New to this? Read it in simple words
- AI agents are becoming part of daily work. The evidence shows they help with parts of jobs, not whole jobs.
- A demo usually shows the successful path. Real teams also pay for retries, approval steps and hidden errors.
- Agents can save time, but checking and fixing their work still costs a lot. Companies need stable costs and a way to recover when a plan fails.
- AI agent
- An AI that takes steps on its own to finish a task, such as booking or coding.
- Demo
- A short example that shows what a product can do.
AI agents are becoming part of daily work. The evidence shows they help with parts of jobs, not whole jobs.
Agents help with parts of jobs
Anthropic’s Economic Index studies millions of AI conversations. It looks at the tasks people do with AI. This gives a clearer picture than claims about fully automatic digital workers.
An agent may write code, summarize a document, or check a plan. It may still be too unreliable to manage the full process. People must review difficult cases, control access, and fix errors.
Some tasks are easier to automate than others. A result is easier to trust when the input is clear and errors can be reversed. Work that needs personal judgment or has serious effects still needs closer human control.
This is why task-level data is useful. It shows where AI saves time and where a person must stay responsible. Good tools make the handover between the agent and the person easy to understand.
Sources1
The cost of an agent includes tool calls, retries, human review, and recovery after errors, not only AI use.
Errors are part of the cost
A demo usually shows the successful path. A real team pays for retries, tool calls, approval steps, and hidden errors. A small error rate can become a large problem when a task has many connected steps.
The business question is not whether an agent can finish one task once. The question is whether it can finish enough supervised tasks at a stable and useful cost.
Human attention is also a cost. A reviewer may need to open every source and check every change. Clear action records, limited permissions, and short evidence notes can make this review faster.
Not every error has the same risk. An error in a draft is easy to reverse. An error in a payment or customer message may be serious. Approval rules should match the possible harm.
Sources2
A better business test
Strong agent products should show when they are unsure. They should save checkpoints and make it cheap for a person to take control. This may matter more than full automation.
Companies need clear ownership, stable costs, and a way to recover when a plan fails. They also need records that show what the agent changed and why.
A useful measure is trusted work divided by total cost. Total cost includes model use, review time, maintenance, and recovery after an error. Agents create real value when all these costs fall together.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
The business-cost analysis is our own conclusion, not a claim by Anthropic.
The story questions whether companies can profit from AI agents, because errors, retries and human checks can make results expensive. How we rate the mood