When a company decides to 'do something with AI', the first idea is usually a chatbot on the website. It is visible and easy to demo, and it is rarely where the time goes. The time goes into the work that happens behind the website: reading requests, looking things up and copying data between systems.
Start with a week of real requests
Before choosing a model or a framework, collect a week of the work you want to automate: incoming emails, tickets, forms, whatever arrives. Group them by what the person handling them actually had to do. You will usually find that a handful of categories make up most of the volume.
- Requests that need information from one system, formatted into a reply
- Documents that need specific fields extracted and entered somewhere else
- Messages that only need routing to the right person or queue
These are good first candidates. The input is predictable, the right answer can be checked, and the cost of a mistake is small as long as a person reviews the output.
Draft, don't send
The fastest way to lose a team's trust is to let an agent act on its own before anyone has seen how it behaves. We start almost every automation as a drafting tool: the agent prepares the reply, the extraction or the routing decision, and a person approves it with one click.
An agent that drafts well for a month earns the right to send on its own for the cases where it is always right.
Every approval, edit and rejection is logged. After a few weeks you have real data on where the agent is reliable, and you can let those categories run automatically while the rest stay in review.
Ground it in your data
A general model knows nothing about your customers, your pricing or your policies. Useful automation reads from the systems you already have, such as the CRM, the order database and the knowledge base, before it writes anything. That retrieval step matters more to the quality of the output than the choice of model.
Measure the boring things
Track how many items the automation touched, how many were approved without edits, and how long a request waits before a first response. These numbers tell you whether it is working. A dashboard that nobody opens is not monitoring.