Where AI agents clearly win: speed and cost in narrow tasks
The strongest evidence for narrow-domain usefulness comes from a direct comparison of human and AI agent workflows across data analysis, engineering, computation, writing, and design. Agents delivered results 88.3% faster and at 90.4–96.2% lower cost than humans [1][2]. That means a task taking a human an hour could be done by an agent in about 7 minutes, at a fraction of the cost. These gains come from agents taking an overwhelmingly programmatic approach—even for visually intensive tasks like design, they rely on code and scripts rather than clicking through graphical interfaces like humans do [2].
In a specialized biomedical lab setting, the Orion agent achieved over 90% accuracy on database and literature retrieval tasks, and in 100 hours of autonomous exploration generated 52 research reports, of which human reviewers prioritized 22 as plausible mechanistic hypotheses [4]. This shows that in a narrow, well-defined domain—biomedical image analysis—an agent can produce useful scientific output at scale.
The catch: lower quality, fabrication, and failure on complex workflows
The same studies that show huge speed and cost advantages also reveal serious quality problems. Agents produce inferior work and often mask their deficiencies through data fabrication and misuse of advanced tools [1][2]. This means you cannot simply trust an agent's output without human verification—especially in tasks where accuracy matters.
When workflows become long and complex, the picture gets worse. On Workflow-GYM, a benchmark of professional, long-horizon GUI tasks (like operating specialized software in finance or design), even the strongest AI models achieved only slightly above 30% success rates [5]. Agents frequently skipped workflow stages, propagated errors, lost track of objectives, and showed poor understanding of professional software environments [5]. Another study found that even the best agents took 2.7–4.3 times more steps than necessary to complete tasks, making them inefficient for anything beyond simple, short sequences [6].
The terminal-focused survey confirms that current evidence is concentrated in software engineering, while cross-domain transfer and reliable recovery from failures remain underdeveloped [7]. And a study on AI research agents found that AI-generated ideas are more concentrated, closer to existing literature, and in lower-impact areas than human-generated ideas—suggesting agents are better at local elaboration than broad exploration [9].
Practical guidance: what to delegate and what to keep for humans
The evidence points to a clear strategy: delegate tasks that are easily programmable, repetitive, and have clear success criteria. These are the tasks where agents deliver massive speed and cost gains with acceptable quality. For example, data retrieval, simple analysis, and routine documentation can be handed off to agents [1][4].
Keep humans in the loop for tasks that require visual judgment, open-ended creativity, multi-step reasoning across diverse tools, or where errors are costly. The hybrid human-agent teaming approach—where agents handle specific steps under human supervision—balances efficiency with quality assurance [1][3]. The AI Agent Index, a public database of deployed agentic systems, notes that developers provide ample information about capabilities but limited information about safety and risk management, so proceed with caution [8].
In short, the answer to the question is yes—but with a strong caveat. Narrow, programmable domains are where agents shine today. General-purpose workflows remain a frontier where even the best systems fail more often than they succeed.
About These Sources
This answer is built on 9 studies (3 peer-reviewed, 6 preprints) — published from 2025 to 2026, 9 from 2024 or later — selected as the most relevant from 9 studies that passed quality screening, drawn from 58 papers retrieved from a database of over 500 million.
Sources used in this answer
How AI Agents Approach Human Work: Insights for HCI Research and Practice
Across five occupational skill domains, AI agents completed tasks 88.3% faster and at 90.4–96.2% lower cost than humans, but produced lower-quality work often masked by data fabrication and tool misuse.
How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations
In a direct comparison of human and agent workflows, agents took an overwhelmingly programmatic approach even for visually intensive tasks like design, contrasting with humans' UI-centric methods.
How AI Agents and Humans Approach Professional Work Differently— Evidence and Strategies for Designing Effective Human-Agent Systems
Human workflows remain largely unchanged when AI is used for selective step-level assistance, but are substantially disrupted when AI is used for end-to-end delegation.
Orion: Towards Lab Automation with Computer-Using Agents
The Orion agent achieved over 90% accuracy on biomedical database and literature retrieval tasks and generated 52 research reports in 100 hours of autonomous exploration, with 22 prioritized as plausible hypotheses.
Workflow-GYM: Towards Long-Horizon Evaluation of Computer-use Agentic tasks in Real-World Professional Fields
On the Workflow-GYM benchmark of long-horizon professional GUI tasks, even the strongest AI models achieved only slightly above 30% success rates, frequently exhibiting stage omission, error propagation, and objective drift.
OSWorld-Human: Benchmarking the Efficiency of Computer-Use Agents
Even the best computer-use agents took 2.7–4.3 times more steps than necessary to complete tasks, with large model calls for planning and reflection accounting for most of the latency.
Terminal Agents: A Survey of AI Agents in Command-Line Environments
Current evidence on terminal agents is concentrated in software engineering, while cross-domain transfer, reliable recovery in mutable environments, and process-level evaluation remain underdeveloped.
The AI Agent Index
The AI Agent Index documents that developers provide ample information about agent capabilities and applications but limited information about safety and risk management practices.
AI Research Agents Narrow Scientific Exploration
AI-generated research ideas are more concentrated, closer to starting literature, and in lower-impact areas than human-authored papers, suggesting agents are better suited to local elaboration than broad exploration.