Expense approval automation with Navan and AI triage
Build an automated expense-policy review pipeline that clears a company's daily approval queue with as little human effort as possible. The goal: Finance teams waste hours rubber-stamping expenses that were never a problem, which buries the few that actually need scrutiny and slows reimbursement. This workflow should sort the queue before a human ever looks at it — bulk-approve the clearly-compliant expenses with simple deterministic rules, and use an AI agent to deep-review only the ambiguous minority, so people spend their attention on the handful of calls that actually need judgement. Trigger it on a daily cron each morning. On each run, pull the pending-approval queue from your expense system (I used Navan via its MCP server — swap in whatever expense/travel tool your team uses). Only consider expenses submitted since the last run, tracked with a persistent cursor so there are no gaps or double-reviews. Structure it as these steps at a high level: Daily trigger — cron step that kicks off the run. Deterministic triage — fetch the queue and split it: auto-approve the clearly-compliant expenses in bulk (conservative rules — when in doubt, send to review, never auto-approve), and pass only the ambiguous ones downstream. AI review agent — an async self-looping agent that deep-reviews only the flagged/ambiguous expenses, pulling detail per expense and classifying each Green / Amber / Red against company policy. Use Anthropic Claude (or your preferred model provider). It writes the finished review to a shared volume. Dashboard — an authenticated web page that reads the latest review and presents it as bulk-approve / quick-fix / needs-judgement, with headline stats and estimated time saved. Chat summary — once the agent finishes, post a morning summary with the counts and time saved to Slack (or your team's chat tool), exactly once per completed review. Decision log + feedback loop — an endpoint the dashboard calls when a reviewer decides on an expense; append it to a durable audit log, and feed recent decisions back into the agent's context so its future classifications align with how the team actually rules. Keep triage conservative so compliance oversight gets stronger, not looser, and make sure every human decision is auditable. Tools used: Navan (via its remote MCP server), Anthropic Claude API, Slack (chat.postMessage), 3B named volumes for persistent state, React + Tailwind for the dashboard.
What this prompt builds
This workflow automates daily expense approval for finance teams by pulling pending items from Navan, applying deterministic rules to bulk-approve compliant expenses, and routing ambiguous cases to a Claude AI agent that classifies them as Green/Amber/Red against company policy. Finance reviews only flagged exceptions via a dashboard instead of the entire queue, with all decisions audit-logged and fed back to improve future classifications. The system reduces manual review time while strengthening compliance oversight and accelerating employee reimbursements.
The problem
Finance teams spend hours every day approving routine expenses—meals, phone bills, taxis—that were never a compliance risk, which buries the few receipts that actually need scrutiny and slows reimbursement for everyone. This workflow automates the daily approval queue: it pulls pending expenses from Navan, bulk-approves clearly compliant items with deterministic rules, and routes only the ambiguous minority to an AI agent that classifies each against company policy as Green, Amber, or Red. Finance reviews a dashboard of exceptions instead of clicking through the entire queue, turning hours of rubber-stamping into minutes of focused judgement.
Solution and impact
The workflow collapses daily approval work into a single dashboard of exceptions: most of the queue is cleared automatically, and Finance spends attention only on the handful of expenses that need human judgement. Because triage is deliberately conservative and every decision is audit-logged, controllers gain stronger compliance oversight while doing less manual work. Compliant expenses are approved same-day instead of sitting in a backlog, so employees are reimbursed faster. A feedback loop captures every human ruling and refines the AI agent over time, so classifications stay aligned with how the team actually enforces policy.