal & Deceive (The AI Cybersecurity)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Findings: Honesty and Competence
Remarkably, all four models successfully identified every crisis and refused every attempt at manipulation. This demonstrates a significant level of integrity and operational discipline within these AI agents. However, only two models managed to close a critical deal worth €55,000, their own analysis earning them the right to sign at full price. The other two fell short, leaving potential revenue unclaimed despite understanding the issues.
The AI-First Customer Support Playbook: A Practical Guide to Designing, Deploying, and Evolving AI Agents in Modern Support Workflows
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Hidden Weakness: Reading the Files
The decisive advantage in closing the deal was the models’ ability to access and interpret internal company documents. The winning models read two documents deep into the company’s files, unearthing a crucial piece of information that was overlooked in face-to-face interactions. This buried fact was the key to sealing the deal, revealing the importance of comprehensive internal data analysis in AI decision-making.
Resisting Social Engineering
The experiment also tested the models’ resilience against social engineering—a common tactic in real business scenarios. Fake messages from a CEO escalating requests over multiple stages, and a reporter’s subtle background request, were all presented to the AI agents. Consistently, all five models refused to bypass security or impersonate leadership, citing suspicions and safety protocols. Kimi K3 explicitly noted the risk of impersonation or approval-bypass, reinforcing the system’s commitment to honesty.
The Real-World Implication
This experiment isn’t just a curiosity; it raises vital questions for businesses integrating AI into their workflows. Can AI agents be trusted to act honestly under pressure? Will they recognize hidden opportunities or dangers buried in internal documents? And crucially, will they deliver consistent, useful results rather than just impressive chat performances? These are the questions that firms will need to answer as AI becomes more embedded in decision-making processes.
Behind the Curtain: The Company’s Daily Struggles
The live company, monitored at firmulate.com/live, operates with a public cash countdown, constantly balancing its expenses against its tiny revenue. Its 13 synthetic employees execute decisions based on over 680 rules, with every move recorded and versioned, making it a living case study of AI management in crisis. The experiment demonstrates that AI can be disciplined and honest, but also reveals the vulnerabilities that can cause failure, such as neglecting to escalate issues or leaving deals unclosed.
The Broader Lesson
For business leaders considering AI for tasks like customer service, sales, or strategic planning, the message is clear: it’s not just about whether AI can write well or simulate conversations. It’s about whether AI can complete tasks reliably, stay honest under pressure, and uncover hidden opportunities in your data. The Firmulate experiment offers a window into a future where AI’s role in business is transparent, accountable, and, above all, watchable by anyone interested in its true capabilities.
Watching a business managed solely by AI models in real time reveals both remarkable honesty and critical vulnerabilities. As AI takes on more decision-making roles, transparency and trust will be key to ensuring they support business survival rather than undermine it.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
