MatIntellect

Lately I've noticed how many businesses that talk to customers want an AI receptionist, support agent or autoresponder. In other words, AI that talks to people directly. That can start to irritate customers: they don't get the answer they need, and the conversation ends badly

Start with a knowledge base

Build a knowledge base for your AI receptionist: questions, topics, prices and the other information the agent is allowed to discuss. If a service is listed but the bot doesn't have its price or conditions, it shouldn't guess. It needs a clear way to say what it doesn't know and pass the question to a member of staff. Test it with questions the knowledge base cannot answer, then change a price and ask again

Someone has to keep that information current. Who changes the price list, when do opening hours change, and how quickly does the agent get the update? If the knowledge base contains old information, a friendly tone won't fix an incorrect answer. Set out what the receptionist must not promise, too, such as the outcome of a procedure or a decision that belongs to a specialist

A person answers a message on a smartphone
Edwin Nava

Have another model check the answer

Add a separate model to check the first model's replies. Without that check, a model can make mistakes, go off topic and lead the conversation somewhere the business didn't intend. The checking model needs rules too: compare the price with the knowledge base, spot promises outside the agent's authority and check whether the customer's question was answered

Say a customer asks about price and available times. The first model drafts a reply, and the checker compares the price with the current price list. Available times need to come from the calendar, not be invented from the conversation history. Give the checker a draft with a wrong price. If the information conflicts or is missing, it's safer to pass the question to a person. The second model can make mistakes too, so important restrictions and actions need separate checks. A valid answer should not be sent back for revision forever

Make the conversation feel human

One of the most important things is a skill that helps the model communicate like a person. A skill is an instruction for an agent, with rules and examples. For a receptionist it can describe the tone, appropriate message length and how to answer a specific question without repeating a canned greeting every time

Test it in a conversation. If the customer asks for a price, answer instead of starting the questionnaire again. If they've already given their name, don't ask twice

A person reads a message while working
Anna Shvets

How it works in a clinic project

I built an AI receptionist for a clinic: its owner was spending time on the same questions on Instagram and needed help with replies and bookings. The bot answers from the clinic's documents in the owner's style. Its knowledge base covers about 114 price items and 48 services

Code supplies prices, dates and the address from the documents and calendar, while a second AI checks claims against the clinic's materials. If the documents don't contain an answer, the question goes to the owner. The bot offers available times from the calendar and sends the owner a client card. The owner confirms the booking. Tested on real questions from the clinic, the system is ready to switch on when the owner decides

Keep a person in the conversation

Complete replacement of a person only works in limited, small parts of the job. The longer a model has to conduct a complicated scripted conversation, the more ways it can go somewhere unexpected. You can't plan for everything, which is why this approach is better suited to short, narrow scenarios. Longer ones create more opportunities for mistakes

Always give customers a way to switch to a real person. Pass the conversation history to the staff member so the customer doesn't have to explain the question from the beginning. Before handing control back to the bot, check that the customer got an answer and the next step is clear. Here's more on what an AI receptionist can handle

Humanizer skill

Humanizer helps spot repetition, stiff wording and canned patterns in English text. Add the skill to your agent and ask it to review a reply without changing the facts or meaning. It helps with style, but it doesn't replace a knowledge check or a person in a difficult conversation

Share:

No comments yet

Leave a Comment

Fields marked with an asterisk (*) are required