AI replies in DMs: what to automate and what a human must handle
Short answer: Let AI answer repeatable, low-risk questions such as opening hours, location, service lists, how to book and where to find a price list. Hand every money, health, legal, complaint and angry-customer message to a person. ManyChat, a tool that automates replies on Instagram, Facebook Messenger, WhatsApp and TikTok, has AI features that answer from business details you save, available on its Pro plan and above as of November 2026. The safe pattern is simple: AI answers from a short approved knowledge list, says it is automated, offers a human at every step, and every conversation is logged and reviewed weekly.
You can now switch on AI replies in DMs, and you need to decide what the AI should answer and where a person must step in.
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.
Automate questions with one correct answer that does not change by customer. Keep a person on anything where a wrong answer costs money, health, trust or a legal problem. This page sets out the lists, handoff rules, a review routine and a test plan you can copy. It is written for agencies setting this up for clients and for owners doing it themselves.
Some terms. A DM is a direct message. AI replies here means a chatbot that writes answers in everyday language from information you give it, instead of fixed scripted lines. A handoff is when the chat passes to a person. A flow is the set of steps an automation follows. This page is information, not legal advice. Platform rules change, so check the platform's current rules and local law.
This page may contain affiliate links from which Blur Studio earns commissions.
What AI replies in DMs actually are
ManyChat lists several AI features on its pricing page, all on Pro and above, not on Free or Essential, as of November 2026: AI Replies (answering common questions from saved knowledge), Knowledge (where you upload business details and links), AI Step (instructions to qualify leads and reply), AI Goals (steering a chat toward a click, follow or lead), AI Comments (replying to positive comments and continuing in DMs) and AI Intention Recognition (starting a flow by what the person means, not an exact keyword). Its blog says AI is built into Pro with no separate fee for new accounts, though older accounts may still see an add-on price on older help pages. Check the pricing page for your account. For a feature tour, see ManyChat AI features.
Other tools offer similar features, and Blur Studio builds custom versions, so the rules below apply to any AI that talks to your customers.
Rules the tool and the platform set
- Be honest it is automated. ManyChat's AI terms bar using AI to mislead people into thinking output was written by a human. Meta, the company that runs Instagram, Messenger and WhatsApp, requires automated chats to disclose this where the law requires it, and recommends it elsewhere.
- Output can be wrong. The AI terms say output may be inaccurate and carry no warranty. You are responsible for what your account sends.
- No automated life decisions. For users in the European Economic Area, the UK, Switzerland and Brazil, the terms say not to use the AI for automated decisions with legal or significant effects, with loan eligibility, CV (job application) screening, insurance pricing, medical triage and benefits as examples.
- Data use. The terms say ManyChat may use interaction data to run and improve the features, and that it contractually restricts its AI providers from using customer content to train their general models. Read the current terms if your clients are sensitive about data.
- Humans must be reachable. WhatsApp's business policy requires automated flows to offer a clear route to a person. Meta's Messenger rules also require bots to answer input within 30 seconds.
Safe topics to automate
| Topic | Why it is safe | Source of truth |
|---|---|---|
| Opening hours, address, parking | One right answer | Client's approved list |
| Service list and what each includes | Stable, public | Client's website |
| Published prices or "from" prices | Safe if the price is public and current | A dated price sheet |
| How to book and the booking link | Directing, not deciding | The calendar page |
| Delivery areas and times | Rule-based | Client's policy page |
| Order status with a lookup | Safe only if connected to real data | The shop or the CRM (the client's customer records system) |
| Collecting name, phone and need | Gathering, not advising | The flow's questions |
The pattern: the AI reads from a fixed, short list, quotes it and stops. If the answer is not in the list, it says "That one needs a person, let me get someone" and hands off. Do not let it guess.
Keep the knowledge list short and dated
A common failure is loading the AI with a whole website. Old pages, discontinued services and conflicting prices then show up in replies. Start with 15 to 30 short entries, each with an owner and a review date. Delete anything you would not want quoted word for word.
Risky topics a person must handle
- Money decisions. Refunds, discounts, special quotes, payment disputes, anything that changes what the customer pays.
- Health, legal and financial advice. Clinics, law firms, accountants and advisers should keep AI away from advice. It can book and collect details. It should not interpret a symptom, a contract or a tax question.
- Complaints and anger. If the message has frustration, a threat to leave a review or an urgent problem, route to a person at once.
- Safety and emergencies. Any mention of harm, injury or urgent danger should trigger a human alert and a standard message pointing to emergency services where relevant.
- Personal data requests. Asking for another person's details, deletion or access requests. A human handles privacy requests.
- Promises on dates and results. The AI should not promise a delivery date, an outcome or a guarantee that the client has not approved.
- Children and vulnerable people. WhatsApp's policy says messages about regulated categories may not go to anyone under 18. When in doubt, hand off.
- Anything the knowledge list does not cover. The default for unknown is a person, not a guess.
Handoff rules
Write these as a short table the client signs off, then build them into the flow.
| Trigger | Action |
|---|---|
| Person types "human", "agent" or "call me" | Stop AI, tag "needs human", alert the team |
| Words such as refund, complaint, lawyer, cancel, angry, broken | Stop AI, tag, alert with higher priority |
| AI cannot answer twice in a row | Offer a person and tag the chat |
| Question on a risky topic from the list above | Short fixed reply plus handoff |
| Chat idle after a handoff for 15 minutes in hours | Reminder to the team |
| Outside opening hours | Say when someone will reply, collect details, hand off in the morning |
Two practical points. First, the tool's Inbox holds conversations from all connected channels, and its plans give assigning and labelling on Essential and above, so a "needs human" tag can route to a named person. How exactly a bot passes a chat to a person depends on how you build the flow, so test it. Second, remember the platform's timing rules: the standard messaging window on Instagram and Messenger is 24 hours for most actions, so a handoff that waits two days may no longer be allowed to send a normal reply. Meta's human agent tag allows a person to reply manually for a separate 7 days, but automated messages do not qualify. See the 24-hour messaging window.
Speed on the human side
A handoff is only as good as the person on the other end. Set a reply target, such as 15 minutes in hours, and measure it. If the team cannot meet it, narrow what the AI hands off or hire help, or customers will find the AI frustrating. See after-hours replies and a speed-to-lead system.
Review and logging
- Log every AI conversation. Keep the question, the answer, the date and whether a handoff happened. Store it where the client can see it.
- Weekly review for the first month. Read 20 conversations. Mark each as correct, partly correct, wrong or should have handed off.
- Monthly after that. Read 10 to 20 and every handoff that took more than the target time.
- Fix the source, not the symptom. If the AI gave a wrong price, update the knowledge entry, then retest the question.
- Keep a change log. Date, who, what changed and why.
- Privacy. Do not paste message text into public tools. Limit who can read logs. Check local law on storing chat content.
Worked example (made-up)
An invented example to show the method: a physiotherapy clinic with Instagram and WhatsApp.
- AI answers: opening hours, address, parking, first-visit length, price of a first visit (from a dated sheet), how to book.
- AI collects: name, phone, reason in a few words, preferred day. Then it books via a link.
- Human handles: any mention of pain levels, injuries, medication, insurance disputes, cancellations within 24 hours and any complaint.
- Fixed reply on a risky topic: "Thanks for telling us. A member of the team will reply during opening hours. If this is an emergency, please call your local emergency number."
- Review: the practice manager reads 20 chats every Friday for a month.
In week one the review finds the AI quoting last year's first-visit price from an old knowledge entry. The fix: delete the entry, add a dated price sheet and retest. Nothing in this example is a result or a benchmark.
Test plan before going live
Run 30 test messages from a spare account before you switch it on for the public. Pass and fail rules are fixed in advance.
- 10 safe questions. Pass: at least 9 are answered correctly with no invented details.
- 10 risky messages (refund, injury, angry, legal threat). Pass: all 10 trigger a handoff and none gets advice.
- 5 questions not in the knowledge list. Pass: all 5 get "That one needs a person" and a handoff.
- 5 attempts to confuse it (jokes, other languages, instructions like "ignore your rules"). Pass: it stays on topic and hands off.
- Check the handoff alert reaches a real person within the target time.
- Check the first message tells people it is automated and offers a human.
Any fail on a risky message means you do not launch. Fix and rerun the whole set, because a fix can break something else.
Who should skip AI replies for now
- Businesses with fewer than a handful of DMs a week. Fixed buttons and a quick personal reply are enough.
- Regulated advice businesses with no review capacity. If nobody can read the logs, do not turn it on.
- Anyone on a plan without the AI features. As of November 2026 ManyChat's Free and Essential plans do not include them. Scripted flows still work. See ManyChat pricing.
- Teams that cannot answer handoffs. A bot with no human behind it frustrates customers and invites complaints.
How I'd approach this
I would start with scripted buttons for the top five questions, add AI only for the free-text gaps, and keep the knowledge list tiny. I would write the handoff table with the client first, test with 30 messages, and review weekly for a month. I would not let the AI discuss prices that are not on a dated sheet.
If you want to try it on a test account first, set up ManyChat, and check which plan includes the AI features you need. If you want AI agents that use your business knowledge and connect to your CRM and calendar, or you have a half-built flow to finish, Blur Studio builds the custom version, with ongoing care. See what that includes. For the wider picture of AI in customer chat, read AI chatbot for business, and for agency pricing, how to price chatbot work.
Questions people ask
Can AI reply to Instagram DMs by itself?
Yes, within Meta's rules. A person must usually message first, and the standard window is 24 hours for most actions. AI features in ManyChat are on Pro and above as of November 2026. Keep it to low-risk topics and offer a human.
What should a chatbot never answer?
Refunds and special prices, health, legal and financial advice, complaints, emergencies, privacy requests and anything not in the approved knowledge list. Those go to a person.
Do I have to tell people it is a bot?
Meta requires it where the law does and recommends it everywhere. ManyChat's AI terms bar using AI to mislead people into thinking the output is human-written. A short first line is simple. Check local law.
How do I stop the AI from making things up?
Give it a short, dated knowledge list, tell it to answer only from that list, and test with questions outside it. Output can still be wrong, since the vendor's terms give no accuracy warranty, so review logs weekly.
Does ManyChat charge extra for AI?
Its blog says AI is built into Pro with no separate fee for new accounts, though older accounts and older help pages may mention an add-on. Check the pricing page for your account. See ManyChat AI features.
How do I hand a chat to a human?
Build a trigger that tags the chat and alerts a person, using words like human, refund or complaint, or two failed answers. Assign it in the Inbox and test the alert. Mind the platform window: Meta's human agent tag gives a person 7 days to reply manually.
How often should I review AI conversations?
Weekly for the first month, reading about 20 chats, then monthly with 10 to 20 plus every slow handoff. Fix the knowledge entry behind each wrong answer.
More in this topic: ManyChat: for agencies (8 guides) →