Most people build their first ManyChat AI flow the way they’d build a vending machine. Press A, get B. Then someone types “hey do you ship to Ireland and can I swap the medium for a large on order 4412?” and the whole thing faceplants into a generic menu.
The difference between an automation that books you clients and one that quietly annoys people comes down to a handful of setup decisions. Below is the exact sequence I use when building one of these, with the numbers and message examples so you can copy the logic straight into your own account.
If you want the broader picture of what the tool does across Instagram, Messenger, and WhatsApp first, there’s a solid primer on automating Instagram, Messenger, and WhatsApp conversations that covers the platform side. This piece is the how-to that comes after it.
Step 1: Give the bot one job, not four
The single biggest failure I see is a bot that tries to be a receptionist, a sales rep, and a support desk at once. The AI step gets vague, it hedges, and it starts answering questions it has no business answering.
Pick one job. Concrete examples that work well:
- Answer pre-purchase questions about sizing, shipping, or fit so people stop leaving the DM thread.
- Qualify leads for a coaching program by asking about goals, timeline, and budget before a human ever replies.
- Re-engage everyone who commented a keyword on a Reel but never opened the DM.
- Collect the four details you need for a booking (service, date, location, name) with zero forms.
A Toronto gym I worked with ran the fourth one. One job only: get a name and a preferred class time. Nothing else. That constraint is what made it work.
Step 2: Feed the AI like it’s a new hire on day two
The AI step in ManyChat is only as good as the material behind it. You have two places to feed it: a knowledge base of your own content (FAQs, shipping policy, pricing, product specs) and the instructions you write in the flow.
Write answers the way you’d text a friend
Take your FAQ page and rewrite every answer out loud. “Returns are accepted within 30 days of delivery provided the item is unworn and in original packaging” becomes “Yep, 30 days. Keep the tags on and we’ll refund you, no drama.”
Short, warm, specific. The AI mirrors the register of the examples it’s given, so if you hand it corporate policy language, that’s what your customers get back.
Give it explicit rules and an exit
Inside your instructions, spell out four things: who it is, how it talks, what it must never do, and when to hand off. Mine usually look something like:
- You are the assistant for a small skincare brand. Friendly, brief, no exclamation marks.
- Never quote a discount. Never promise delivery dates. Never give medical advice.
- If someone mentions a rash, a reaction, or a refund over $75, stop and escalate to a human.
- Always end by asking one question to keep the conversation moving.
That last rule matters more than it looks. A bot that asks nothing gets one message and silence.
Step 3: Wire the AI into the flow properly
Here’s a layout that handles most cases without getting complicated:
- Trigger: comment or keyword (“SIZE”, “PRICING”, “TOUR”).
- One open question: “What are you trying to sort out first?” No menu. Menus kill momentum.
- AI step: answers using your knowledge base and instructions.
- Handoff: if the AI flags a human need, tag the contact, notify your team in Slack, and send one holding message.
The holding message is the part people skip. “Got it, let me grab someone who can sort this properly. Back within the hour.” Without it, a ten-minute wait feels like being ignored.
Step 4: Test it with the messy stuff, not the polite stuff
Run every one of these before you turn the flow on. They break more builds than anything else.
- A message with typos and no capital letters: “do u ship 2 ireland”
- Two questions in one message: “how much is it and does it come in black”
- A question you never wrote an answer for, like a restock date three months out
- An annoyed customer: “this is the third time I’ve asked”
- Someone fishing for a discount, to check it doesn’t invent one
- An emoji-only message, and a voice note, to see how the transcription handles accents
Fix the answers, not just the wording. If the AI keeps saying “I’m not sure” to the same question, the knowledge base is missing that page, and no amount of prompt tweaking will patch it.
Step 5: Track three numbers, not thirty
Once it’s live, ignore the vanity metrics. Watch these instead:
- Containment rate: the share of conversations the AI finishes without a human. Anything above 55% on a pre-sales flow is healthy.
- Handoff rate: how often it escalates. If this is above 40%, your instructions are too cautious or your knowledge base is thin.
- Reply rate on the opening message: how many people answer your first question. Under 30% and your trigger or opener needs work.
Review these weekly for the first month. Small edits to the instructions move containment by 5 to 15 points, which is far more than most people expect.
A worked example, start to finish
The gym flow again, with real figures from a six-week run.
A Reel with a 30-second kettlebell drill pulled 41,200 views. The comment trigger was “TOUR”. 412 people commented. 318 opened the DM, and 274 answered the opening question about their training goal.
The AI sorted them into three lanes: general questions (handled fully, 61% of the 318), schedule and pricing (answered from the knowledge base, then pushed to booking), and existing members with billing problems (escalated immediately, 19 conversations).
End result: 41 booked trials in six weeks from a single Reel, with no ad spend and roughly two hours of setup. The gym’s owner spent her time on the 19 billing issues, which is exactly where a human belongs.
Mistakes that quietly kill the results
- Letting the AI answer pricing questions with detail it half-knows. Give it a fixed line and nothing else.
- Using a menu of numbered options. People type what they want; let them.
- Forgetting the follow-up message at 24 hours for people who go quiet.
- Writing instructions once and never touching them. These flows are living documents.
- Testing on your own account only. Your phone number is a friendly test subject, not a real one.
What to do in week two
Open the transcript log and read the last 50 conversations end to end. You’re looking for two things: questions the AI answered badly, and questions it answered well that you hadn’t thought to write down.
The first list becomes new knowledge base entries. The second becomes new trigger keywords, because if enough people ask a question in the DM, they’ll comment it on a post too. Ten minutes of reading, one or two edits, and the flow gets measurably sharper each week. That compounding is where the real return sits, and it never shows up in the setup video.