Manychat

Flows write into the marketing stack through named connectors. Contacts collected in DMs go to HubSpot, Klaviyo, Mailchimp, ActiveCampaign, Kit and Flodesk, chat data logs to Google Sheets, and Hotmart sells courses inside Instagram and WhatsApp conversations. Zapier reaches more than 7,000 apps and Make builds scenarios across them.

The External Request action calls any HTTPS API with GET, POST, PUT or DELETE, custom headers and a JSON body, and maps the response into contact fields. Developers can also package an API as a Manychat Application with its own auth, actions and option sources. The AI reaches these systems through the flow it sits in, rather than choosing a tool itself in the middle of a conversation.

Sourcemanychat.com/integrationsread 2026-10-09

Flow Builder chains triggers, messages, conditions, delays, actions and other flows on one canvas, with a Randomizer that splits contacts across up to 12 weighted paths, sequences that send over days, and rules that run actions on events.

An AI Step drops into any flow: the brand states a goal and background, the AI turns them into tasks, carries the conversation and saves answers into fields that later steps branch on.

On Instagram, up to three AI Goals follow an AI reply in a priority order the brand sets, and a lead capture ends by sending a message, starting another automation or assigning the conversation to a person.

Sourcehelp.manychat.com/hc/en-us/articles/14281187288860-manychat-ai-stepread 2026-10-09

A Knowledge library feeds AI Replies. Brands add text entries of up to 250,000 characters each and public web pages, scanned one page at a time, shown for review and editing, and active only once saved. Link entries can be refreshed when a page changes, entries can be updated or deleted at any time, and edits apply from the next conversation.

In the AI Playground every test answer lists the entries it drew on. Only visible page text is extracted, so images, video and PDFs are skipped, social profiles and Google Docs are refused, and file uploads are marked coming soon. The AI Step instead works from up to 10,000 characters of context written into the step.

Sourcehelp.manychat.com/hc/en-us/articles/25626595060124-manychat-ai-knowledgeread 2026-10-09

AI Behavior sets the AI's role, brand voice and explicit guardrails, the topics it should never touch and what it should never say or do, starting from a draft written off the account's bio and recent comments. Before AI Replies goes live, the brand tests it in the AI Playground and picks who takes conversations the AI cannot answer, and Assign to a person routes those to a team member or group.

Each skill has its own toggle, only Admins and Editors can change Manychat AI, and staff can pause automation for a contact from 30 minutes to indefinitely and resume it from the Inbox. AI Comments is tuned by approving or rejecting five sample replies. The product page says brands approve every AI comment reply, while the help center shows approval only at setup.

Sourcecommunity.manychat.com/product-updates/your-ai-now-speaks-your-language-meet-ai-behavior-9491read 2026-10-09

Accounts have five roles with set permissions: Owner, Admin, Editor, Inbox Agent and Viewer. Inbox Agents work conversations, tags and contact fields without touching automations, Viewers see statistics read only, invitation links are single use and expire in 24 hours, and only Admins and Editors reach Manychat AI.

Manychat holds ISO/IEC 27001:2022 certification from BSI, a SOC 2 Type II report from A-LIGN available on request, ISO/IEC 42001:2023 for AI management, a CSA STAR listing and PCI DSS compliance under SAQ A for its payment integrations. Data is encrypted in transit with TLS 1.2 and 1.3 and at rest to FIPS 140-2, penetration tests run yearly and a private bug bounty runs on Bugcrowd. Single sign on, SCIM and two factor sign in do not appear.

Sourcemanychat.com/securityread 2026-10-09

Every message the AI sends sits in the contact's Inbox conversation beside the automations that ran, and an Overview tab in Manychat AI reports on AI performance. In the AI Playground each test answer shows the Knowledge entries behind it, and testers rate answers up or down with a reason. Inbox Analytics tracks response times, team efficiency and conversation volume. Live conversations do not carry that source view, and there is no log of agent decisions, tool calls or configuration changes for tracing a single reply after the fact.

Sourcecommunity.manychat.com/product-updates/better-faster-stronger-introducing-new-updates-to-ai-playground-9311read 2026-10-09

Inside a conversation the AI keeps the thread, adapting to what the follower says over several turns and waiting up to 10 hours for a reply before it stops. Answers it gathers, such as an email or a budget, are saved to system or custom contact fields that later flows and the Inbox read, so they belong to the contact record rather than to a memory the agent keeps.

AI Comments studies the account's past comment replies to match its tone, and AI Behavior holds the persona and guardrails as settings. Nothing carries what the agent learned about a person from one conversation to the next with its own scope, lifetime or deletion control.

Sourcehelp.manychat.com/hc/en-us/articles/14281227789468-power-up-your-chat-marketing-with-manychat-airead 2026-10-09

Manychat runs as a hosted service only, and all production data sits in AWS data centers in Frankfurt, with no other region on offer and no private cloud or on premises install.

AI features run on generative AI providers OpenAI, Anthropic, Google and Microsoft, all listed as processing in the US, and transfers from the EU, UK and Switzerland rest on the Data Privacy Framework and standard contractual clauses.

Full daily database backups are kept apart from the main data center, accounts inactive for 18 months are flagged for deletion, and personal data leaves active systems when an account is deleted.

Sourcemanychat.com/securityread 2026-10-09

Quick Automations are ready made setups a brand picks and fills in step by step, among them auto DM links from comments, generate leads with stories, grow followers from comments and a delayed welcome to new followers. Templates install whole automations, fields and tags from a link, creators and agencies build and protect their own, and Pro templates need a paid plan.

Manychat AI ships as separate skills: AI Replies answers questions, AI Comments answers positive comments, and AI Goals extends either by sharing a link, growing followers or capturing leads. The Flow Builder assistant drafts a whole automation from a plain description of the goal.

Sourcehelp.manychat.com/hc/en-us/articles/16654065283100-quick-automation-auto-dm-links-from-commentsread 2026-10-09

Inbound social activity starts the work. Comments on posts and reels, DMs matching a keyword or a described intention, story replies and mentions, new followers, clicks on Instagram, Facebook and WhatsApp ads, Facebook Shop messages, QR codes and ref links each open an automation, and AI Replies and AI Comments answer as messages arrive, around the clock.

Sequences and broadcasts run on schedules, and outside apps start flows by posting to Manychat's webhook endpoint. Channels span Instagram, TikTok, Messenger, WhatsApp, Telegram, SMS and email. AI Replies, Comments and Goals run on Instagram only, the AI Step on Messenger, Instagram, WhatsApp, TikTok and Telegram, and intention recognition on all of those but TikTok.

Sourcehelp.manychat.com/hc/en-us/articles/23018283889180-manychat-ai-repliesread 2026-10-09

The model behind Manychat AI is not named, and brands cannot pick one for AI Replies, Comments, Goals or the AI Step. The subprocessor list names OpenAI, Anthropic, Google and Microsoft for generative AI, and the AI terms cite OpenAI's policies and the Azure OpenAI code of conduct. Separately, flows can call ChatGPT, Claude or DeepSeek on the brand's own API key, choosing the model, prompt, temperature and token limit and saving the answer to a field. Those calls are single generations from the latest message, billed by the provider, and sit outside the Manychat AI skills.

Sourcemanychat.com/legal/service-providersread 2026-10-09

A public REST API at api.manychat.com, with a Swagger console and Bearer token auth, covers the account (tags, custom and bot fields, flows, growth tools), sending (content and whole flows) and contacts (create, update, find, tag and set fields). Reads run at up to 100 requests a second, sends at 25 and writes at 10, and an account over the limit may be cut off for 24 hours.

Keys are generated under Settings on paid plans. The Dev Program adds Manychat Applications, JSON defined apps with an App Key per install, actions, option sources and triggers that outside systems fire through a webhook, with public apps listed at apps.manychat.com after review. Manychat publishes no MCP server of its own.

Sourcehelp.manychat.com/hc/en-us/articles/14959510331420-how-to-generate-a-token-for-the-manychat-api-and-where-to-get-parametersread 2026-10-09

The AI Playground lets a brand chat with its AI as a follower would before going live, with suggested or custom questions. It flags questions the Knowledge library cannot answer and offers to add the missing information on the spot, shows which entries each answer drew on, and takes a thumbs up or down with a reason. Brands retest after each change, then go live.

Flows can be previewed before publishing, and the Randomizer splits contacts across weighted paths for A/B tests. There is no saved test set, scoring run or comparison between AI versions, and the Randomizer reports no results by path, so a change is judged by reading test conversations.

Sourcecommunity.manychat.com/product-updates/ai-playground-test-your-ai-detect-knowledge-gaps-and-go-live-with-confidence-9189read 2026-10-09

Manychat works through the official messaging platforms as a Meta Business Partner and TikTok Marketing Partner, so DMs, comments and messages travel through the Instagram, Messenger, WhatsApp, TikTok and Telegram channels. No browser, desktop or computer is driven by the AI. Public web pages enter Knowledge as extracted text, a fetch for answers rather than an agent operating a site.

Sourcemanychat.comread 2026-10-09