Clinics, salons, law offices, home services and real estate — the question comes at 9 pm and nobody answers until morning.
Ten to five hundred people, a queue full of the same questions and a cost per contact that keeps rising.
Visitors ask about price and availability; you want budget, timing and a booked call, not a form.
Opening hours, prices, delivery, parking — the same twenty answers all day on chat, email and WhatsApp.
Numbers from Gartner, Genesys, Deloitte and Gorgias — chatbots fail on trust and accuracy, not on technology.
of customers give a virtual agent no more than three attempts before giving up.
of support contacts are fully resolved in self-service — everything else still reaches a person.
of customers say a route to a human must be available when generative AI handles service.
average cost of an agent-assisted contact, up from $6.70 in 2024.
of service leaders report a positive financial return from AI so far.
median share of conversations resolved by AI on the Gorgias platform — roughly half still need an agent.
Put your monthly volume into the calculator and see hours and dollars.
An AI chatbot for business is a workflow that answers customer messages from your own knowledge — FAQ, price list, policies, order and booking systems — around the clock, and hands anything it cannot answer to a person with the full history. Most of what reaches a support inbox is the same twenty questions, a status check or a request for a slot. That is the part chatbot automation takes over; the judgement calls stay with your team.
We build it around the tools you already use rather than a chatbot platform you have to move into. A message from the site widget, WhatsApp, Instagram or Messenger is classified by a language model, matched against your documents in a vector store and answered only from what was found. Order status comes from Shopify or your ERP after the customer is verified; a booking is made in Calendly or Google Calendar; a qualified lead lands in HubSpot or Pipedrive with budget and timing filled in. A request for a human, a complaint or a low-confidence answer opens a ticket in Zendesk, Intercom or Gorgias instead of a reply.
Questions get answered at 9 pm instead of 9 am, leads are qualified while they are still interested, and the team spends its time on refunds, exceptions and sales rather than opening hours. Every night the workflow reports which questions it could not answer, so the knowledge base keeps improving and the same gap is not missed twice. Chatbots automation done this way is measured in resolved conversations and booked calls, not in deflected customers.
The run path of one conversation. The model may only answer from your documents and your systems; anything it is unsure about goes to a person with the transcript.
Fires on a message from the site widget, WhatsApp Business API, Instagram or Messenger; a missed call can start the same run with an SMS.
Looks the sender up by phone or email in your CRM or Shopify, so the bot knows whether it is talking to a lead, a customer or a patient.
Reads the message and labels it: question, order status, booking, lead, complaint or out of scope. Sentiment is scored at the same time.
Searches your FAQ, prices and policies in the vector store (Pinecone or pgvector) and keeps only the passages that match the question.
Writes the reply strictly from the retrieved passages and the customer record, in your tone. No source, no answer.
Low confidence, medical, legal or financial advice, refunds and anything about pricing exceptions do not go out — they go to a person.
For “where is my order?” pulls the live status from Shopify, your ERP or the carrier after the customer is verified.
Offers free times from Calendly, Google Calendar, Acuity or Square Appointments, books the one chosen and sends the confirmation.
For a qualified lead, writes budget, timing and need into HubSpot, Pipedrive or GoHighLevel and pings the owner in Slack or SMS.
A request for a person, a complaint or negative sentiment ends the automated part immediately.
Opens the ticket in Zendesk, Intercom, Gorgias or Front with the full history, so the customer never repeats themselves.
Every night walks through the day’s unresolved threads and groups them by topic.
Posts the gaps in your knowledge base to Slack, so the content gets fixed and the same question is not missed twice.
Typical minutes per conversation for an administrator or support agent. Your own numbers go into the calculator below.
Based on 9.5 min by hand and 1 min with the flow per item, from the table above.
The bot answers, looks up, qualifies and books. People keep the exceptions, the promises and the review of what the bot says.
The bot explains the policy; a person decides the exception.
Negative sentiment or a request for a human ends the automated part at once.
Out of scope by design — the bot books a consultation instead of answering.
A weekly sample of conversations is read by your team and the knowledge base is corrected.
The bot quotes only what is in your documents; anything else is confirmed by a person first.
The bot collects budget and timing; the sales conversation is yours.
Flip a switch to hand a step to the flow or take it back.
Benchmarks from Gartner, Salesforce and Gorgias platform data as compiled by Macha: the best deployments resolve about half of conversations on their own and hand the rest to a person cleanly.
We map the process as it runs today, count the minutes and agree what the flow must never do on its own.
A working flow on your real data, in a sandbox. You see every run and every exception.
Edge cases, approvals and alerts, then the switch-over — with the old way kept as a fallback.
Monitoring, fixes when a vendor changes a format, and a monthly report of hours saved.
Every answer is written from passages retrieved from your own FAQ, price list and policies, and the model is instructed to refuse when it finds no source. Low-confidence answers go to a person instead of the customer. The Air Canada ruling, where a tribunal held the airline to its chatbot’s wrong answer, is exactly the case this design prevents.
It depends on how repetitive your questions are and how good your documents are. Platform data puts the median AI resolution rate around 45%; the rest should reach a person quickly, which is why the hand-off is built in from day one.
On request, on negative sentiment or when the bot is unsure, the full transcript is opened as a ticket in Zendesk, Intercom, Gorgias or Front, or posted to Slack. The customer never repeats themselves.
Yes. One workflow serves all channels through the WhatsApp Business API, Instagram and Messenger APIs and a site widget, with the same knowledge base and the same rules.
Model usage and messaging fees, which scale with conversation volume and length, plus hosting on your own n8n instance. There is no per-resolution pricing — put your volume into the calculator above to see the hours side.
The workflow runs in your cloud, conversations are stored in your systems and API calls to the model provider are made under terms that exclude training on your data. The model can also run locally so conversations never leave your network.
The tone, greeting and limits are written with you and tested on real past conversations before launch. Because answers come from your own documents, the bot says what your team would say — only faster.
Describe it in two sentences — we reply within a day with a workflow sketch.
hello@wireclad.com →