What this agent does
An Intercom auto-responder is a customer-facing surface, which means it has to do two contradictory things well. It has to respond quickly enough that customers feel heard, and it has to be wrong rarely enough that the wins outweigh the embarrassments. Retrieval-augmented generation, gated by a strict grounding policy and a topic allow-list, is the configuration that makes both true.
The agent reads inbound conversations, retrieves relevant passages from a curated knowledge base, drafts a reply that cites the passages, and either auto-sends (for low-stakes topics after calibration) or routes to a teammate. The draft is always visible to the teammate. The customer always sees an AI label.
For the broader pattern of "agent that handles a defined task," see what an AI agent can actually do. For the safety-mindset prerequisites, see AI agent safety and guardrails.
Intercom permissions
Intercom apps run as OAuth integrations. The agent installs as an app and operates as a teammate inside the workspace.
- Read conversations. To see new and ongoing conversations and their full message history.
- Read help-center articles. To pull article metadata when the retrieval result points to a public docs page.
- Write replies as a teammate. Configured as a non-Operator teammate named "AI Helper" or similar, with the AI-label badge enabled on outgoing messages.
- Read team and admin metadata. Read-only, used to route handoffs to the right team based on the conversation's existing assignment.
The agent does not get write access to user attributes, custom data, or company records. Those are different scopes with much higher blast radius, and the agent has no reason to need them.
Knowledge base curation
"Train on all our docs" is the wrong shape. A help center accumulates outdated articles, internal-facing pages, and conflicting versions of the same answer. An auto-responder trained on the raw corpus will eventually surface the wrong version of an article it found.
The right shape is a curated index. A defined set of help-center articles, marked as canonical, indexed for retrieval. Other articles exist for human-only search. The index is rebuilt on a schedule (we run nightly) and stale-article checks happen weekly.
Each indexed passage carries metadata: the source article's URL, the last-updated date, the article's topic tag, and a confidence-required threshold (0.8 for general help, 0.95 for billing-adjacent, 1.0 for legal language). Higher thresholds make the agent more likely to refuse and route, which is the right behaviour for higher-stakes content.
Curation is a continuing investment, not a one-time setup. Plan for half a customer-support engineer's time per week to keep the index honest. The investment pays back many times over in reduced false answers.
Grounded retrieval and citation
The agent's response pipeline has three stages.
- Retrieve. Take the customer's message, retrieve the top five passages from the index with hybrid search (dense embedding similarity plus BM25 keyword match). If no passage clears the topic's threshold, refuse and route.
- Draft. Generate a reply with the constraint that every sentence must cite one of the retrieved passages by inline marker. The generator is prompted to refuse rather than improvise if the retrieved passages do not cover the question.
- Validate. A post-generation check verifies that every sentence has a citation and that the cited passage actually says what the sentence claims (token-level overlap check; deep semantic check on the lowest-overlap sentences). Sentences that fail validation are stripped before the reply is queued.
The customer sees the reply with a "Sources" footer linking to the help-center articles cited. This earns trust faster than a smooth ungrounded answer, in our testing and in the general literature on retrieval-augmented support assistants.
Human handoff rules
The agent hands off to a human under five conditions.
- No retrieval above threshold. The most common case. The agent posts a short "I don't have this answer, routing to a teammate" and assigns to the appropriate team.
- Topic is on the never-auto list. Billing, account changes, security, legal. The agent drafts a starting point in the teammate sidebar but does not send anything to the customer until a teammate approves.
- Customer asks for a human. Any variant of "talk to a person" or "speak with someone" triggers immediate handoff. The agent does not negotiate.
- Sentiment is hostile. If the inbound message scores above a configurable hostility threshold (frustration markers, escalation language), the agent routes to a senior teammate rather than attempting a first response.
- Repeat customer with prior complaint. Customers with an open complaint in the last 14 days always route to the teammate handling that complaint, not to the agent.
Guardrails
Six guardrails for a customer-facing agent.
- AI label always visible. Intercom's "AI sender" badge or an equivalent. Customers know who they are talking to.
- No pricing in free-text. Pricing comes from the canonical pricing page, by link only. Inventing prices breaks contracts.
- No PII in retrieved passages. The knowledge base index excludes any document containing customer identifiers. The agent has no business surfacing another customer's data even by accident.
- Topic allow-list. The agent can only auto-send on topics that have explicitly graduated past calibration. New topics start at "draft only."
- Daily false-positive review. A teammate reviews five random auto-sent replies per day and grades them. The grade feeds the per-topic threshold.
- Customer escalation always works. "Talk to a human" is non-negotiable, even if the agent thinks its answer is right.
Common mistakes
Treating the knowledge base as a dump. Indexing every help article and every internal wiki page gives the retriever too many wrong-looking-right options. Curate.
Hiding the AI label. Customers feel deceived when they discover after the fact. The short-term close-rate gain is not worth the long-term trust loss.
Auto-sending on billing. The single fastest way to acquire angry chargebacks. Billing is human-only.
No citation footer. A grounded answer without a citation footer looks the same as an ungrounded one to the customer. Show the sources.
Letting hostile threads stay with the agent. A customer who is already frustrated needs a human, not another AI response. Sentiment detection earns its keep here.
Frequently asked questions
What does an Intercom auto-responder agent actually do?
It listens for inbound Intercom conversations, retrieves the most relevant passages from a curated knowledge base, drafts a reply citing the passages, and either auto-sends or hands the draft to a human depending on confidence and topic sensitivity. It always inserts a 'see full answer here' link to the docs page when one exists. It never invents pricing, account-specific data, or policy details.
Which Intercom permissions does the agent need?
An Intercom OAuth app with scopes for reading conversations, posting messages as a configured teammate, and reading the help-center articles. The agent should be configured as a teammate (not as the default Operator bot) so customers see a clearly labelled AI sender and routed conversations are auditable per teammate ID.
How does the agent avoid hallucinating answers?
Retrieval-augmented generation with a strict grounding policy. The generator only emits sentences that cite a retrieved passage. If retrieval returns no passage above the relevance threshold, the agent does not draft a reply; it routes the conversation to a human. The threshold is configurable per topic and tightened for billing, security, and compliance questions.
Should the agent answer billing questions?
Only the general-policy ones (how to cancel, where invoices live, which tier includes which feature). Specific billing questions about a customer's account, refunds, or pro-ration are always handed off because the agent does not have access to the billing system as a source of truth and inventing dollar amounts is the fastest way to lose customer trust.
How long before the agent auto-sends without human review?
Two weeks of shadow drafts, then graduated auto-send by topic. The lowest-stakes topic (where to find a settings page) graduates first, then password-reset instructions, then onboarding questions. Billing, security, and account-specific topics never auto-send. The agent's job for those is to draft a starting point a human ships in one click.
Three takeaways before you close this tab
- Grounded or routed. The agent never improvises.
- AI label always. Trust is built by labelling, not by hiding.
- Topic allow-list. Billing and security stay human-only.
Sources
- Intercom Developer Docs, "Authentication and scopes", retrieved 2026-05-11, developers.intercom.com/docs/build-an-integration/authentication
- Intercom Developer Docs, "Conversations API reference", retrieved 2026-05-11, developers.intercom.com/docs/references/conversations
- Lewis et al., "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", NeurIPS 2020, retrieved 2026-05-11, arxiv.org/abs/2005.11401
- NIST, "SP 800-53 AC-6 Least Privilege", retrieved 2026-05-11, csrc.nist.gov/sp800-53/AC-6
- Aryan Agarwal, "Gravity support-agent guardrails", internal v1, May 2026, About
Related agent patterns
The same shape, applied to other tools and surfaces:
- AI agent for inbox triage, the email parallel to chat-based support routing.
- AI agent for Slack triage, classifying and routing channel messages.
- AI agent for Mailchimp segmentation, the outbound-side audience pattern.
- AI agent safety and guardrails, the principles every customer-facing agent respects.
- AI agent tool use explained, how an agent gets connected to Intercom and its help center.
- AI agent failure modes, the cases a support agent has to defend against.
- How we test AI agents with 80 tests per capability, the calibration methodology.