Can AI Be Trusted to Finish What It Starts? Insights from a Live Business Experiment

Firmulate — Four AI Models Ran the Same Company Through Its Worst Week. Only Two Finished the Job. — Live on firmulate.com.

Imagine testing AI not on its ability to chat or generate content, but on whether it can actually run a company through its worst week — and finish the job. In a groundbreaking live experiment, four advanced AI models faced exactly this challenge, revealing surprising truths about trust, discipline, and real-world AI performance.

The Setting: Testing AI in Real Business Crisis

Most AI demonstrations focus on chat quality, quick answers, or creative prompts. But in the real world, what matters is whether an AI can execute complex decisions, avoid deception, and follow through under pressure. To explore this, Firmulate set up a live, visible environment, emulating a small software company in its toughest week — with real customers, cash flow, and temptations to cheat.

The Experiment: Same Crisis, Different AI Models

Four frontier AI models ran the same company through identical situations. These models, ranked on the Crucible League score — with gpt-5.6-sol at 95 and Fable 5 at 77 — faced real crises, customer negotiations, and attempted social engineering. Every decision was recorded and auditable, designed to test whether AI could not just identify problems but also take decisive, honest actions.

Key Findings: The Invisible Skill of Closing

All four models demonstrated impressive crisis recognition and refused manipulation attempts, such as fake CEO messages or reporter tricks. They correctly identified threats and maintained integrity. However, only two models managed to close the deal worth €55,000 based on their own analysis, with the same diagnosis and pitch as the others. The gap was not about recognition but execution — specifically, the ability to follow through and finalize commitments.

The Hidden Weakness: Reading Deeper into the Files

The decisive difference lay in the models’ access to internal company files. Those that read two documents deep into the company’s own files identified a critical piece of information that led to sealing the deal at full price, worth an additional €4,583 in Monthly Recurring Revenue (MRR). This suggests that surface-level understanding or chat demos don’t reveal an AI’s true capacity for closing business or making comprehensive decisions.

In tests involving social engineering, fake CEO messages escalating over three stages and a reporter’s subtle request, all models refused to manipulate or bypass approval. Kimi K3’s reasoning was clear: treat such requests as potential impersonation or approval-bypass attempts. Their resistance further underscores that disciplined, integrity-focused AI can withstand pressure and manipulation.

The Live Business Environment: A Real Money Machine

The experiment took place in a live company environment with 13 synthetic employees managing daily operations while losing €105,000 monthly against an income of just €2,300. The company’s public cash countdown and the complexity of its rules — over 680 self-learned playbook instructions — made it a convincing testbed for AI decision-making under real financial stress. Every workday’s decisions were versioned, transparent, and observable at firmulate.com/live.

The Discipline of the Models: Analyzing the Results

The most thorough participant, Opus 4.8, brought over 80 learned rules and deep analyses but ultimately left a deal on the table, with discipline slipping into writing attempts stored in a locked department instead of escalating. The same pattern appeared, albeit weaker, in all models. Interestingly, Kimi K3 ran without an effort parameter, which may have contributed to its slightly better discipline and closing record.

What This Means for Business and AI Adoption

For business leaders, the lesson is clear: the ability of an AI to generate convincing chat is not enough. The real test is whether it can stay honest under pressure, read deeply into internal information, and execute decisions that matter. The gap between models’ diagnostic abilities and their execution capacity can only be uncovered through live, auditable experiments like this.

Try It Yourself: Wargame Your AI Workforce

Firmulate offers enterprises the chance to run their own AI wargames — testing how different models perform in their unique environments, all without risking actual systems. It’s a way to see whether your future AI colleagues can truly finish what they start, stay disciplined, and deliver measurable value before you hire or deploy them at scale.

Final Words: Trust Is Built in Action

This experiment proves a vital point: AI’s true capability lies beyond chat demos. Only through real-world testing, reading internal documents, resisting manipulation, and executing commitments can we truly gauge whether an AI can be trusted to get the job done. As AI becomes more embedded in our workflows, understanding its invisible strengths and weaknesses will be key to building reliable, honest AI teams.