A language model is an artificial intelligence system trained on large amounts of text to understand and produce natural language. It is the technology behind many of today's conversational assistants. In customer service it sounds appealing: it answers fluently, at any hour, in everyday language. But it's worth looking calmly at what it does well and where it needs limits.
What it can do well
- Understand free-form written messages. People don't always write like a form. A model can interpret incomplete sentences, typos or different wording.
- Answer with information you give it. If it connects to your FAQs, catalogs or documents, it can write clear answers based on that material.
- Summarize conversations. When a case moves to a person, it can hand over a summary of what happened so the customer doesn't have to repeat everything.
- Sort and route. Detect what a request is about and send it to the right person.
- Suggest drafts. Prepare a reply that an advisor reviews and adjusts before sending.
- Handle several languages or styles. Adapt the tone to the situation, if told how.
What it can't do, or shouldn't do alone
- Know what it hasn't been given. If it doesn't have your up-to-date information, it may answer something that sounds reasonable but is wrong. This is known as a "hallucination."
- Make decisions that require judgment or authority. Approving refunds, making exceptions or committing the company to an agreement are human decisions.
- Handle emotionally charged cases. A customer hurt by a serious problem needs real empathy and the ability to judge.
- Give professional advice. Medical, legal or financial topics require a professional. The assistant can point to whom to consult, not replace them.
- Always answer the same way. These models can vary their answers to the same question, so they need clear instructions and review.
How to set limits
- Define which sources it answers from. It should rely on your documents and say "I don't have that information" when it can't find it.
- Set which topics are out of its scope. For those cases, the assistant must hand the conversation to a person.
- Always leave a way out to a human. The customer must be able to ask to talk to someone and get that service.
- Say it's an automated assistant. Transparency builds trust.
- Review real conversations. It's the best way to find errors and improve the instructions.
Your customers' data
Conversations include personal data. Define what is stored, who sees it and what happens to that data if you use third-party services. Law 1581 of 2012 applies. You can go deeper in Law 1581 and artificial intelligence: how to protect the data your solution processes.
How to know if it fits
Ask yourself three questions. Does your team often answer the same inquiries? Do you have clear, up-to-date information an assistant can answer from? Will a person review what the assistant does? If all three answers are yes, it is probably worth trying a small case.
Start small
The sensible move is to choose one type of inquiry, test it with a small group and expand based on the results. That way you avoid promising your customers something the system doesn't yet do well.
How we support you
At ParqueSoft Crea we build AI assistants connected to your systems, with clear limits and a path to your human team, and we support you through implementation and ongoing support. To see how it fits the most-used channel, read WhatsApp chatbots for business: what can be automated and what can't.
Facing a similar challenge? Get a quote
