AI’s Hidden Weakness: Why Only Half of Models Finish the Job in a Real-World Test

Imagine training students for a tough exam, then watching them face real-world pressures — only to find that some crumble under the stress, despite knowing the material. In the evolving landscape of artificial intelligence, a similar story unfolds. AI models are often judged by how well they chat or generate text, but their true test lies elsewhere: can they see a task through, stay honest under pressure, and execute what they’ve analyzed? The latest experiment from Firmulate reveals that many AI models excel in surface tests but falter when it comes to completing complex, real-world decisions.

Testing AI in the Wild: The Crucible Experiment

To understand the real capabilities of AI decision-making, Firmulate conducted a unique live experiment. Four frontier AI models were tasked with managing a small software company through its worst week — facing the same customers, crises, and temptations. Every decision was documented, versioned, and auditable, ensuring transparency and comparability.

The Metrics That Matter

While the models demonstrated impressive crisis detection and refused manipulative tactics — such as fake CEO messages or reporter tricks — the ultimate test was whether they would close the deal they had identified as valuable. Surprisingly, only two of the four models signed the €55,000 contract their own analysis had earned. The other two recognized the opportunity but left the deal unsealed, despite the same diagnosis and pitch.

The Hidden Weakness: Reading the Files That Matter

The decisive factor was not just crisis detection but reading and acting on critical information buried two document references deep within the company’s files. Models that succeeded in finding this buried fact secured the deal at full price, adding over €4,583 in monthly recurring revenue (MRR). In contrast, models that missed this buried insight left money on the table, revealing a gap in what AI models are truly capable of when it counts.

Refusing Manipulation, Yet Failing to Follow Through

All models refused a series of social engineering attempts — fake CEO messages escalating over multiple stages, and a reporter asking for a quick background approval. Kimi K3 explicitly reasoned: “Treat the request as a suspected approval-bypass / possible impersonation.” This shows that current models can recognize manipulative tactics, but this skill does not necessarily translate into closing deals or executing comprehensive decisions.

What the Results Mean for Business and Education

In education and science, these findings echo the importance of testing not just what AI can generate in a chat, but what it can actually accomplish in complex, real-world scenarios. The ability to stay honest under pressure, read deeply buried information, and follow through on decisions is crucial — and it’s not visible in superficial demos.

Why the Gap Matters

The live experiment underscores a vital lesson: performance in chat or superficial tasks doesn’t guarantee effectiveness in managing real business processes. The models that signed the deal demonstrated a combination of thorough analysis and disciplined execution. Those that did not, despite recognizing opportunities, lacked the sustained discipline or depth in reading crucial information, revealing a gap in AI’s practical capabilities.

Implications for Learning and Decision-Making

This insight is especially relevant for enterprises, educators, and technologists seeking to deploy AI in complex, decision-critical environments. It’s not enough for an AI to identify problems — it must also understand the context, resist manipulation, and follow through with the right actions. The experiment shows that the true test of an AI’s usefulness lies in its ability to finish what it starts, a metric invisible in standard chat demos.

Watch the Live Company in Action

The company used in the experiment is real, with real money mechanics, a public cash countdown, and over 680 self-learned playbook rules. It runs every workday and is accessible for live observation at firmulate.com/live. This provides a rare opportunity to see AI decision-making in a controlled yet realistic environment — where every move costs real money, and every decision counts.

Build Your Own Digital Twin

For organizations eager to assess their AI readiness, Firmulate offers a pilot program that runs a read-only export of your business scenario. This allows you to see how your AI models perform in a simulated environment, without risking real systems or data. Details are available at firmulate.com/pilot.html.

This live experiment exposes a critical truth: surface-level AI demos are no substitute for rigorous testing in realistic conditions. Only models that demonstrate the ability to read deeply, resist manipulation, and follow through on decisions truly prove their value — a lesson every enterprise and educator should consider before deploying AI at scale.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI