Your support inbox has a rhythm. If you sell coffee subscriptions, you already know Monday brings the “where is my bag?” messages, Wednesday brings the “can I skip next month?” requests, and every third Thursday someone asks whether the decaf is Swiss Water processed. Those aren’t mysteries. They’re repeatable questions with repeatable answers, which makes them ideal candidates for automation.
What follows is the path from a blank Kommunicate account to a live AI assistant handling real conversations, using that coffee business as the running example.
Step 1: Choose the three questions your bot will own first
New teams try to automate everything in week one and end up with a 40-node flowchart nobody wants to maintain. Start smaller. Export your last 200 tickets and tag them by intent. For the coffee store, the tally looked like this: 34% order status, 19% shipping delays, 14% subscription changes, 11% product questions, 22% everything else.
Four intents covering 78% of volume. Pick the top three with the simplest answers and leave the rest to humans.
Write the answer before you touch the builder
For “where is my order?”, the ideal reply isn’t a paragraph. It’s a tracking link plus a delivery date, delivered in under ten seconds. If you can’t write that answer cleanly on a sticky note, no flow will rescue it. Fix the answer first, then build.
Step 2: Pick the right bot type inside Kommunicate
The platform gives you three ways to power a conversation, and choosing the wrong one is the most common early mistake.
- Bot Builder — drag-and-drop flows for menus, lead capture, and anything with a fixed path. No code required. Ideal for “choose your issue” routing.
- Answer Bot — trained on your help center articles, policy PDFs, and resolved tickets so it can handle free-form questions like “do you ship to Ireland?”
- External bot — bring your own Dialogflow, Amazon Lex, Rasa, or OpenAI-backed endpoint when you need reasoning over live order data.
Most small teams launch with the first two and graduate to the third once they have real transcripts to learn from. For the bigger picture of how these pieces fit together, there’s a useful primer on Kommunicate AI’s approach to combining live chat with intelligent automation.
Write a welcome message that names options
“Hi! I’m a bot. How can I help?” gets ignored. Try: “Hi, I can check your order, pause a subscription, or pass you to Priya on the team. What do you need?” Naming real services and a real person cuts drop-off on that first message significantly. Add a two-second delay before it appears so the chat doesn’t feel like a wall slamming open.
Step 3: Build the human handoff before the fun stuff
Bots break trust fastest when a frustrated customer can’t escape them. Build the exit before the entrance.
- Trigger handoff on explicit language: “human”, “agent”, “manager”, “real person”.
- Trigger after two consecutive fallbacks, so the bot stops guessing instead of looping.
- Trigger on sentiment signals: profanity, all-caps messages, the word “refund”, or a second complaint in the same session.
- Always render a visible “Talk to a human” button. Never bury it inside a menu.
- Pass context to the agent: full transcript, the email the bot collected, and which flow the customer came from.
Set business hours too. At 11pm, “Thanks for the details. The team replies from 8am and I’ve emailed you a summary” beats silence every time.
Step 4: Connect the channels where customers already are
Start with the website widget, which is a single script tag, then add WhatsApp, Messenger, or Slack if that’s where your buyers live. Keep the same welcome flow and the same handoff rules on every channel. A WhatsApp customer who gets a thinner experience than a website visitor will notice, and they’ll say so publicly.
Give the bot a name and an avatar. “Ana from Northwind” resolves more conversations than “Bot #4”, because people match tone to identity.
Step 5: Train the AI layer on your actual documentation
Feed the Answer Bot your help center, your shipping policy, and answers pulled from resolved tickets. Then test it the way customers really type: “my bag never showed”, “did my payment go through”, “is decaf swiss water”, “cancel please”. Note every miss. Nine times out of ten the fix lives in the source document rather than the model. An article titled “Subscription Management Overview” cannot answer “how do I skip a month?” Rewrite the heading and the opening line, retrain, retest.
Step 6: Review unresolved queries every week
Block 20 minutes each Monday. Open the unresolved queries report and sort by frequency. Every entry lands in one of three buckets.
- A missing intent you can add as a flow.
- An existing intent with a confusing source doc that needs rewriting.
- Something that should never be automated, such as a billing dispute, a legal question, or a churn-risk customer.
Move the top three items out of the first bucket each week. Stores that keep this loop running typically climb from roughly 35-40% containment at launch to 65% or better within two months. Skip it and containment flatlines by week three.
Step 7: Track three numbers, not thirteen
Ignore the dashboard buffet. Watch containment rate (bot resolved without a human), bot-specific CSAT, and first response time. Keep CSAT split between bot and human conversations, because a blended average hides the case where your bot scores 4.5 while the human queue sits at 3.8, or the reverse. Review that split monthly and adjust routing, not just flows. Clean reporting across both sides is one reason teams end up choosing live chat and automation on the same platform in the first place.
Where these setups usually go sideways
Two failure patterns show up constantly. First, the bot answers correctly but slowly: three messages of “Sure! Let me help with that. One moment. Okay!” before the actual answer arrives. Strip the filler and lead with the useful part.
Second, nobody owns the bot. It gets built during a sprint, then never updated while shipping policies change and product names shift. Assign one person, even part-time, with the weekly review sitting on their calendar.
Then read your first fifty bot transcripts end to end. Not the analytics, the conversations. You’ll spot the exact message where customers get confused, and it’s almost always earlier than the dashboard suggests. That single hour of reading is worth more than any feature you could bolt on that same week.