The same nine questions consume a senior person's week
Each team gets the same questions every week, and they always land on the same two people. How do I request access? What is the deployment process? Who is responsible for this service? Where is that document stored?
The questions are not difficult. They are costly because they interrupt, and because the answer lives in one person's memory instead of in a place you can search.
An assistant helps only when it is honest about what it covers. If an internal answer is confidently wrong, people trust it at once. Nobody checks an internal tool, so a wrong answer costs more than the interruption it prevented.
The other half is knowing which questions a bot should not answer. Questions that involve a person, a decision that has not yet been made, or a policy with exceptions belong with a human. The assistant must say so promptly.
How the agent is built in Agent Studio
Permission scoping comes before capability. The agent has access only to what the requesting team is authorized to see. When teams have different entitlements, a separate agent per scope is safer than one agent that reasons about access.
MCP connections point at the internal documentation, read live rather than copied. The system prompt requires every answer to cite its source, and to return an explicit "not documented" when the source does not address the point, rather than inferring.
Prompt skills hold the glossary: internal names, service nicknames, and acronyms. Those are what distinguish an assistant that understands the question from one that only matches keywords.
The flow sits where people already are. When a Slack node receives the mention, the Agent Chat node passes the question on. The reply goes back to the thread.
The valuable result is the set of undocumented answers. A Google Sheets node records them. The list becomes the documentation backlog, ranked by how often people ask.
Sessions are durable, so a follow-up in the same thread keeps the previous context. The interaction feels like a colleague, not a system.
What this agent is built from
- Permission scope: one agent per entitlement boundary rather than one agent reasoning about access.
- MCP connections: internal documentation, read live rather than copied.
- System prompt: cite the source on every answer, and return “not documented” rather than inferring.
- Prompt skills: your glossary of internal names, nicknames, and acronyms.
- Slack (Communication) plus Agent Chat (Util): puts the assistant in the thread where the question was asked.
- Google Sheets (Integration): logs unanswered questions as a ranked documentation backlog.
- When an ActionFlow is enough: If the nine questions are genuinely fixed, a keyword-matched reply workflow answers them with no risk of invention.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.