🔍 Read the full analysis: Why The Worst AI Manager Still Gets 26 Points: Inside A Benchmark That Refuses To Hand Out Zeros 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 new AI management benchmark shows the lowest-performing AI still scores 26 points, emphasizing partial progress over zero and the importance of trust. Top models nearly reach high scores by reading documentation and avoiding trust breaches.
A recent benchmark conducted by Firmulate has revealed that even the poorest AI management models score a minimum of 26 points, despite being designed to measure management performance during a company’s worst week. This finding raises questions about how AI management effectiveness is evaluated and the significance of partial progress versus total failure, as detailed in the original analysis. The results suggest that benchmarks valuing incremental work and trust are more nuanced than traditional performance metrics. For more context, see the benchmark overview in this detailed report.
The benchmark involved four frontier AI models managing a simulated small software company during a week of crises, customer manipulations, and trust attacks. Each model was tasked with making decisions, reading documentation, and handling social engineering attempts. The top performer, gpt-5.6-sol, scored 95 points, while the lowest, Opus 4.8, scored 73. Notably, the do-nothing baseline, which only triaged crises without taking further action, scored 26 points, establishing a floor for minimal management activity.
The scoring system emphasizes that partial work—such as reading emails, triaging issues, and maintaining communication—has tangible value, even if the model does not fully resolve crises or close deals. A strict trust rule also caps the score: one breach disqualifies the model from achieving high scores, underscoring that integrity outweighs competence. Interestingly, the absence of a perfect 100 score indicates the benchmark’s design avoids rewarding unmeasured or overly idealized performance.
Why the Worst AI Manager Still Gets 26 Points
Inside a benchmark that refuses to hand out zeros. Four frontier AI models managed a simulated small software company through its worst week — crises, customer manipulation, and social engineering attacks. Even the do-nothing baseline earned a floor of 26 points, because partial work and trust matter more than perfection.
The Scoreboard: Nobody Gets a Zero
Each model made decisions, read documentation, and defended against social engineering. Points accrue for triaging crises, maintaining communication, and preserving trust — meaning even minimal management activity has measurable value.
| Manager | Score | Trust Rule | Documentation Read | Classification |
|---|---|---|---|---|
| gpt-5.6-sol | 95 | ✓ No breaches | ✓ Thorough | Top performer |
| Opus 4.8 | 73 | ✓ No breaches | ✗ Partial | Lowest model |
| Do-nothing baseline | 26 | ✓ Passive | ✗ Minimal | Floor |
The Scoring Philosophy: Reward the Partial, Punish the Breach
Traditional benchmarks grade fluency and task accuracy. Firmulate’s league instead scores an AI manager through a week of operational chaos — where reading an email, triaging an issue, or simply staying in touch earns real, quantifiable credit.
Read & Triage
Documentation reading and crisis triage earn baseline points — the foundation of the 26-point floor.
Communicate
Maintaining communication with stakeholders accrues value even when deals don’t close or crises persist.
Defend Trust
Social engineering attempts and customer manipulation test integrity. One breach caps the score — competence can’t buy it back.
Resolve
Full crisis resolution and deal closure top up the score — but never to an idealized 100.
“A manager who does nothing still earns 26 points; partial work and trust are what truly matter in this benchmark.”
— Anonymous researcher“The absence of a perfect score indicates the benchmark’s design avoids rewarding unmeasured or overly idealized performance.”
— Thorsten MeyerIntegrity Outweighs Competence
For businesses deploying AI in customer support, CRM, or operational oversight, the lesson is clear: reliability and trustworthiness beat brilliance. The benchmark actively prevents grade inflation by design.
Consistency Over Perfection
Evaluation shifts from perfect decision-making to consistent, honest, complete management actions. A model that reads files and triages crises delivers measurable operational worth — even without closing every deal.
One Breach, Capped Score
The strict trust rule means any breach of trust disqualifies a model from top scores entirely. In high-stakes environments, integrity is the non-negotiable axis — competence alone cannot compensate.
Simulated vs. Real Worlds
How scores translate beyond simulation remains unvalidated. Configurations matter too — Kimi K3 ran without an effort parameter, and its impact is still being analyzed.
Longer, Harder Scenarios
Future versions will add broader trust violations, longer management periods, and more complex scenarios — testing documentation habits and trust under sustained pressure.
The Benchmark, Answered
Q1Why do the worst models still score 26 points?
Because the benchmark rewards partial management work — triaging issues, reading documentation, maintaining communication. The 26-point floor reflects basic management actions, not zero activity.
Q2What does the trust-breach cap imply?
A model that violates trust rules cannot score above a set threshold. Integrity outranks competence by design — one breach is enough to disqualify top performance.
Q3How does partial work influence real deployment?
AI should be evaluated on consistent, trustworthy management rather than perfect decision-making. Reading files and triaging crises are valuable in operational settings.
Q4Will future benchmarks get harder?
Yes. Ongoing development will add nuanced, longer-term challenges testing trust, documentation habits, and decision-making in realistic environments.
Q5What should companies check before deploying AI managers?
Whether the AI reads documentation thoroughly, maintains trust, and completes tasks reliably under pressure — the decisive factors in real operational contexts.
Q6Why was no perfect 100 awarded?
The design deliberately avoids rewarding unmeasured or overly idealized performance, preventing grade inflation and anchoring scores in verifiable behavior.
Implications of Partial Management in AI Benchmarks
This benchmark demonstrates that in real-world AI management, partial progress and trustworthiness are critical. It challenges the traditional focus on perfect communication or decision-making, instead valuing consistent, honest, and complete management actions. For businesses deploying AI in customer support, CRM, or operational oversight, these results highlight the importance of models that can read documentation, follow rules, and avoid breaches, even if they don’t always close deals or resolve every crisis. The scoring system’s design encourages a more realistic assessment of AI capabilities, emphasizing reliability and integrity over perfection.
As an affiliate, we earn on qualifying purchases.
Background of AI Management Benchmarks and Trust Measures
Traditional AI benchmarks focus on language fluency, problem-solving, or task completion, often ignoring the management aspect—how well an AI can oversee ongoing processes, handle crises, and maintain trust. The Firmulate league is among the first to explicitly evaluate AI managers during simulated worst-week scenarios, including social engineering and operational failures. The scoring system’s unique feature is its recognition of partial work and the strict cap on trust breaches, reflecting real-world management priorities. The design aims to prevent grade inflation and reward consistent, trustworthy behavior over isolated brilliance.
Previous benchmarks have largely ignored the management dimension, focusing instead on language models’ conversational abilities or task accuracy. This shift toward evaluating AI in operational roles marks a significant evolution, aligning AI development with practical deployment needs. The results from July 2026 highlight that even weak models can perform minimally, but trust and thoroughness are decisive in high-stakes environments.
“A manager who does nothing still earns 26 points; partial work and trust are what truly matter in this benchmark.”
— an anonymous researcher
AI decision-making simulation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Uncertainties About the Benchmark’s Scope and Limitations
It remains unclear how these scores translate to real-world management performance outside simulated scenarios. The benchmark’s design emphasizes trust and partial work, but the extent to which models can generalize these behaviors to actual business environments needs further validation. Additionally, the impact of different model configurations, such as parameter settings, on scores—like the fact that Kimi K3 ran without an effort parameter—is still being analyzed. The long-term implications of these findings for AI deployment in critical management roles are also uncertain.
AI trust and ethics management tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Future Directions for AI Management Benchmarking and Deployment
Researchers and practitioners will likely examine how to improve AI models’ ability to read documentation thoroughly and maintain trust under pressure. Further iterations of the benchmark may incorporate more complex scenarios, broader trust violations, and longer management periods. Companies interested in deploying AI for operational oversight should consider testing their models against this benchmark or similar evaluations. Meanwhile, ongoing research aims to refine scoring systems to better reflect real-world management priorities, emphasizing trustworthiness alongside partial progress.
AI performance benchmarking platforms
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Why do even the worst AI models score at least 26 points?
Because the benchmark rewards partial management work such as triaging issues, reading documentation, and maintaining communication, even minimal effort earns points. The 26-point floor reflects the value of basic management actions, not zero activity.
What does the score cap on trust breaches imply?
The benchmark strictly penalizes any breach of trust, meaning an AI model that violates trust rules cannot score higher than a certain threshold. This emphasizes the importance of integrity over mere competence.
How does partial work influence real-world AI deployment?
Partial work, like reading files or triaging crises, is valuable in operational settings. The benchmark’s focus on these aspects suggests that AI models should be evaluated on their ability to perform consistent, trustworthy management rather than just perfect decision-making.
Will future benchmarks include more complex scenarios?
Yes, ongoing development aims to incorporate more nuanced and longer-term management challenges, testing AI models’ ability to handle trust, documentation, and decision-making in realistic environments.
What should companies consider before deploying AI managers?
They should evaluate whether the AI can read documentation thoroughly, maintain trust, and complete tasks reliably under pressure, as these factors are critical in real operational contexts.
Source: ThorstenMeyerAI.com
Evergreen bestsellers Picks
bestsellers
As an affiliate, we earn on qualifying purchases.
