# Front-line answers on customer channels, grounded in your knowledge base

> Governed agents answer customers on WhatsApp, Instagram, Messenger, Slack, and email — KB-grounded, tool-gated, with human takeover and per-conversation spend ceilings.

- Canonical: https://www.brahmalabs.io/use-cases/customer-support/
- Department: Customer Support & CX
- All departments: https://www.brahmalabs.io/use-cases/

## The reality of the support queue

Support teams run the same loop every day: the same twenty questions, spread across five channels, answered from documentation that drifts a little further from the macros each quarter. Deflection bots promise relief and add a new problem — answers invented from nowhere, no path to a human, no record of what the bot told the customer. Most teams conclude, reasonably, that an unsupervised model in front of customers is a liability.

That conclusion is correct. The fix is not a smarter model. It is a narrower one.

## What changes with governed agents

Brahmalabs puts agents on WhatsApp, Instagram, Messenger, Slack, and email. Each channel has a verification lifecycle — it goes live only once connectivity is proven — and email inbound works whether your stack is Gmail, Microsoft 365, or plain IMAP. What an agent can say and do on those channels is constrained by mechanism, not by prompt:

- Answers are grounded through search_kb, a read-only tool whose scope is fixed server-side. An agent answering an anonymous WhatsApp visitor can search only the knowledge bases it was explicitly granted; the tool cannot widen that scope, regardless of what the model asks for. Retrieval returns passages with source files and relevance scores, so every answer is citable.
- The knowledge base itself is markdown and text files, chunked and embedded into a vector store inside the platform database. Indexing runs on the same durable substrate as workflows.
- Tools must be explicitly marked channel-safe before an agent facing external parties can use them. The customer-facing tool surface is a deliberate subset, not a default.
- Every conversation carries a spend ceiling, and per-conversation spend is captured and surfaced. A chatty thread cannot run away with the token budget.
- Channel agents were red-teamed before ship: a ship-gate suite of adversarial tests, not a hope.

Every thread — every reply, every takeover — is visible in the Conversations inbox; escalations and approvals land in the append-only audit log. When someone asks what the agent told a customer, the thread is there to read.

## What stays human

The Conversations inbox is the seam between agent and team. Operators see every thread across every channel and can step into any of them; the agent yields, the human resolves, the agent takes the thread back. Escalations go through a named approver. New knowledge base articles are reviewed before they are indexed. The agent absorbs the twentieth repeat of a known question; your team keeps the customer whose problem is in no KB. Agents propose; humans decide — on live chat, not just in workflows.

## The workflows

### 1. Front-line answers on customer channels

- **Trigger:** A customer message arrives on WhatsApp, Instagram, Messenger, Slack, or email.
- **Flow:** The agent retrieves passages from its granted knowledge base through the read-only search_kb tool, grounds the reply in cited sources, and works from a channel-safe tool surface under a per-conversation spend ceiling.
- **Human gate:** Any operator can take over mid-conversation from the Conversations inbox and hand back to the agent when done.
- **Output:** A grounded reply on the customer's channel; the conversation tracks open, pending, or resolved in the shared inbox.

### 2. Ticket triage and escalation with SLA timers

- **Trigger:** A new ticket arrives via webhook, or an operator takes over a conversation and assigns it to a support lead.
- **Flow:** The agent classifies severity and topic, checks the knowledge base for known issues, drafts a first response, and routes the ticket to the right queue.
- **Human gate:** A support lead approves each escalation; SLA timeouts push unanswered approvals up the escalation chain.
- **Output:** A triaged ticket with a drafted reply, plus an escalation record when the timer fires.

### 3. KB gap detection

- **Trigger:** Operators log the questions that forced a takeover to a Data Table; a schedule picks up the new rows.
- **Flow:** The agent clusters the unanswered questions, checks each against existing articles, and drafts a markdown article for every genuine gap.
- **Human gate:** A support lead reviews and approves each draft before it is indexed; nothing enters the knowledge base unreviewed.
- **Output:** Approved drafts are indexed into the KB; the gap list lands as rows in a Data Table for tracking.

## Platform pieces this leans on

- Channel agents (WhatsApp/IG/Messenger/Slack/email)
- Conversations inbox: takeover and handback
- Vector knowledge base + search_kb
- Channel-safe tool gating
- Per-conversation spend ceilings
- HITL approvals with SLA timeouts and escalation
- Data Tables
- Append-only audit log

---

Brahmalabs — the AgentOps platform for production AI agents. https://www.brahmalabs.io