Vendor Patterns, Thresholds, and a Category Taxonomy
Good expense categorization isn't AI magic — it's a well-designed ruleset that AI happens to apply fast and consistently. Before you touch a prompt, you need three things nailed down: what your categories actually are, what patterns identify a vendor, and what thresholds matter. Get these right and AI becomes reliable. Skip this step and you'll get plausible-sounding categorization that's wrong in ways you won't notice until month three.
1. Your category taxonomy. Most orgs need far fewer categories than they think. A workable starting set:
- Travel
- Software & Subscriptions
- Meals & Entertainment
- Office Supplies & Equipment
- Professional Services (legal, consulting, contractors)
- Facilities & Utilities
- Uncategorized / Needs Review
Notice the last one isn't optional — a taxonomy without an explicit "needs review" bucket forces every ambiguous transaction into a category that's probably wrong, and you'll never know it happened.
2. Vendor patterns. Real transaction data is messy — the same vendor shows up as "AMZN Mktp US," "Amazon.com," and "AMAZON.COM*AB1C2." A good ruleset defines the pattern, not just the name: "Any vendor string containing AMZN or AMAZON → Office Supplies, unless amount exceeds $500, in which case flag for review." That second clause matters — a $14 Amazon order is almost certainly supplies; a $2,400 one might be equipment that belongs in a different bucket, or might need a PO check.
3. Amount thresholds. Thresholds are where a lot of the real judgment lives. Useful threshold questions to answer up front:
- What amount, regardless of vendor, always gets flagged for a second look? ($1,000? $5,000? — set this based on your own materiality, not a generic number.)
- Are there vendors where any amount over a much lower bar is unusual? (A $600 office-supplies transaction is normal for some teams, a red flag for others.)
- Do round numbers or repeated identical amounts deserve a second look? (Could indicate a duplicate charge or a recurring subscription worth confirming.)
Putting it together. A ruleset isn't a prompt — it's a short, written spec: category list, vendor-pattern-to-category mapping for your most common vendors, and threshold rules that override the pattern match. Write this down before you bring AI in. AI's job in the next lesson is to help you draft and refine this spec quickly — not to invent your categories from scratch with no input from you.
▶️ Try this
Write your own starting taxonomy: 5–7 categories plus an explicit "needs review" bucket, in a plain list. Then pick your three most frequent vendors and write the pattern-to-category rule for each, including any amount threshold that should override it. This becomes the spec you hand to AI in the next lesson.