Expense report automation.

Receipt OCR extracts: merchant name, amount, tax breakdown, date, line items where present, currency, payment method. AI categorizes into expense category (meals, lodging, transport, software, supplies, client entertainment) based on merchant + line items + employee context. Multi-currency normalized to company base currency at the transaction-date FX rate. Per-diem entries handled separately — no receipt required, fixed amount per day per location.

Within-policy expenses below the auto-approval threshold (typically $75–$200 depending on category, configurable). No human review required — it's a known-good pattern with corporate-card matching evidence. 60–75% of typical expense volume routes here once the policy is well-tuned. The auto-approval threshold is the highest-leverage tuning lever; raising it from $50 to $150 typically saves more time than any other policy adjustment.

Auto-approved expenses match against corporate card transactions for verification — same merchant, same amount, same date confirms the receipt is real. Daily 5% random audit sample reviewed by finance for false-positive 'within policy' approvals. Sample audit is what keeps the auto-approve lane honest; without it, policy drift goes undetected and edge-case violations compound into significant policy-leakage over quarters.

Above auto-approval threshold or category requiring business purpose justification (client entertainment, T&E for non-routine travel, professional development) routes to direct manager. Manager sees the AI's policy assessment inline + the receipt + employee's stated business purpose. Approve, reject, or request clarification. Average review time: 90 seconds vs 5–8 minutes from cold review. Manager's inbox doesn't fill with $12 coffee receipts; only the items needing actual judgment.

Manager reviews flag patterns: repeat employees with high-spend pattern, repeat categories that always need approval (signal that policy threshold is wrong, not that employees are bad), unusual location/time patterns. Aggregated weekly into manager dashboard — managers see their team's expense behavior trends. Policy-tuning signal: if 80% of flagged-for-review items get approved, the threshold is set too low.

Out-of-policy expenses, items above the manager threshold (typically $1,500+), or anything flagged by AI as suspicious (duplicate submission, suspicious merchant patterns, employee history flags) route to finance. AI provides full context: policy violation specifics, employee's prior submission patterns, similar expense comparisons, suggested resolution. Finance approves with policy exception note, rejects with reason, or requests additional documentation.

AI fraud detection runs on flagged items: receipt forgery patterns (cropped images, mismatched fonts), duplicate submissions across periods, mileage claims that don't match map distance, hotel claims for the same night by multiple employees in different cities. Confirmed fraud cases route to People-team + Legal. Quarterly fraud-pattern review surfaces emerging schemes. Audit trail captures everything for SOX-relevant cases — every decision, who made it, what evidence was reviewed.

Reimbursement paid through next payroll cycle (or direct ACH if you've separated reimbursement from payroll). Employee notified with a clean breakdown — what was approved, what was reimbursed, payment date. Average time-to-reimbursement compresses from typical 14–28 days to 3–7 days. Faster reimbursement is one of the highest-leverage employee experience moves you can make; it removes a quiet ongoing irritation that compounds.

Rejected expense routes back to employee with specific reason — 'meal exceeded $75 city-tier limit; please add business purpose justification' rather than vague 'not approved.' Employee can re-submit with corrected information, justification, or accept the rejection. Most rejections are fixable with one round-trip if the rejection reason is specific. Vague rejections create back-and-forth that erodes employee trust in the process.

Rejection patterns aggregate to the policy-update queue. If a category gets rejected 70% of the time, the policy threshold is probably wrong — the spending is clearly happening for legitimate reasons. Quarterly policy review sees the patterns; finance + People-team adjust limits and exception rules. The expense policy is a living document tuned by actual spending behavior, not a 12-month-old PDF nobody reads.