Skip to main content

​
AI auto-reply for WhatsApp from your knowledge base

When a customer messages your WhatsApp channel, Orbit can answer within seconds using only your own knowledge base. Every inbound message runs through transcription (for voice notes), intent classification, and retrieval over your published content, and a reply is sent only when the grounded answer clears the confidence threshold you set. Below the threshold, nothing is sent and the conversation waits for your team like any other inbound. The feature is off for every organization until an owner or admin enables it. This guide covers enabling and tuning it per organization, testing it safely before go-live, and the day-to-day loop of monitoring and closing knowledge gaps.

​
How an inbound message flows

  1. Inbound message. A customer’s text message arrives on your WhatsApp channel and opens or continues a conversation.
  2. Voice notes are transcribed. Voice messages are transcribed with Deepgram and the transcript enters the same pipeline, so customers who prefer voice notes get answers too.
  3. Intent classification. The message intent (billing, support, complaint, and so on) is recorded on the conversation and applied as a tag when auto-tagging is on.
  4. Retrieval. The five chunks most relevant to the question are retrieved from your knowledge bases by semantic similarity.
  5. Grounded answer and confidence check. An answer is drafted strictly from the retrieved chunks. If the chunks do not contain the answer, the draft is discarded and the confidence falls below any threshold you can set.
  6. Reply or handoff. Above the threshold, the reply is sent on the WhatsApp thread and appears in the inbox like any outbound message. Below it, no bot message is sent at all and your team takes the conversation.
Very short messages (fewer than 20 characters, such as “hi”) skip the AI pipeline entirely and go straight to your team.

​
Prerequisites

  • A connected WhatsApp channel. If you have not connected one yet, follow the WhatsApp getting-started guide.
  • A populated knowledge base. Auto-reply is only as good as the content it can retrieve. Upload, chunk, and embed your content as described in the knowledge base lifecycle guide. A knowledge base with no embedded chunks always ends in a handoff.
  • An owner or admin to change the organization’s enablement, described below.

​
Enable and configure it per organization

Read the current state:
cURL
Node SDK
Enable it and set the threshold in one call. Both fields are required, and the change needs an owner or admin role:
cURL
Node SDK
The threshold accepts 0.5 to 0.95 and defaults to 0.75. Higher is more conservative: more conversations go to humans, and the ones the bot does answer are more reliably backed by your content. Disabling (enabled: false) takes effect immediately; no further auto-replies are sent.

​
Intent tagging for your taxonomy (optional)

Classification reads its own organization settings block, inbox_ai, separately from the auto-reply enablement. By default each intent maps to a platform tag (for example cancellation maps to churn-risk). You can rename individual labels to your internal taxonomy without losing the other defaults by setting an additive intent_tag_map in your organization settings:
The same settings block exposes classify_enabled, sentiment_enabled, translate_enabled, and auto_tag_enabled if you want to switch off classification, sentiment escalation, or inbound translation for your organization. Those switches do not gate the knowledge-base auto-reply; the auto-reply lives behind its own per-organization enable only.

​
What a healthy auto-reply looks like

A customer asks: “Do you have parking at the downtown store?” and the knowledge base contains a store-info article. The customer receives a one-to-three sentence answer drawn from that article, for example: “Yes, the downtown store has free parking behind the building on Main Street.” On your side, the sent message is marked as an auto-reply and stores its confidence score on the message record, so your team can always see which messages the bot sent and how confident the retrieval was. When confidence is below your threshold, the customer sees nothing from the bot: no placeholder and no “I don’t know” message. The conversation simply lands in your team’s queue like any other inbound.

​
Test it safely before enabling

Before flipping the enable on live traffic, preview retrieval with the knowledge base search endpoint, the same retrieval that grounds the bot’s answers. Run a sample set of 20 to 30 real questions taken from your contact reasons and review the top chunks each question returns:
  • If the right article lands in the top five for most questions, retrieval is ready.
  • Questions that return nothing, or the wrong article, are knowledge gaps to close in the knowledge base before enabling, not after.
Then start conservative: enable at 0.8 to 0.85 for the first week, review what went to handoff, and step down toward 0.75 as the gaps close.

​
Disclosure, opt-out, and quiet hours

  • Disclosure. Depending on your markets you may need to tell customers that an AI assistant can answer them. Orbit ships AI disclosure settings for exactly this; see the AI disclosure setup guide for what applies to your organization.
  • Reaching a human. Any question the knowledge base cannot answer confidently goes to a human, and the organization’s enable is an immediate off-switch. On RCS, every auto-reply additionally carries a one-tap “Talk to an agent” action.
  • Quiet hours. Auto-reply never initiates contact; it only answers a message the customer just sent. Scheduled sends and campaigns keep honoring your quiet hours configuration.

​
Monitor it and close the loop

  • In the inbox, auto-replies appear in the conversation thread with their confidence, so supervisors can review what the bot said.
  • Watch what the bot could not answer: the knowledge gap surfaces cluster unanswered and escalated questions, and the knowledge gap miner turns them into a ranked worklist.
  • Feed those clusters back into the knowledge base. The auto-draft loop can turn mined contact reasons into review-ready draft articles, so the next customer who asks gets an answer that exists.
If your team spots a wrong answer, raise the threshold while you fix the underlying article. If too many answerable questions wait for humans, check whether the retrieved chunks actually contain the answer: improve the article wording when they do not, or lower the threshold when they do.

​
RCS uses the same engine

RCS auto-reply reuses the same retrieval and confidence core with its own per-organization enablement (rcs_kb_enabled plus a matching threshold in organization settings), so enabling WhatsApp does not enable RCS and vice versa. RCS replies add suggested actions, “That helped, thanks” and “Talk to an agent”, so the contact can confirm resolution or hand off in one tap.