A chatbot is not an agent
There is a difference worth drawing from the start. A chatbot answers from a script: if the customer says A, it replies B. An AI agent understands natural language, queries the company's systems to answer with real data, and decides what to do with each case. The first one attends; the second one resolves.
Most disappointments with AI in customer service come from installing the first while expecting the second.
Connected to the data, or useless
An agent that is not connected to the management system can only speak in generalities. "When does my order arrive?" has a single useful answer, and it lives in your database, not in the model. That is why, in every project we do, integration with the company's system is the core of the work, not an add-on. The language model is the interface; the information is yours.
This changes what the agent can do: check the status of a delivery, verify an account, take an order with the right products and prices for that customer, log a complaint with all the data the relevant area needs.
Handing off to a person is part of the design
The best agent knows when to stay quiet. Some conversations must go to a human: a sensitive complaint, a negotiation, an angry customer. We design that handoff from the start, with clear criteria and the full context of the conversation delivered to the person who takes it over. A customer should never have to repeat what they already explained.
The measure of a good agent is not how many conversations it resolves alone, but how many it resolves well and how many it knows to hand off in time.
What it solves well
- Repetitive queries with an answer in the system: order status, opening hours, coverage, list prices, balances.
- Simple order taking for registered customers, with product and quantity validation.
- Classification and routing: understanding what the message is about and sending it to the right area with a summary.
- After-hours service, without promising what nobody will be able to deliver until the next day.
Common mistakes
- Letting the model make things up. An agent without clear rules about what it does not know ends up promising discounts or deadlines that do not exist. The instructions must say what to do when there is no answer.
- Launching to every customer on day one. First a small group, review the real conversations, adjust, and only then open it up.
- Not measuring. How many conversations it resolved alone, how many it handed off, how many ended badly. Without those numbers there is no way to improve it.
- Forgetting it is just another channel. Whatever the agent records has to show up in the same system the rest of the company uses.
Where to start
With the three or four questions that repeat most on your WhatsApp today. That defines the first scope, which data needs to be connected and how success is measured. At a water distribution company we work with, the agent started by handling customer queries and routing cases; today it is part of an ecosystem that includes delivery, plant and online store. It did not get there in one step, and that is the most important lesson.
