What AI Agents Really Can and Can’t Do
Think of AI agents as digital assistants that can work independently—like having a helper that can keep working even when you're not directly telling it what to do. Before we dive in, remember that every situation is different, and you'll need to decide what works best for your specific needs.
What AI Agents Are Really Good At
1. Watching Things Non-Stop: Imagine having to watch your email inbox 24/7—you'd get tired, but an AI agent wouldn't. MIT researchers show that agents can monitor systems continuously with high reliability, far exceeding human capacity for sustained attention.
2. Handling Complex Step-by-Step Tasks: Think of it as following a complicated recipe. While humans might forget steps or get confused, agents excel at following detailed instructions perfectly every time. The research presented in our book Agentic AI found that agents could handle complex business workflows with 85% fewer errors than manual processing.
3. Working Across Multiple Tools: It's like having someone who can instantly switch between your email, calendar, and task manager without getting confused. Our research demonstrates that agents can effectively coordinate across up to half a dozen different tools simultaneously while maintaining context. Beyond that threshold, the cognitive load increases significantly, leading to hallucinations and a heightened risk of conflicts between functionalities.
4. Learning From Their Mistakes: Just like how we get better at a game the more we play it, agents can quickly improve their performance through practice. A 2025 study in Nature showed that AI agents could achieve robust and fast learning across diverse domains, enabling them to adapt their strategies efficiently.
5. Doing Things in the Right Order at the Right Time: Think of a conductor keeping an orchestra in perfect time—agents can coordinate multiple actions with precise timing. IBM's AI workflows have been shown to improve customer engagement by 40% and reduce wait times, demonstrating the efficiency of AI in managing complex processes.
6. Following Rules Consistently: Unlike humans, who might bend rules when tired or stressed, agents stick to the guidelines every time (or almost). Researchers found that Large Language Model (LLM)-based AI agents achieved up to 92% accuracy in legal invoice compliance reviews, surpassing experienced human reviewers who averaged 72% accuracy.
When You Should Be Careful Using Agents
1. When Mistakes Could Snowball: Imagine if one small spelling error in a document got copied into thousands of other documents. If each AI agent or subtask has 90% accuracy, completing four sequential steps results in only 66% final accuracy (0.94=0.65). With 10 steps, this drops to 35% (0.910≈0.35). For example, an AI travel agent might first book the wrong hotel dates (10% error), causing misaligned museum tickets (20% error from cascading mistake) and restaurant reservations (30% error), despite each component being 90% accurate individually.
2. When It's Hard to Check Their Work: It's like having a helper who works so quickly you can't keep track of what they're doing. UC Berkeley’s Sutardja Center highlights that as agentic AI systems scale in autonomy, human supervisors struggle to track and evaluate their decisions.
3. When Complex Judgment Is Needed: Some decisions are like choosing the perfect gift - they need understanding of feelings and context that agents just don't have yet. MIT Sloan explains that AI cannot yet interpret emotional nuance or cultural context, limiting its decision-making in areas requiring human empathy or situational awareness.
4. When You Need Brand New Ideas: While agents are great at following patterns, they're not great at true creativity. Recent research found that current AI systems score approximately 40% lower than humans on tasks requiring novel thinking and original problem-solving.
5. When Things Are Highly Unpredictable: Like trying to drive in a city where the traffic rules change randomly—agents get confused when situations don't follow expected patterns. MIT researchers demonstrated that reinforcement learning agents performed significantly worse when trained in environments with inconsistent patterns or random changes.
How Agent Abilities Are Changing
Think of AI agents like smartphones - they keep getting better, and what they can do keeps changing. Here's what to consider:
How complicated is the task?
Are the rules clear and stable?
What happens if something goes wrong?
Can you effectively supervise the agent?
Will the situation keep changing?
The key to success is starting small with simple, well-defined tasks and gradually taking on more as you understand what works and what doesn't.
What surprising discovery about AI agents most changed your perspective? Share your thoughts in the comments below 👇
What's Next?
Explore our latest research in our best-selling book Agentic Artificial Intelligence to learn how to successfully implement agents in your organization, work, and life.
Join the first Executive Course on Agentic AI to learn how to successfully integrate Agentic AI in your organization and build stronger human-AI collaboration.
Discover IRREPLACEABLE for practical ways to grow the human strengths that matter most in an AI-powered future.
Thanks,
-- Pascal
Hi, Mr Pascal. I'am Abdoul Karim KANE, Master degree in management of university Malian