# Answer the helpdesk on Slack. Keep every access decision human.

> Channel agents answer employee IT questions on Slack from a scoped knowledge base; access reviews, incident runbooks, and vendor recon run with human gates and an audit trail.

- Canonical: https://www.brahmalabs.io/use-cases/it-operations/
- Department: IT & internal operations
- All departments: https://www.brahmalabs.io/use-cases/

## The reality in internal IT

The helpdesk queue is the same twenty questions on rotation. Password resets, VPN setup, which license to request, why the printer driver broke again. The answers exist — in a wiki nobody opens. Meanwhile the structural work slips: quarterly access reviews get compressed into the week before the audit, the SaaS inventory is whatever the card statement implies, and incident response depends on whoever remembers where the runbook lives. And the team's automation scripts accumulate credentials — pasted into code, copied between repos, rotated never.

## What changes with governed agents

The wiki becomes retrievable. Markdown and text docs are chunked, embedded, and indexed in the platform's own vector store; agents query them through a read-only search_kb tool that returns passages with source files and relevance scores. Scope is fixed server-side: an agent granted the IT knowledge base can search that base and nothing else, and it cannot widen its own reach — not even if the model asks.

Channel agents put those answers where employees already are. Each conversation carries a spend ceiling, and the tool surface on channels is restricted to tools explicitly marked channel-safe. A thread the agent cannot close moves to a human without losing its context.

Credentials follow one rule everywhere: agents receive short-lived, scoped references injected at call time, never raw secrets. An exfiltrated prompt has nothing to leak. Agent steps run in sandboxed containers behind egress allowlists, and the run page shows which domains a run actually reached — the security team verifies instead of trusting.

Recurring hygiene moves onto durable scheduled runs. A run parked on an approval waits days without burning compute; a run interrupted by a restart resumes where it stopped.

## What stays human

Every access decision — the agent assembles evidence, the system owner decides. Every remediation step that touches infrastructure — one approval gate per mutating stage. Every escalated conversation — an operator takes over, the agent steps back. The append-only audit log records the lot across auth, config, run, and approval events, so the trail exists whether or not anyone remembered to write it down.

## The workflows

### 1. Employee IT helpdesk on Slack

- **Trigger:** An employee posts a question in the IT help channel on Slack.
- **Flow:** The channel agent searches the IT knowledge base through the read-only search_kb tool and answers in-thread, grounded in retrieved passages with their source files and relevance scores.
- **Human gate:** Unresolved threads escalate to the Conversations inbox, where an operator takes over mid-conversation with the full history and hands back to the agent when done.
- **Output:** A resolved Slack thread, or an assigned conversation in the inbox carrying the complete exchange.

### 2. Quarterly access reviews with human sign-off

- **Trigger:** A quarterly schedule starts a review run for each system in scope.
- **Flow:** The agent pulls account and permission listings via integration actions, diffs them against the roster in a Data Table, and drafts a per-system evidence packet flagging orphaned and over-scoped accounts.
- **Human gate:** System owners approve or reject every flagged account in the reviewer inbox; the agent never revokes access — all decisions are human.
- **Output:** A versioned evidence artifact per system and a decision trail in the append-only audit log.

### 3. Incident runbooks with staged approvals

- **Trigger:** A monitoring webhook fires, or an operator triggers the runbook manually.
- **Flow:** The agent gathers diagnostics in a sandboxed step, matches symptoms against runbook docs in the knowledge base, and proposes the next remediation stage with its supporting evidence.
- **Human gate:** Every stage that mutates infrastructure parks at an approval gate; on-call approves stage by stage, with SLA timeouts and escalation chains if nobody responds.
- **Output:** A run page showing each stage, decision, and captured outbound domain, plus a post-incident artifact.

### 4. SaaS and vendor inventory recon

- **Trigger:** A weekly schedule kicks off reconciliation against the vendor Data Table.
- **Flow:** The agent reads the vendor table, checks entries against billing exports and SSO app listings via integration actions, and writes discrepancies — unknown apps, lapsed owners, seat drift — back as flagged rows.
- **Human gate:** The IT lead reviews flagged rows and decides what to offboard or renegotiate; the agent proposes, it never cancels anything.
- **Output:** An updated Data Table with flagged rows and a recon summary posted to the team's Slack channel.

## Platform pieces this leans on

- Channel agents (Slack + email)
- Conversations inbox with takeover and handback
- Vector knowledge base + search_kb
- Data Tables
- HITL approvals with SLA timeouts and escalation
- Credential vault (scoped references, never raw secrets)
- Append-only audit log
- Egress-controlled sandboxes

---

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