Turn a raw expense export into an anomaly-flagged summary before it goes to whoever approves it

You are analyzing a raw expense transaction export to prepare it for whoever reviews or approves it — the goal is to surface anything unusual before that person has to find it themselves, not to approve or reject anything yourself.

TRANSACTION DATA
03/02 Delta Airlines $612.40 'flight'; 03/04 Marriott $890.00 'hotel 3 nights'; 03/04 Marriott $890.00 'hotel 3 nights' (again); 03/09 Staples $1,240 'office supplies'.

EXPENSE POLICY LIMITS (if any)
Meals capped at $75/day, hotel capped at $300/night, no single office supply purchase over $500 without pre-approval.

COMPARISON PERIOD
Prior month: travel $2,100 total, office supplies $180 total.

WHO THIS GOES TO
Direct manager who has to co-sign expense reports over $1,000.

STEP 1 — CATEGORIZE
Group every transaction into a category based on merchant name and description. If a transaction's category is genuinely ambiguous from the description alone, put it in an "uncategorized — needs description" bucket rather than guessing a category that might be wrong; a wrong category is worse than an honest gap.

STEP 2 — FLAG ANOMALIES
Flag, specifically and by transaction: potential duplicate charges (same amount, same or adjacent date, similar merchant), any single transaction that exceeds a stated policy limit, and any category whose total is a significant departure from the comparison period if one was given. For every flag, state the specific reason in one line — never a vague "this looks off," always the concrete pattern that triggered it (exact duplicate amount, which limit was exceeded and by how much, percentage change versus the prior period).

STEP 3 — DO NOT DECIDE
Do not label any flagged item as fraudulent, unauthorized, or a policy violation — you don't have the context to know if it was pre-approved. Use "flagged for review" language only, and note explicitly that a flag is not a conclusion.

OUTPUT FORMAT
1. Category totals table.
2. Numbered list of flagged items, each with the specific reason.
3. Uncategorized bucket, if any, listed by transaction so they can be manually assigned.
4. One line stating how many transactions were reviewed and how many were flagged, for a quick sanity check against the source file.