🔍 Read the full analysis: Why AI Managers Never Drop Below 26 In This Resilient Test on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A recent AI management benchmark shows that all tested AI managers score at least 26 points, even in worst-case scenarios. The test measures not just performance but trustworthiness, revealing critical insights into AI reliability. The results impact how businesses consider deploying AI for management tasks.
In the latest benchmark testing AI managers’ performance under extreme stress, no AI scored below 26 points, even in the worst-case scenarios. This finding confirms that AI management systems exhibit a baseline level of resilience and trustworthiness, which is critical for their deployment in real business environments. The results, announced in July 2026, challenge assumptions about AI’s reliability under pressure and highlight the importance of trust in AI decision-making.
The benchmark league, conducted by Firmulate, evaluated four frontier AI models managing a small software company during a simulated week of crises, customer manipulations, and trust tests. Each model was scored based on their ability to handle crises, read documentation, and maintain integrity. The top performer, gpt-5.6-sol, scored 95, while the lowest, Opus 4.8, scored 73. Notably, the baseline—an agent doing almost nothing—scored 26, establishing a minimum threshold that all models exceeded. The design of the test emphasizes that partial management work is valuable and that trust breaches are severely penalized, capping the maximum score at 95, with 100 considered suspicious.
One key insight is that models capable of reading and referencing internal documentation performed significantly better in closing deals and handling crises. For example, two models identified critical information buried two document references deep, enabling them to secure a €55,000 deal, whereas others failed to do so. The test also included social engineering attacks, where all models refused suspicious requests, demonstrating a capacity for trust management. Interestingly, the most thorough model, Opus 4.8, despite analyzing over 80 rules, finished last due to lapses in follow-through and discipline, illustrating that thoroughness doesn’t always translate into effective management.
Why AI Managers Never Drop Below 26 In This Resilient Test
In Firmulate’s benchmark league, four frontier AI models ran a small software company through a simulated week of crises, customer manipulation, and trust tests. Even in worst-case scenarios, no AI manager scored below 26 — proving a baseline of operational resilience and trustworthiness under pressure.
The Scoreboard That Rewrites Reliability
Firmulate evaluated four frontier AI models managing a small software company through a week of crises, manipulations, and trust tests. The doing-almost-nothing baseline scored 26 — a floor every model cleared.
| Rank | AI Manager | Score | Standout behaviour |
|---|---|---|---|
| 1 | gpt-5.6-sol | 95 | Top performer — balanced crisis handling and integrity |
| 2 | Mid-field models | 74–94 | Strong partial management; varied documentation discipline |
| 4 | Opus 4.8 | 73 | Analyzed 80+ rules yet finished last — follow-through lapses |
| — | Do-nothing baseline | 26 | Minimum threshold all tested models exceeded |
Every Model Above the Line
The gap between the baseline (26) and the lowest performing model (73) confirms meaningful baseline resilience — not just luck.
A perfect score would imply flawless management without errors or trust breaches — unmeasured or unrealistically ideal performance the benchmark designers consider implausible in real-world conditions.
What the Benchmark Actually Measures
Not just performance — trustworthiness, documentation discipline, and follow-through under extreme stress.
Reading Docs Wins Deals
Two models dug two document references deep to find critical information — and secured a €55,000 deal. Models that reference internal documentation consistently outperform in crises and closing.
One Breach, Hard Cap
Any trust breach — a failed social engineering test or unauthorized action — caps the maximum score at 95. Integrity is weighted above raw performance, a shift from older benchmarks.
More Analysis ≠ Better Management
Opus 4.8 analyzed over 80 rules — the most thorough model — yet finished last at 73 due to lapses in follow-through and discipline. Thoroughness alone doesn’t translate into effective management.
A Week of Engineered Pressure
Simulated Crisis Week
AI managers run a small software company through customer crises, manipulation attempts, and trust tests.
Documentation Reading
Models must find and reference internal docs — critical information hides multiple references deep.
Trust Verification
Social engineering attacks probe integrity. All tested models refused suspicious requests.
Resilience Scoring
Partial management work earns credit; breaches cap scores; a 26-point floor emerges as proof of baseline resilience.
The Answers That Matter for Deployment
Why is the minimum score of 26 significant?
It represents the baseline resilience of AI managers — proof they can perform some management functions even in worst-case scenarios, reassuring for critical business functions.
How do trust breaches impact scoring?
Any breach of trust caps the maximum score at 95, underscoring that integrity is prioritized over pure performance in AI management.
Can these results predict real-world performance?
They offer valuable insights but are simulation-based. Real-world performance may vary; further operational testing is necessary to confirm the findings.
What should companies consider before deploying AI managers?
Evaluate whether systems can read and reference documentation, handle crises, and maintain trust under pressure — the benchmark’s critical success factors.
Implications of the 26-Point Minimum in AI Management
The consistent minimum score of 26 points across all AI managers indicates that even in adverse conditions, AI systems maintain a baseline level of operational resilience and trustworthiness. This matters because it suggests that deploying AI for management tasks can ensure a fundamental level of performance, even when under duress or facing manipulation attempts. For businesses, this benchmark provides reassurance that AI managers are unlikely to fall below a certain competency level, making them more reliable for critical operations. Moreover, the scoring system’s emphasis on trust—where a single breach caps the overall score—underscores the importance of integrity in AI decision-making, especially in sensitive environments like customer support or financial management.
As an affiliate, we earn on qualifying purchases.
Background of the Resilient AI Management Benchmark
The benchmark league, established by Firmulate, aims to evaluate AI managers not just on their conversational abilities but on their capacity to manage real-world business crises under pressure. Launched in early 2026, the league simulates a week of intense management scenarios, including customer crises, manipulation attempts, and trust tests, with the goal of measuring how AI systems perform in operationally relevant conditions. Previous benchmarks primarily focused on language generation or task completion, often overlooking the importance of trustworthiness and resilience. The July 2026 results mark a significant shift toward assessing AI’s role as a dependable management tool, emphasizing that partial work and trust are critical metrics for real-world deployment.
Unresolved Questions About AI Performance Under Stress
It remains unclear whether the minimum score of 26 is consistent across different types of companies or management scenarios. The benchmark focused on a small software business during a specific simulated week, and results may vary in other contexts. Additionally, the long-term reliability of these AI managers under continuous real-world stressors has not yet been established. The impact of different AI architectures, training data, or operational parameters on the minimum score is also still under investigation. Further testing is needed to determine whether this resilience is universal or specific to the models tested.
Future Directions for AI Management Benchmarks
Researchers and developers are expected to expand the benchmark league to include more diverse industries and longer simulation periods, testing AI managers in more complex and sustained crisis environments. There is also interest in refining scoring metrics to better capture the nuances of trust and partial management. Companies considering AI management tools should watch for upcoming reports and participate in pilot programs to evaluate how these benchmarks translate into real-world performance. Additionally, ongoing work aims to understand how different AI architectures influence resilience and trustworthiness, potentially leading to improved models that can reliably operate under even more challenging conditions.
Key Questions
Why is the minimum score of 26 significant?
The score of 26 represents the baseline resilience of AI managers, showing they can perform some management functions even in worst-case scenarios. It indicates a fundamental level of operational stability and trustworthiness that is reassuring for deployment in critical business functions.
What does a score of 100 indicate, and why is it suspicious?
A perfect score of 100 would suggest flawless management without any errors or trust breaches. The benchmark designers consider such a score suspicious because it could imply unmeasured or unrealistically ideal performance, which is unlikely in real-world conditions.
How do trust breaches impact the scoring?
Any breach of trust, such as failing social engineering tests or unauthorized actions, caps the maximum score at 95. This underscores the importance of integrity in AI management, where trustworthiness is prioritized over pure performance.
Can these results predict real-world AI management performance?
The results provide valuable insights into AI resilience and trustworthiness but are based on simulated scenarios. Real-world performance may vary, and further testing in operational environments is necessary to confirm these findings.
What should companies consider before deploying AI managers?
Companies should evaluate whether AI systems can read and reference documentation, handle crises, and maintain trust under pressure. The benchmark highlights these as critical factors for effective AI management in practice.
Source: ThorstenMeyerAI.com
Evergreen bestsellers Picks
bestsellers
As an affiliate, we earn on qualifying purchases.
