How to Draft (and Approve) Support Replies in Intercom with AI
Summary
- Auto-sending AI replies presents material risk: one confidently wrong answer can trigger a customer relationship reset, internal investigation, and cleanup, so draft with AI and gate every customer-facing message behind human approval.
- Classify intent before drafting: transactional queries such as order status, receipts, and password resets are safe candidates, while refunds, account access, upset customers, and billing inquiries need caution or direct human routing.
- Production guardrails matter: Intercom returns up to 500 conversation parts, the reply endpoint has no idempotency key, and rate limits enforce 10-second windows with 10,000 calls/min for private apps (25,000 ceiling), so check the latest part ID before sending and pause on 429 using
X-RateLimit-Reset. - Start with one or two high-volume, low-risk query types and run every draft through approval in the first month; Jinba Flow provides
WAIT_FOR_APPROVALas a standalone workflow node and captures each run as an audit-ready chat thread.
Repetitive first-response questions are a measurable cost. When every incoming conversation requires a support agent to read, interpret, and type a reply from scratch, ticket volume dictates headcount. The obvious solution is Intercom automation, and AI makes full automation appear within reach.
Auto reply support messages carry real risk. A single confidently wrong reply wastes the time it saved and triggers a customer relationship reset, an internal investigation, and cleanup work that compounds. Support teams that have tried full AI auto-send consistently identify the same failure mode: the model produces a plausible answer to the wrong question, and the error is not caught before it reaches the customer.
The effective approach drafts the reply with AI, gates it on human approval, and sends it only after a person has signed off.
This article examines each step of that workflow, built on Intercom's conversation API and a platform that provides the WAIT_FOR_APPROVAL primitive needed to make human-in-the-loop approval a requirement rather than an intention.
Why "Draft and Approve" Outperforms Auto-Send
The case for keeping a human in the loop is structural.
Intercom's evolution of its native AI, Fin, reflects this. Intercom moved from measuring "resolutions" to measuring "outcomes" because, as Intercom states directly, solving a customer problem with full automation is not always appropriate; the goal is to reach the right result safely and efficiently. The commercial signal is equally clear: Intercom prices a Fin Procedure handoff, where Fin does initial work and then passes control to a human, at $0.99 per outcome, the same rate as a full AI resolution. The handoff is a first-class outcome rather than a fallback.
The implication for teams building custom Intercom automation is straightforward. The draft-and-approve pattern is the recognised, production-grade model for AI in customer support rather than a compromise on automation.
The practical starting point is classification. Before generating any draft, the workflow must determine whether a query is safe to draft at all. Billing disputes, account access requests, refund conversations, and anything with a frustrated tone require human handling from the first reply. Transactional queries, including order status, password reset confirmation, and receipt requests, are candidates for AI drafting. Establishing that boundary is consistently the decision that makes AI support workflows reliable.

Step 1: Read the Conversation
The workflow begins by fetching the inbound conversation from Intercom.
The Retrieve Conversation endpoint is GET /conversations/{conversation_id}, authenticated with a Bearer token and an Intercom-Version header. The conversation ID accepts the literal string "last" to retrieve the most recent conversation.
The response returns the full conversation model with several fields that drive routing decisions:
state: open, closed, or snoozedpriority: whether the conversation is flagged as urgentwaiting_since: the timestamp of the customer's last reply; null if an admin replied lastconversation_parts: the full message history, capped at 500 parts
The 500-part limit applies to any account with long-running or multi-session conversations. Only the most recent 500 parts are returned; older history must be retrieved separately if context depth matters for classification.
In Jinba Flow, this step maps to an INTERCOM_GET_CONVERSATION workflow node, which returns the structured conversation object for downstream steps to consume.
Step 2: Classify Intent Before Drafting
The conversation text goes into an AI classification step before any reply is drafted. This is the gate that keeps sensitive queries out of the auto-draft path entirely.
In Jinba Flow, this is an ANTHROPIC_INVOKE call using Claude Sonnet 5.5, the default model in Jinba's Claude-native environment. The prompt instructs the model to assign the conversation to one of a predefined set of categories. A practical starting set:
- Transactional: order status, receipt requests, password reset confirmation
- Billing Inquiry: invoice questions, subscription changes (draft with caution)
- Technical Bug Report: reproducible errors, integration failures
- Refund Request: route to human queue, do not draft
- Upset Customer: route to human queue, do not draft
- Account Access: route to human queue, do not draft
The classification output controls a branch in the workflow. Conversations in the "route to human" categories skip the drafting step entirely and are assigned directly to a support agent or specialist inbox. Only conversations in the safe-to-draft categories proceed to Step 3.
This design addresses the persistent question teams face when evaluating AI for support: which tickets are suitable for automation. Ticket type alone does not determine the answer. A classification step that enforces the boundary before the AI writes a customer-facing word does.
Intercom's Fin takes a comparable approach. Its answer engine uses retrieval-augmented generation (RAG) over a centralized content library specifically to avoid hallucinations, and scopes answers by audience targeting and custom instructions. For custom workflows, the equivalent is combining a tightly scoped classification prompt with knowledge base context retrieved before the drafting call.
Step 3: Draft the Reply and Await Human Approval
For conversations that pass the classification gate, a second ANTHROPIC_INVOKE call generates the draft reply. The prompt receives the conversation text, the classification label, and any relevant knowledge base excerpts as grounding context. Providing that context directly reduces the risk of the model producing an answer that is technically coherent but factually incorrect for the specific account or product version.
The draft output is routed to a WAIT_FOR_APPROVAL step rather than sent to Intercom.
In Jinba Flow, WAIT_FOR_APPROVAL is a workflow node that pauses execution entirely. The drafted reply is presented to a designated support agent via an in-chat approval card or the Jinba run console, with explicit approve and reject controls. The workflow does not advance until a human acts.
This is the architectural difference between AI-augmented support and AI-autonomous support. The customer never receives text that has not passed a human review.
Intercom does provide a native human-in-the-loop primitive: the "Loop in teammate / agent" step inside Fin Procedures. It pauses a Fin Procedure until a teammate provides input or a timeout is reached. However, this step is exclusive to Fin Procedures. For custom, API-driven workflows built outside Fin, there is no equivalent native approval gate. A platform like Jinba Flow provides WAIT_FOR_APPROVAL as a standalone workflow node that can be used in any automation sequence.
Every workflow run in Jinba is captured as a reviewable chat thread. This creates an automatic audit log: what was drafted, who reviewed it, and what decision was made. When a reply later requires review, the record is already there.
Step 4: Send on Approval, Escalate on Rejection
When the approval step resolves, the workflow branches on the outcome.
On approval, the workflow calls the Intercom Reply to Conversation endpoint: POST /conversations/{conversation_id}/reply. The payload for an admin reply is:
{
"message_type": "comment",
"type": "admin",
"body": "The AI-drafted and human-approved reply text.",
"admin_id": "YOUR_ADMIN_ID"
}
In Jinba Flow, this is the INTERCOM_SEND_MESSAGE node, which accepts the approved body text and the conversation ID from earlier in the workflow.
One production detail requires attention: by default, a bot or Operator reply to an inbound email conversation is not delivered as an email to the customer. It is stored as an unnotifiable bot comment and remains invisible to the customer until an Intercom account team enables the relevant feature flag. Workflows that include email-initiated conversations must account for this. In the draft-and-approve model, the assigned human agent is already in context and can confirm delivery, which is another reason the pattern is more reliable than full auto-send.
On rejection, the workflow branches to a human escalation path. The conversation is assigned to a senior agent or moved to a review inbox, with the rejected draft preserved in the run log for context.
Production Considerations
Two operational constraints apply at scale.
Idempotency. The POST /conversations/{id}/reply endpoint does not support an idempotency key. If a webhook fires twice for the same event, the same reply can be sent twice. The solution is to check the ID of the most recent conversation part before executing the send step, and skip execution if the workflow has already replied to that part.
Rate limiting. Intercom enforces API rate limits in 10-second windows. Private apps share 10,000 calls per minute per workspace, with a ceiling of 25,000 per minute. Exceeding the limit returns an HTTP 429 error with error code rate_limit_exceeded. The X-RateLimit-Reset header carries the timestamp of the next window reset. Workflow logic must catch 429 responses and pause before retrying rather than looping immediately.
Jinba Flow runs at $5–20 per month at scale, using an 80% deterministic execution model. The majority of the workflow (fetching the conversation, routing by classification, constructing the payload) runs as rule-based logic. Only the AI invocation steps consume model compute. This keeps costs predictable and avoids the reliability problems that emerge when stochastic AI agents control branching logic.
Where to Go From Here
The four-step structure is the foundation: read the conversation with INTERCOM_GET_CONVERSATION, classify intent with ANTHROPIC_INVOKE, draft and gate with WAIT_FOR_APPROVAL, then send or escalate based on the human decision via INTERCOM_SEND_MESSAGE.
Start with one or two of the highest-volume, lowest-risk query types. Build the classification categories around the query types the team already knows are safe to answer without escalation. Run the approval step for every draft in the first month. The approval log will show where the AI is reliable and where it is not, and that record is the evidence base for any future decision to expand the automation scope.

The goal of Intercom automation is to remove the repetitive, low-judgment work that keeps agents from handling the conversations that require them, while retaining humans in customer-facing communication.
Frequently Asked Questions
What is human-in-the-loop approval in Intercom automation?
Human-in-the-loop approval means an AI-generated reply cannot reach a customer until a support agent explicitly approves it. The AI drafts the message, the workflow pauses at an approval gate, and only an approve action triggers the send. In a custom Intercom workflow, this is enforced with a WAIT_FOR_APPROVAL node that halts execution entirely. Execution does not advance until a person acts, so the customer never receives text that has not been reviewed.
Why is auto-sending AI replies risky in customer support?
A single confidently wrong reply costs far more than the time it saved. The typical failure mode is a plausible answer to the wrong question, delivered before anyone can catch it, rather than obvious nonsense. That triggers a customer relationship reset, an internal investigation, and cleanup work. This is why Intercom shifted from measuring "resolutions" to measuring "outcomes," and why a Fin Procedure handoff to a human is priced as a first-class outcome rather than a fallback.
Which support tickets are safe to automate with AI drafting?
Transactional queries are the safe starting set: order status, receipt requests, and password reset confirmations. Billing inquiries can be drafted with caution, while refund requests, account access issues, upset customers, and anything involving a dispute should route to a human from the first reply. Ticket type alone does not set the boundary. An AI classification step enforces the routing before the model writes a single customer-facing word.
How do I send an AI-drafted reply to a conversation in Intercom?
After approval, call POST /conversations/{conversation_id}/reply with message_type: "comment", type: "admin", the approved body text, and the relevant admin_id. The conversation ID can also accept the literal string "last" when it is retrieved earlier in the workflow. In Jinba Flow, this maps to the INTERCOM_SEND_MESSAGE node, which takes the approved body and the conversation ID carried through from Step 1.
Does Intercom have a native approval step for AI-generated replies?
Yes, but only inside Fin Procedures. The "Loop in teammate / agent" step pauses a Fin Procedure until a teammate provides input or a timeout is reached. For custom, API-driven workflows built outside Fin, there is no equivalent native approval gate. Teams therefore use a platform like Jinba Flow that exposes WAIT_FOR_APPROVAL as a standalone workflow node that can be used in any automation sequence.
How do I avoid sending the same Intercom reply twice?
Check the ID of the most recent conversation part before executing the send, and skip if the workflow has already replied to it. This matters because the reply endpoint does not support an idempotency key, so a webhook that fires twice for the same event can produce a duplicate customer-facing message. A pre-send check on the last part ID is the reliable guard.
What are Intercom's API rate limits for conversation automation?
Private apps share 10,000 calls per minute per workspace, with a ceiling of 25,000 per minute, enforced in 10-second windows. Exceeding the limit returns an HTTP 429 with error code rate_limit_exceeded. Read the X-RateLimit-Reset header for the next window timestamp and pause before retrying rather than looping immediately, especially inside approval-heavy workflows where several API steps run per conversation.
Why are bot replies to email conversations not delivered to the customer?
By default, a bot or Operator reply to an inbound email conversation is stored as an unnotifiable bot comment and stays invisible to the customer until an Intercom account team enables the relevant feature flag. Any workflow that handles email-initiated conversations must account for this. In the draft-and-approve model, the assigned human agent is already in context and can confirm delivery, which is another reason the pattern is more reliable than full auto-send.
How much does it cost to run an Intercom AI reply workflow?
Jinba Flow runs at $5–20 per month at scale, using an 80% deterministic execution model. Most of the workflow (fetching the conversation, routing by classification, constructing the payload) runs as rule-based logic, and only the AI invocation steps consume model compute. That keeps costs predictable and avoids the reliability problems that appear when stochastic AI agents control branching logic.