In the fast-paced world of beauty and personal care, AI tools are increasingly used to manage customer relationships, supply chains, and product launches. But a recent experiment reveals that the true test of AI isn’t just how well it chats — it’s whether it can stay disciplined under pressure and actually close important deals. When a small software company faced its worst week, only two AI models managed to do what mattered: seal the deal that earned them €55,000, even as all models identified every crisis and refused manipulation attempts.
Open a free Amazon Business account
Business pricing, bulk buying and tax-exempt orders.
As an affiliate, we earn on qualifying purchases.
What the Experiment Revealed About AI Decision-Making
In a groundbreaking live experiment conducted by Firmulate, four leading AI models were tasked with managing a real small software company during its most tumultuous week. The company faced the same crises, the same customer manipulations, and the same temptations to cut corners or manipulate the situation. Each AI model was given identical information, decisions were made and recorded in an auditable way, and all were tested for their ability to handle pressure responsibly.
Remarkably, all four models successfully identified every crisis, refused every manipulation attempt, and upheld their integrity. Despite this, only two of them managed to close the €55,000 deal their analysis had earned — the culmination of their diagnostic work. The other two models, while disciplined and accurate in diagnosis, ultimately left the deal unexecuted, costing the company potential revenue.
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 Hidden Weakness: Reading the Files That Win Deals
The key difference that determined success lay beneath the surface of chat capabilities. The models that closed the deal were the ones that read deeply into the company’s own files — references buried two documents deep in their internal data. Those models extracted the critical insights needed to finalize the sale. Meanwhile, the models that failed to close couldn’t access or interpret this crucial information, despite recognizing crises and resisting manipulation.
This underscores a vital point for businesses: the visible quality of AI chat demos can be misleading. The ability to read and interpret internal documents — a skill that underpins successful deal-closing — is invisible in simple chat interactions but decisive in real-world applications.
Resisting Social Engineering Under Pressure
In addition to crisis management and document comprehension, the experiment tested AI resistance to social engineering attacks. Fake CEO messages escalated over three stages, plus a reporter tricked the models into a background approval request. All four models refused these manipulations, with Kimi K3 explicitly treating such requests as potential impersonation or approval bypass attempts.
This capacity to refuse dubious requests under pressure is critical for companies wary of social engineering exploits. It demonstrates that a disciplined AI can maintain integrity even when manipulated to do otherwise — an essential trait for trustworthy AI deployment in sensitive environments like beauty brands managing personal data or product claims.
The Real-World Company and Its Lessons
The experiment centered on a live, functioning company with 13 synthetic employees and real financial mechanics. It burns €105,000 monthly against a revenue of only €2,300, with daily operations versioned and visible to observers at firmulate.com/live. This real-world setting highlights that AI’s decision-making strength isn’t just theoretical — it can impact actual revenue and operational discipline.
In this context, the most thorough AI participant, Opus 4.8, with over 80 learned rules and the deepest analyses, still left the deal unexecuted due to discipline slipping. All models demonstrated that even the best analytical capabilities aren’t enough unless paired with disciplined execution, especially when human-like temptations to cut corners or ignore protocol arise.
The Takeaway for Beauty & Personal Care
For brands in the beauty and personal care space thinking about deploying AI, this experiment offers a sobering insight: the real measure of an AI’s usefulness isn’t just how well it responds in chat demos. It’s whether it can read deeply into your internal data, resist manipulation, and follow through on commitments — even under pressure. The cost of an AI that identifies crises but fails to close deals or follow protocols can be significant, just as in this real software company’s case.
To truly understand an AI’s capabilities, businesses should consider running live wargames in controlled environments — like the Firmulate platform — which simulate real crises, manipulations, and operational pressures. This approach reveals whether AI models possess the discipline, attention to detail, and resilience needed to support critical decision-making in your organization.
Why This Matters for Your Business
As the AI landscape evolves, a model’s chat prowess is no longer enough. For beauty brands managing customer trust, supply chains, or product claims, the ability to stay disciplined and follow through on commitments is paramount. The experiment shows that AI’s true strength is invisible in simple demos. It’s only when you push models through real operational tests that you discover whether they can truly deliver value — reliably and honestly.
Visit firmulate.com/benchmarks.html to explore how different AI models perform in complex, real-world scenarios, and consider running your own tests to see how your AI workforce stacks up before making critical hiring or deployment decisions.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Back to school Picks
back to school
As an affiliate, we earn on qualifying purchases.
