An AI receptionist for your inbound messages: WhatsApp, Instagram and live chat

An AI receptionist answers the people who write to your business first: a WhatsApp message, an Instagram direct message or comment, a Messenger chat, or a question typed into the live chat widget on your site. It reads the question, answers from the facts you gave it, asks for what is missing, and hands the conversation to a person when it cannot answer. In one hotel's published case, the agent's median first reply to a guest was 41 seconds, and it created no booking on its own.

Most writing about AI and messaging is about outreach, which is writing to strangers. This post is about the other half: the messages that arrive when you are asleep, in a meeting or busy with a customer. Those people already want something, and the only question is whether they get an answer while they still care.

What does a receptionist do that a chatbot does not?

The words get mixed up, so here is the working distinction.

• A chatbot follows a menu or a script. It works while the question matches a branch, and it fails with free text.
• An AI receptionist reads what the person wrote, in their words and language, answers from your documents, asks for missing details, remembers the conversation so far, and decides when to pass it to a human.
• A human receptionist has judgement and authority, and is available a third of the day.

The receptionist sits between the two. It handles the questions with a known answer, such as hours, prices from a rate card, availability, how something works, and where to go next. It leaves the ones that need judgement to a person. The longer comparison is in live chat, chatbot or AI concierge.

Which channels count?

The same agent logic can sit behind several inboxes, and each has its own habits:

WhatsApp runs through a WhatsApp account you connect, and this post makes no claim about outbound messaging there. Instagram is mostly an inbox and a replies channel today: the agent answers direct messages, comments and story mentions. The Instagram DMs for B2B post goes deeper on that channel.

What do you give it so it answers correctly?

An AI receptionist is only as good as what it can read. Before you switch it on, put these in the knowledge base:

1. The facts people ask about most. Prices or how pricing works, hours, locations, delivery times, what is included, cancellation and refund terms.
2. The rules for quoting. What to quote, what to ask for first (dates, quantity, size), and what never to promise.
3. The things it must not decide. Discounts, exceptions, legal or medical claims, complaints, anything with a contract.
4. Your voice. Two or three real replies you are proud of, in the languages you use.
5. The next step. A booking link, a form, or a phone number for a person.

The hotel case shows why the second item matters. Once the agent was told to quote as soon as it had dates and party size, the share of its replies that carried a price went from 50% to 92%. The change came from a rule, not from a better model.

What did one hotel see?

Hotel Pirin SKI & SPA in Bansko is a 68 room hotel. From 22 July 2026 a reservations agent answered guests by email, and from 24 September on Facebook Messenger and WhatsApp, in the guest's language, quoting from the hotel's live availability system or, when that check failed, from the official rate card.

The published numbers, with what each counts:

• 41 seconds median first reply to a guest, over 54 guest inquiries. Messenger was 12 seconds and email 49 seconds.
• 122 of 123 agent answers to a guest came within 2 minutes.
• Guests who wrote at 00:16 and 23:01 had an answer in 49 and 56 seconds.
• The agent checked live availability in 30 of the 54 threads.
• 8 of 63 leisure guests booked, and seven reservations followed an agent's first reply. The agent created no reservation. Four were closed by staff, the owner or the reply desk, two on the hotel website at the rate the agent quoted, and one on Booking.com.

The case study also says what the numbers do not prove. It is not a causal comparison against staff, since staff join threads that already look promising. And group requests still needed a person: staff answered them in a median of almost four hours across 18 threads. A receptionist helps most with the first minute, and the follow through is a staffing question.

Where must a person step in?

Plan the handoff before you plan the answers.

• The agent cannot answer with a fact. Better silence plus an alert than a guess.
• Money or exceptions. A discount, a refund, a group price, a custom scope.
• A complaint or an upset customer. Tone matters more than speed.
• The person asks for a human. Always honour it.
• A high value moment. The person is ready to book or buy.

In Sellinger a campaign can run in approval mode, where every message waits for a person to approve it, and a person can take over any conversation. The agent stops on its own when it cannot answer. The case study's lesson is blunt: after a guest was handed to staff, the wait ran from three hours to 14 days, so a handoff needs an alert and a clock. See the handoff on WhatsApp for the setup.

How do you start safely?

1. Pick one channel, the one where you lose the most messages.
2. Load the knowledge base with the five items above and test it with 20 real questions from your own inbox.
3. Run in approval mode for the first week and read every draft.
4. Set the handoff rule and an owner, with an alert that reaches a phone.
5. Count what matters: time to first reply, the share of conversations that reach a person, and the number of bookings or leads. Not the number of messages sent.
6. Add a second channel when the first is stable.

A checklist

• [ ] I know which channel loses me the most inbound messages.
• [ ] The knowledge base holds the facts, the quoting rules and the no go list.
• [ ] I tested it with real questions in each language I use.
• [ ] Approval mode is on for the first week.
• [ ] Every handoff alerts a named person.
• [ ] I track time to first reply and outcomes, not message volume.