Every one of the nine frontier AI models tested got worse at allocating scarce resources once you fed them charts and dashboards. That is not a typo. Adding visual business evidence made budget, capital, and capacity decisions measurably worse, while it improved almost every other executive task at the same time.

Turns out, “more data is better” is a dangerous assumption when you put AI in charge of high-stakes calls. This new study, built by researchers including the well-known NLP scientist Preslav Nakov, is one of the first to test AI CEOs the way we actually work: with a mix of written reports and visual evidence, not just text. And the result should make any leader pause before wiring every chart in your company into your decision-support agent.

Why This Paper Matters

Most AI benchmarks test models on text alone. That is a clean experiment, but not how business works. Your operations team reads a situation report and a forecast chart together. So the researchers behind C-SUITEBENCH built the first controlled benchmark that drops frontier models into the CEO chair across 50 real scenarios, tested two ways: text only, and text plus the visual evidence a real executive would see.

This matters because companies are rushing to stand up executive AI copilots and decision-support agents. The assumption underneath almost all of them is that feeding the AI more context makes it smarter. This paper is the first hard evidence the opposite can be true, exactly where you care most. Its a lesson any leader building these tools should file away.

That makes it an AI Assessment for companies, done properly: not “can the model read a chart,” but “does the chart make the decision better.”

Methodology, Explained Simply

Think of C-SUITEBENCH as a flight simulator for AI executives. Nine frontier models each played CEO across five tasks: diagnosing what’s wrong, prioritizing actions, allocating constrained resources, forecasting risk, and justifying a decision to the board. Every scenario came in two versions: a written situation report, and that same report paired with the charts a real CEO would actually look at.

The setup isolates one variable. The only difference is whether the model saw visual evidence, so when results diverge the researchers can trace exactly what the extra input did.

Here is where it gets interesting. Each visual channel helped on its own. Risk charts improved forecasting. Trend graphs improved reasoning. But pile them together, and the model’s ability to respect hard constraints fell apart. The authors call this signal crowding. Like giving someone ten clues at once: helpful individually, but together they drown out the thing you needed to hold onto.

The paradox is structural, not lazy. Visual perception and constrained action are separate bottlenecks, and fixing one can break the other.

Results and Practical Insights

Lead with the number that stopped me: every one of the nine models got worse at resource allocation once visuals were added, averaging -0.08 and deepening to -0.12 under tension. Not a fluke of one vendor. A pattern across the entire frontier.

At the same time, the same visuals delivered the largest, most reliable gains exactly where executives need defensible calls. Risk forecasting improved by +0.24. Board-facing justification lifted +0.17. The pattern is consistent: visuals help the reasoning-heavy, narrative work, and hurt the constrained-planning work.

For a Chief AI Officer, the rule writes itself: match the data to the decision. When the call is about weighing risk or building a board narrative, give the agent the charts. When it is about allocating a fixed budget or capacity under hard limits, be careful how much visual noise you add. Selective grounding, not maximal data, is the design principle.

Key Takeaways for Operations and Finance Leaders

The Evidence on the Page

Thanks to All Authors