How to fix chatbot flows and finish a half-built chatbot or automation
Short answer: A half-built chatbot (a program that answers customer messages on Instagram, a photo and video app; WhatsApp, a messaging app; Messenger, Facebook's chat; or your website) usually stalls for the same reasons: flows (the maps of messages and questions) that dead-end, no hand-off to a person, leads that go nowhere, accounts owned by someone who left, and no tests. To finish it, list every path, fix dead ends and the hand-off, connect the lead data, confirm who owns the accounts, then test with a second account before you switch it on. If more than half the flows need redoing, rebuilding is often cheaper than patching.
Someone started the chatbot, then the project stalled. It half works, nobody trusts it, and you are not sure whether to fix it or start again.
I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.
Most half-built chatbots are not broken in a mysterious way. They are unfinished in predictable ways. The fastest route is to inventory what exists, mark each part keep, fix or bin, and finish the plumbing that was skipped: the person hand-off, the lead data, the tests and the ownership. This guide gives you the problem list, a checklist, a decision rule for repair or rebuild, and what to ask if you bring in help.
The examples use tools such as ManyChat, a service that runs automatic replies in Instagram, Facebook Messenger, WhatsApp, TikTok and other apps, and Meta, the company that owns Instagram, Facebook and WhatsApp and sets the rules for them. The method fits any chat tool. Platform rules and prices change, so as of November 2026 check each vendor's current pages.
The common problems in a half-built chatbot
| Problem | What it looks like | Why it hurts |
|---|---|---|
| Dead ends | A button or answer leads to nothing, or to a message that says nothing useful | The customer waits for a reply that never comes |
| No hand-off to a person | The bot keeps answering when someone says "I need a human" | Meta's rules expect a clear route to a person, and customers leave |
| Leads go nowhere | Names and numbers sit inside the chat tool, not in your CRM (customer relationship management tool, where you keep contacts and deals) or calendar | Someone must copy them by hand, so some are missed |
| Triggers that overlap or never fire | Two automations answer the same comment, or none does | Double replies or silence |
| Test content left in | Placeholder text, a test link, "lorem ipsum" in a live flow | Looks careless to customers |
| Wrong owner | The account was connected by a freelancer, a former staff member or an agency | You lose access when they leave |
| Plan and limits ignored | Free plan live-automation cap, contact limit, empty wallet (the prepaid balance some tools use for WhatsApp and email) | Parts of the bot switch off or bills surprise you |
| No consent record | You cannot show when someone agreed to be messaged | Risk with the platform and with local law |
| Nobody owns it | No one checks it weekly | Silent failures stay for months |
Which of these apply to you? If three or more rows match, the project probably needs a finish plan, not another round of small edits.
Step 1: inventory what exists
Before you change anything, write down what is live. A one page note is enough:
- Which channels are connected (Instagram, Facebook Messenger, WhatsApp, website, TikTok) and who is the admin of each.
- Every automation, its trigger (what starts it) and whether it is live.
- Where lead data goes: a spreadsheet, a CRM, an email, or nowhere.
- Which paid plan the account is on and what it includes.
- Any connected tools, such as a connector like Zapier or Make that passes data between apps.
Screenshot the flows before you edit. If you can duplicate a flow before changing it, do that. In ManyChat the status page and the plan page also tell you quickly if the problem is on the platform side. ManyChat's status page currently lists an open Meta-side incident for the Follow to DM trigger (the welcome message sent to new followers), so a welcome message that misfires may not be your build. See the not-triggering fix checklist for the order of checks.
Step 2: confirm who owns the accounts
This is the problem that hurts most when it is missed. If an agency or freelancer built your chatbot, check that you, the business owner, are listed as an administrator on the chat tool and on the connected Instagram, Facebook and WhatsApp accounts. ManyChat's terms say an agency must identify the business owner as administrator on the account and on the connected pages, and that the agency keeps the rights to the flows, templates and sequences it built. So the content can be theirs even when the account is yours. Ask for an export or a written copy of the flows if the relationship ends. For contract wording, see contracts for chatbot work: platform risk and ownership.
Step 3: map every path and fix dead ends
Draw the conversation on paper. Start at each trigger and follow every button, keyword and answer to the end. Mark each end as one of four: a useful answer, a booking or lead capture, a hand-off to a person, or a dead end. Fix every dead end. Then check these rules:
- A way out. Every path offers a route to a human. Meta's policy asks automated chats to give a clear route to human support, and says that where the law requires it, a bot should say it is an automated service. Check local law and Meta's current rules.
- A default reply. If the bot does not understand, it says so and offers a short menu, not silence.
- Reply speed. Meta's Messenger policy says automated bots must respond to input within 30 seconds, and non-compliant bots get a period to fix it before messaging may be limited. A flow that waits on a slow step can fall foul of this.
- The 24 hour window. On Messenger and Instagram, standard replies fit inside a window that lasts 24 hours for most user actions. Follow-ups after that need an approved route. See what the 24 hour messaging window is.
- Comment-to-DM limits. Comment-to-DM sends a private message to people who comment on a post. Meta's private reply rule gives one message per commenter within 7 days.
Step 4: finish the lead data
A bot that collects names and then leaves them inside the chat tool is half done. Decide the fields you need (name, phone, email, need, source, consent), store them in the chat tool, and pass them on. The options are a direct integration, a connector, a spreadsheet bridge or a custom API call (a direct software-to-software connection). The details and a field map are in send chat leads to your CRM and calendar automatically. Whatever you pick, compare daily counts of leads in the chat tool and leads in the CRM. If they differ, something broke.
The 12-point finish checklist
- You are an administrator on every connected account and on the chat tool.
- Every automation is listed with its trigger, owner and purpose.
- No two automations answer the same trigger.
- Every path ends in an answer, a booking, a lead capture or a person.
- Every path offers a way to reach a human.
- The bot says what it is, where law or platform rules ask for it.
- All placeholder text and test links are gone.
- Lead fields reach the CRM or sheet, and a person is alerted.
- The plan includes every feature in use, and limits and wallet are checked.
- Consent is recorded with a date where you message for marketing.
- The test plan below has passed on a second account.
- One named person reviews it weekly, and a note says how to switch it off.
Test plan with pass and fail rules
- Every trigger. From a second account, fire each trigger once. Pass if the first message arrives within about 30 seconds. Fail if it is late or silent.
- Every button. Tap every button. Pass if none leads to a dead end.
- Odd input. Send emojis, a voice note, a photo and a random word. Pass if the bot handles each without looping.
- Human request. Type "talk to a person". Pass if a person is alerted and the bot stops.
- Lead data. Complete a lead path. Pass if one correct record appears in your CRM or sheet.
- Time test. Message after hours. Pass if the reply says when a person will answer.
- Limits. Check the count of active people against your plan. Pass if you are inside the limit or you set a cap.
Repair, rebuild or hire help: decision rules
Repair it yourself if fewer than a third of the flows have problems, the structure is clear, you are the admin, and the fixes are about wording, dead ends and a few connections.
Rebuild if more than half the flows need redoing, nobody can explain the structure, there are many overlapping triggers, or the original design never matched the business. A fresh, smaller build of the three jobs that matter often beats patching a large one. Start with the three jobs: answer the top questions, capture the lead, book the call.
Bring in help if you need CRM or calendar connections, AI replies that use your own business information, several channels, or ongoing monitoring, and you do not have time to build it. Ask any helper for: a written list of what they will finish, who owns the accounts and the flows at the end, a test report, and how fixes after launch work. For a cost guide in a related area, see how much it costs to finish a half-built AI app and whether AI can finish it or you need a developer.
When a custom build is the better fit than patching a chat tool, see AI chatbot development services and AI chatbot for business.
Who should skip a full finish project
- If the bot gets a handful of chats a week, a simple away message and a clear contact line may beat a big flow. Meta Business Suite, Meta's inbox tool, offers instant reply, away message and FAQ automations. See ManyChat vs Meta Business Suite automated replies.
- If the account is at risk of a restriction, fix the behaviour first. ManyChat's help center says it cannot reverse Meta's bans or restrictions, and names too many automated messages too quickly and aggressive copy-paste comment automation as common triggers. No official safe limits are published, so I would not trust any number you see in a blog.
- If your real problem is slow human follow-up, a bot will not fix it. Start with a lead rule and an owner.
How I'd approach this
I would not add features to an unfinished bot. I would cut it down to the three jobs that matter, finish those so they work end to end, test them from another account, and only then add more. Get ownership and the lead hand-off right first, since they are the two things that are expensive to fix later. Put a short weekly check in someone's diary, because a chatbot that nobody watches goes stale.
If you want help finishing it, Blur Studio finishes half-built flows and builds the custom version: AI agents that use your business knowledge, CRM and calendar connections, and ongoing care. You can read what that includes here. If you only need a second opinion on whether to repair or rebuild, the inventory note from Step 1 is the best thing to bring.
Questions people ask
How do I know if my chatbot is worth finishing?
Count the flows that need redoing. If fewer than a third do, and you are the admin on every account, repair it. If more than half do, a smaller rebuild is usually cleaner.
Why does my chatbot work in tests but not for customers?
Often the cause is a trigger rule, such as one private reply per commenter or the 24 hour window, or two automations clashing. See the fix checklist.
Who should own the accounts and the flows?
You, the business owner, should be an administrator on the chat tool and on every connected account. ManyChat's terms say an agency keeps rights in the flows it built, so agree in writing what you get at the end.
Does the chatbot need a way to reach a human?
Meta's policy asks automated chats to offer a clear route to a person, and to say it is automated where the law requires. Check local law and Meta's current rules.
Can AI finish a half-built chatbot for me?
It can help draft answers and review flows, but it cannot fix ownership, connections or consent. See whether AI can finish it or you need a developer.
How long does it take to finish a small chatbot?
It depends on the number of flows and connections, so I would not give a fixed time. Start with an inventory and ask any helper for a written list of what will be finished.
What should I test before switching it back on?
Every trigger, every button, odd input, a human request, a full lead path to your CRM, an after-hours message and your plan limits, all from a second account.
More in this topic: ManyChat: for small business owners (8 guides) →