# Every inquiry extracted, enriched, scored, and routed — with citations a rep can check

> Governed AI agents for sales: cited lead enrichment, deterministic scoring, pipeline hygiene sweeps, N-of-M discount approvals, and channel replies behind human gates.

- Canonical: https://www.brahmalabs.io/use-cases/sales/
- Department: Sales
- All departments: https://www.brahmalabs.io/use-cases/

## The reality in most sales teams

An inquiry lands in a shared inbox. Someone copies it into the CRM, guesses at the missing fields, and moves on. Research happens per rep, per deal, in browser tabs that close when the call starts. Two reps quote the same account different discounts because approval lives in a chat thread nobody can find later. And the pipeline decays quietly: deals with no next step, stage dates that stopped matching reality, forecasts built on records last touched a quarter ago.

None of this is a talent problem. It is a plumbing problem. Extraction, enrichment, scoring, and routing are repetitive, judgment-shaped work — and reps do them badly because reps are paid to sell.

## What changes with governed agents

The plumbing becomes workflow. Extraction is an agent step with a defined output schema, so a WhatsApp message and a web form produce the same lead record. Enrichment runs inside a sandbox with a network egress allowlist, and every claim it adds carries a source the rep can open — cited research, not model recall. Scoring is deterministic rules, not model mood: the same lead scores the same twice, and your platform team can read the rules.

The anchor flow on this page is not hypothetical. It runs in production for a manufacturing client: inquiry in, cited brief and score out, and incomplete leads get a follow-up email asking exactly the questions the form left blank.

Everything rides a substrate the security team reviews once. Credentials stay in the vault; agents get short-lived scoped references, never raw keys. PII redaction runs before inference. Customer-facing channel agents get a restricted, channel-safe tool surface and a per-conversation spend ceiling. Every run writes to an append-only audit log, and each run's actual outbound domains are captured on the run page.

## What stays human

Selling. Agents never negotiate, never promise, and never send to a prospect without a named human gate. Discounts above threshold need real consensus — N-of-M, recorded, with who approved what and when. A rep can take over any live conversation mid-thread and hand back when done. The agent's job is to put a scored, cited brief in front of the rep before the call; the call is still theirs. Agents propose; humans decide.

## The workflows

### 1. Inbound lead intelligence

- **Trigger:** A new inquiry arrives — web form webhook, email, or a WhatsApp channel message.
- **Flow:** An agent extracts company, contact, and intent fields; a research step enriches from the web with cited sources; deterministic rules score the lead; for incomplete leads the agent drafts a follow-up email asking exactly the missing questions.
- **Human gate:** Sales reviews the scored brief and its evidence trail before outreach; drafted follow-up emails only send after a named approver signs off.
- **Output:** A scored lead row with citations in a Data Table or the CRM, a versioned brief artifact, and an approved follow-up sent from the connected email channel.

### 2. Pipeline hygiene sweep

- **Trigger:** A weekly schedule with a skip overlap policy walks every open deal.
- **Flow:** An agent pulls open deals through the CRM integration, flags records with no recent activity, missing next steps, or stage dates that stopped matching reality, and drafts the suggested correction for each one.
- **Human gate:** Nothing writes back on its own; every suggested CRM update waits in the reviewer inbox until the deal owner accepts or rejects it.
- **Output:** A stale-deal report in a Data Table, approved corrections applied to the CRM, and every change recorded in the audit log.

### 3. Quote and discount approval

- **Trigger:** A rep submits a quote through a form or a workflow webhook.
- **Flow:** The workflow checks the discount against threshold rules; below the line it proceeds, above it a HITL node routes the deal context to N-of-M approvers with a dynamic approval form, and the run parks durably until they decide.
- **Human gate:** Above-threshold discounts need N-of-M consensus — sales leadership plus finance — with SLA timeouts and an escalation chain if nobody responds.
- **Output:** A recorded approve-or-reject decision in the audit log and a CRM record updated only after humans sign off.

### 4. Meeting brief prep

- **Trigger:** A scheduled run ahead of booked calls, or a rep requests a brief for a named account.
- **Flow:** An agent pulls account history, searches the knowledge base through the read-only search_kb tool, runs cited web research inside an egress-controlled sandbox, and assembles a one-page brief with a source for every claim.
- **Human gate:** The rep reads and judges; the brief is preparation, not a script, and nothing sends or updates anywhere unless the rep acts.
- **Output:** A versioned brief artifact with passage-level citations, linked from the run page.

### 5. Channel inquiry replies

- **Trigger:** A prospect messages a connected WhatsApp or email channel.
- **Flow:** A channel agent answers routine product questions grounded in the sales knowledge base via search_kb, on a restricted channel-safe tool surface with a per-conversation spend ceiling.
- **Human gate:** Replies can be gated on approval before sending, and a rep can take over any conversation from the shared inbox mid-thread, then hand back to the agent.
- **Output:** Answered inquiries in the Conversations inbox with open, pending, or resolved state and per-conversation spend visible.

## Platform pieces this leans on

- Channel agents (WhatsApp/IG/Messenger/Slack/email)
- Conversations inbox with takeover and handoff
- Vector knowledge base + search_kb
- Data Tables
- N-of-M approvals with SLA timeouts
- Durable long-running runs
- Egress-controlled sandboxes
- Append-only audit log

---

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