In 2019, I built an internal chatbot. Back then, customers wanted more accurate answers. Nobody asked a machine to do the work on its own. That sounded more like science fiction than a product requirement.
The question has changed. People now ask how much they can hand over to an agent. We have moved from what a machine knows, to what it can do, and finally to what we can trust it to do.
Yet projects do not improve as quickly as the AI does. Agents usually fail somewhere other than the model. They cannot reach the data they need, lack permission to operate business systems, or have no way to reverse a bad action. In the end, someone still has to own the outcome.
Accuracy was not the hardest problem with that old chatbot either. The real problem was that it answered confidently when it was unsure. We eventually made it stop answering below a confidence threshold and show relevant documents instead. The answer rate fell, but trust rose. Users forgive a machine for not knowing. They do not forgive it for pretending to know.
The stakes are higher with agents. Giving a wrong answer is one thing. Sending an email or changing data based on a wrong judgment is another.
Adopting an agent is not a project to add another chat window. It is a decision to delegate part of a company's work to software. Before choosing a model, we need to define the job, the data it may see, the actions it may take, and the moment it must hand control back to a person.
Whether to adopt agents will soon stop being the interesting question. Most organizations eventually will. The question that lasts longer is this: have we already defined what it means to do our work well?