Build an IT access-request resolver that takes a plain-English request (e.g. "give Jordan admin on Salesforce"), grounds it in a CMDB, decides what to do against an editable policy, acts on that decision, and closes out an incident — so an IT/IAM service desk gets consistent, auditable, policy-driven handling of routine access tickets instead of doing every step by hand. Trigger it with an HTTP route: POST an access request as JSON { text, requester }. Keep the route private to space members unless otherwise specified. Wire it as a short linear flow: Receive — validate the incoming request and normalize it for the rest of the flow. Understand — use an LLM to extract who is asking, which application, the intent, and urgency; then look those up in a CMDB / asset-inventory system so the decision is based on real records rather than the raw text. Decide — use an LLM to apply an editable policy (kept in one file a human can easily change) to the enriched facts, returning a strict decision such as auto-approve, needs-manager-approval, or deny, along with the action to take and rule-citing reasoning. Act — create/update an incident ticket, then carry out the decision (grant access, request approval, or reject), and post a notification. Its output is the HTTP response summarizing what happened. For the external systems, name my choices but keep them swappable: use an LLM (I used Anthropic's Claude — any capable model works), a CMDB + ITSM data service for lookups, records, and access-group changes (mine is a combined CMDB/ITSM API — swap in ServiceNow, Jira Service Management, or your own), and post notifications to Slack (or your team's chat tool). Make notifications optional: if chat isn't configured or a call fails, fall back to logging and a note in the result so the loop still closes. This workflow is read-write — it creates/updates tickets and changes access-group membership. Tools used: Anthropic, CMDB/ITSM data service, Slack