What Agents Cannot Do

I talk about what AI agents can do a lot. Let me be honest about what they cannot do. Knowing the limits is what keeps you from wasting time and money on the wrong problems.

They Cannot Replace Judgment

Agents execute. They do not decide. They can research options, present tradeoffs, and recommend paths. But the decision is yours. If you give an agent full autonomy to make business decisions, you will eventually regret it.

They Cannot Handle Novel Situations

Agents are great at structured, repeatable tasks. They are bad at handling things they have never seen before. If a situation falls outside the playbook, a good agent stops and asks. A poorly built one improvises, and that is when things break.

They Hallucinate

This has gotten better but it has not gone away. If an agent cannot find real information, it will sometimes make something up that sounds plausible. This is why verification steps in your prompts are mandatory, not optional.

They Need Maintenance

APIs change. Models update. The data your agent relies on shifts. An agent you built six months ago might need prompt adjustments today. They are not set and forget forever. They are more like a car. Reliable, but they need oil changes.

They Are Not Cheap for Everything

Running agents costs money. API calls, compute time, tokens. For high-volume tasks the cost adds up. Make sure the automation actually saves you more than it costs. Sometimes the answer is “this task is not worth automating yet.”

What They Are Great At

Research and data gathering. Repetitive text processing. Monitoring and alerting. Draft generation. Code execution with verification. Pipeline automation. These are the sweet spots. Build there first.