🔍 Read the full analysis: The Reason The Worst AI Managers Still Manage 26 Points on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Recent AI management benchmarks show the lowest-performing models still score 26 points, indicating partial progress and highlighting trust and completion issues. The results shed light on AI’s capabilities and limitations in managing real-world business tasks, as detailed in the original analysis.
The final results of a novel AI management benchmark reveal that the lowest-scoring AI model, labeled as the do-nothing baseline, still managed to earn 26 points out of a possible high score, raising questions about the scoring system and the true capabilities of current AI management tools. This analysis is discussed in the original analysis. This benchmark, conducted by Firmulate, tested several frontier AI models over a simulated week of business crises, with the results emphasizing the importance of trust, completeness, and partial progress in AI management systems. For more details, see the coverage on the original source.
The benchmark involved four AI models managing a small software company during a simulated week filled with customer crises, manipulative attempts, and social engineering attacks. Each model’s decisions were fully auditable, and the scoring was designed to reflect real-world management tasks, including trustworthiness and task completion.
The top performer, gpt-5.6-sol, scored 95 points, while the second-place model, Kimi K3, scored 93. Despite the high scores for the leaders, the baseline model—designed to do almost nothing—still scored 26 points. This baseline was not zero, illustrating that even minimal management efforts are recognized and valued in the scoring system. The reason: partial work, such as triaging emails or reading documents, counts towards the score, reflecting real management activities.
Interestingly, the scoring system also includes a strict trust cap: if an AI breaches trust even once, it cannot score above 90, regardless of other performance. The absence of a perfect 100 score indicates that the system is intentionally designed to flag unmeasured or suspiciously perfect results. The results highlight that models which read their own documentation and avoid manipulation attempts tend to perform better, especially in closing deals and maintaining trust during social engineering tests.
The Reason The Worst AI Managers Still Manage 26 Points
A novel benchmark by Firmulate put frontier AI models in charge of a small software company through a simulated week of customer crises, manipulation attempts, and social engineering attacks. The lowest scorer — the “do-nothing” baseline — still walked away with 26 points. Here’s why partial work counts.
The Scoreboard: 95 to 26
Every decision was fully auditable. Scoring reflects real-world management: effectiveness, trustworthiness, and task completion. Note the gap between leaders and baseline — and the trust cap that no model can breach.
Why the Baseline Isn’t Zero
The scoring system deliberately recognizes partial work. Even minimal management activity — reading documents, triaging inboxes, acknowledging a crisis — mirrors what real managers do, and it earns points.
Triage Counts
Reading documents, sorting emails, and responding to inquiries are genuine management activities — the benchmark scores them, so doing “almost nothing” still yields 26 points.
Integrity Is a Hard Limit
Breach trust even once and the model cannot score above 90, regardless of other performance. Trustworthiness outranks raw output in this system.
No 100 By Design
The absence of a perfect score is intentional: the system flags suspiciously perfect or unmeasured results. Perfection signals something was missed.
“Trustworthiness is a hard cap in the scoring system; even the best models can’t surpass it if they breach trust once. This underscores the importance of ethical AI behavior.”
— Thorsten Meyer
From Simulation to Score
Crisis Week
Models run a small software firm through customer crises, manipulation, and social engineering attacks.
Auditable Decisions
Every management decision is logged and transparent for review by the scoring system.
Partial Credit
Triage, reading documentation, and replies all count — the baseline lands at 26.
Trust Check
One trust breach caps the score at 90. Integrity is non-negotiable.
Final Ranking
Effective, trustworthy, complete management rises to the top: 95 for gpt-5.6-sol.
Implications & Background
Traditional AI benchmarks measure language fluency or narrow tasks. The Firmulate league — launched July 2026 — is among the first to evaluate AI management skills in realistic, auditable business scenarios.
Reliability Over Fluency
Companies should select AI management tools based on trustworthiness, follow-through, and task completion — not surface-level performance. Partial progress is valuable, but trust remains paramount.
Docs & Deflection
Models that read their own documentation and resisted manipulation attempts performed best — especially in closing deals and maintaining trust during social engineering tests.
Does It Translate?
It remains unclear whether simulated scores map to real-world outcomes like customer satisfaction, revenue growth, or crisis resolution efficiency in live settings.
Harder Scenarios
Future iterations will test broader model ranges, enterprise integrations, and refined scoring that better separates partial from full management success.
At a Glance: FAQ
Q1Why does the lowest baseline still get points?
The scoring recognizes partial work — triaging emails, reading documents — which are valuable management activities. Even near-zero effort earns 26 points.
Q2What does the absence of a 100 score imply?
Designers treat perfection as suspicious or unmeasured. Trust and completeness cannot be faked, so a perfect score is intentionally out of reach.
Q3How do trust breaches affect scores?
One breach caps the maximum score at 90, regardless of other performance. Trustworthiness is non-negotiable in this evaluation.
Q4Can AI fully manage real business operations?
Not yet. Models handle tasks and crises, but trust, reading comprehension, and follow-through under pressure remain areas needing improvement.
Q5What should businesses take away?
Prioritize reliability, trustworthiness, and task completion over language fluency. Benchmarks like this reveal real operational capability before deployment.
Implications of Partial Progress in AI Management
The results demonstrate that current AI models can perform meaningful management tasks, such as crisis triage and trust maintenance, even if they do not achieve full success. The fact that the lowest baseline scores 26 points shows that partial work—like reading documents, responding to inquiries, and triaging crises—is recognized and valued, which is crucial as AI tools are increasingly integrated into business operations.
Moreover, the strict trust cap emphasizes that integrity and trustworthiness are non-negotiable in AI management. A model that breaches trust cannot reach top scores, underscoring the importance of ethical behavior in AI decision-making. For businesses, this benchmark offers a more realistic picture of AI capabilities, highlighting that partial progress is valuable but trust remains paramount. This could influence how companies select and evaluate AI management tools, prioritizing reliability and integrity over raw performance alone.
As an affiliate, we earn on qualifying purchases.
Background of AI Management Benchmarks
Traditional AI benchmarks focus primarily on language capabilities, such as conversation quality or information retrieval. Few tests measure AI’s ability to manage ongoing business processes or handle crises under pressure. The firmulate.com league, launched in July 2026, is among the first to explicitly evaluate AI models’ management skills in realistic scenarios.
The benchmark involved simulating a week of business crises, manipulative social engineering attacks, and trust challenges, with models making decisions that were auditable and scored based on effectiveness, trustworthiness, and task completion. The scoring system was intentionally designed to recognize partial work, with a minimum score of 26 for doing almost nothing and a maximum of 95 for near-perfect management.
Prior to this, most AI evaluation efforts either measured language fluency or narrow task-specific performance, leaving a gap in understanding AI’s capacity to manage complex, real-world operations. The firmulate league aims to fill that gap by providing a transparent, auditable, and realistic assessment of AI management capabilities.
“Trustworthiness is a hard cap in the scoring system; even the best models can’t surpass it if they breach trust once. This underscores the importance of ethical AI behavior.”
— Thorsten Meyer
AI decision-making simulation platforms
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unresolved Questions About AI Management Scoring
It remains unclear whether the scores accurately reflect real-world AI management capabilities outside simulated environments. The benchmark’s design emphasizes partial work and trust, but how these scores translate into operational effectiveness in live business settings is still under discussion.
Additionally, the absence of a perfect score raises questions about whether models can ever fully meet the benchmark’s standards for trust and completeness, or if these are inherently limited by current AI technology.
Further research is needed to determine how these scores correlate with actual business outcomes, such as customer satisfaction, revenue growth, or crisis resolution efficiency.
AI trustworthiness assessment tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Future Directions for AI Management Evaluation
Next steps include expanding the benchmark to more complex business scenarios and testing a broader range of AI models, including those integrated into real enterprise systems. Researchers and developers will likely focus on improving models’ ability to read and understand documentation, maintain trust, and complete tasks reliably under pressure.
There is also interest in refining scoring systems to better distinguish between partial and full management success, possibly incorporating real-world performance metrics. Additionally, companies may start adopting similar benchmarks to evaluate their AI tools before deployment in critical operational roles.
Ultimately, ongoing benchmarking efforts aim to establish clearer standards for AI management, emphasizing trustworthiness and task completion as key success factors.
As an affiliate, we earn on qualifying purchases.
Key Questions
Why does the lowest baseline score still get points?
The scoring system recognizes partial work, such as triaging emails or reading documents, which are valuable management activities. Even doing almost nothing earns 26 points, reflecting minimal but real effort.
What does the absence of a 100 score imply?
The benchmark designers treat a perfect score as suspicious, indicating unmeasured or unachievable perfection. It emphasizes that trust and completeness are critical and cannot be faked.
How do trust breaches affect AI scores?
If an AI breaches trust even once, it caps its maximum score at 90, regardless of other performance measures. Trustworthiness is a non-negotiable in this evaluation.
Can current AI models fully manage real business operations?
While they can handle certain tasks and crises, the scores suggest that full management—especially under pressure—is still a challenge. Trust, reading comprehension, and follow-through remain areas needing improvement.
What are the implications for businesses deploying AI?
Businesses should prioritize AI tools that demonstrate reliability, trustworthiness, and the ability to complete tasks, rather than just language fluency or superficial performance. Benchmarks like this provide valuable insights into real capabilities.
Source: ThorstenMeyerAI.com
Columbus Day / Indigenous Peoples' Day Picks
long weekend sales
As an affiliate, we earn on qualifying purchases.
