Skip to content

AI & Automation1 min read

AI agents struggle with professional tasks in new benchmark

SemiAnalysis reports that current AI models fail most complex tasks in its new GDPx benchmark, affecting businesses planning autonomous workflows.

by Fini MironFounder @ Balkaris
An abstract composition printed flat in deep teal, pale sand, muted slate blue and bright marigold orange: one thick line made of four straight segments turning at right angles as it crosses from the left edge down to the lower right, with a small square at each turn.
Drawn for this article on our own machine. It is not a photograph.

The-decoder reports that analytics firm SemiAnalysis has released a new benchmark called GDPx to test AI agents in simulated work environments. The test includes 206 tasks across eleven domains, such as chip design, law, and cybersecurity. According to the report, no model achieved a score higher than 12 per cent. Claude Opus 5.5 sits at the top of the leaderboard with a 12 per cent result, followed by Gemini 4 Carbon at 10.9 per cent.

Why it matters

These results suggest that current AI agents are not yet ready for autonomous professional workflows. If you plan to integrate agents into complex processes like legal review or financial modelling, you cannot rely on them to complete tasks without human oversight. The high failure rates in specific domains indicate that technical accuracy remains a significant hurdle for enterprise adoption.

  • Expect high error rates in specialised fields like law or chip design.
  • Budget for human verification to catch mistakes in spreadsheets or clinical notes.
  • Consider the cost-to-performance ratio, as cheaper models like Gemini 4 Carbon offer high utility at lower prices.

What to watch

You should monitor how providers improve their agentic tooling, as the benchmark results depend heavily on the software used to run the models. Watch for more transparent grading methods to verify these scores. For your own projects, prioritise testing specific sub-tasks rather than assuming a model can handle a full professional workflow autonomously.

Autonomous workflows will remain manual until models can execute multi-step reasoning without error. Within a year, agents will only manage professional tasks if they master domain-specific logic.

Key takeaways

  • Human oversight is mandatory for complex professional workflows.
  • Technical accuracy in specialised domains remains a primary barrier to adoption.
  • Model performance depends heavily on the software used to run them.

Frequently asked

How should we select a model for specific business tasks?

You should test individual sub-tasks rather than assuming a model can manage a full workflow. Consider the cost-to-performance ratio of models like Gemini 4 Carbon for specific utilities.

When will AI agents be capable of working without supervision?

Autonomous workflows will likely remain manual until models can execute multi-step reasoning without error. This requires mastering domain-specific logic.