How to Tell If an AI Agent Is Worth Building
A blunt way to tell the difference between an agent that will earn its keep and a demo that will die after the first meeting.
I do not start with the model. I start with the job.
A lot of agent ideas sound good because the demo is fun. The useful ones usually look more boring: a repeated task, a pile of context, a decision someone has to make, and a place where the result needs to go.
Lead qualification is a good example. The work is bigger than writing a reply. Someone has to check the company, read the form, look at CRM history, judge fit, and decide whether sales should spend time on it. That is exactly the kind of messy middle where an agent can help.
The build gets risky when nobody can review what the agent did. I want to see the sources, the score, the reasoning, and the proposed action. If the output is just a confident paragraph, I do not trust it. A good agent makes review faster. It does not ask people to stop thinking.
Failure cost matters too. Drafting a reply for approval is one thing. Sending refunds, deleting records, or making legal claims is another. The more expensive the mistake, the more boring the system around the agent needs to be: permissions, logs, approvals, and fallbacks.
Context is where many agent ideas fall apart. A lead agent without CRM data is guessing. A support agent without policy docs is guessing. A finance agent without vendor rules is guessing. The model might sound fluent, but fluency is not context.
The best project briefs are plain. When this happens, collect this context, make this decision, prepare this output, ask this person to approve it, and write the result here.
If we can say that clearly, it is worth scoping. If we cannot, the work is not ready for code yet.
Got one of these problems?