AI-assisted weekly vendor spend review¶
AI usage: AI drafts the vendor-by-vendor summary, flags anomalies, and produces a leadership-ready memo. A human (you) cross-checks every number, verifies the source data, and signs the final memo. AI does not decide what is material; people do.
What you will do¶
- Run a weekly vendor-spend review in about 30 minutes instead of 1-3 hours.
- Produce a vendor-by-vendor summary with anomalies flagged.
- Produce a leadership-ready memo reviewed by a human.
- Keep the workflow repeatable: the same inputs, prompts, and output template every week.
Before you start¶
- A weekly spend export (vendor, invoice date, amount, cost center, notes).
- Access to an AI assistant you are allowed to use with company data (check your org's data policy first).
- Last week's summary for comparison.
- A stable output template for the memo.
Steps¶
1. Prepare the input table (5 min)¶
Export the week's spend to a flat table with columns: vendor, invoice date, amount, cost center, notes. Remove duplicates and obvious errors. Keep the same column contract every week — it is what makes the assistant's output comparable week to week.
2. Set the assistant's context (2 min)¶
Start a fresh session for the week. Paste a short context block:
You are my vendor-spend analyst. I give you a table of this week's vendor spend. You will respond with (1) a vendor-by-vendor summary in a table, (2) anomalies or changes vs last week, and (3) a draft memo for leadership. Do not invent numbers: if a number is not in the table you give me, say so.
3. Have the assistant draft (5 min)¶
Attach/paste the input table and last week's summary, then run:
Here is this week's vendor spend table and last week's summary. Draft the three-part analysis. For each vendor with a change over 15% week-over-week, call it out. Round totals to whole dollars. Mark anything you are unsure about as [UNSURE] instead of guessing.
4. Cross-check the numbers (10 min)¶
You are the final reviewer. This step is not optional. Verify:
- The total row in the draft equals a sum you independently compute.
- Every flagged anomaly exists in the source table.
- [UNSURE] flags have a correct name.
If the assistant reports a total that does not match the table, correct it and note the discrepancy in the memo. A total mismatch means the memo must not ship until resolved.
5. Publish the memo (5 min)¶
Copy the reviewed summary into the leadership memo template: headline, table, flags, and the one thing you are doing about it next week. Sign it. Archive the assistant exchange and the data table alongside the memo for audit.
Review checklist¶
- Numbers independently summed and verified against source table
- Every flagged anomaly exists in the source data
- [UNSURE] items are flagged, not guessed
- Data policy: assistant access is permitted for this data
- Date range covers the full week
Failure modes and fixes¶
| Symptom | Likely cause | Fix |
|---|---|---|
| Assistant totals never match | Pasted table formatting / hidden rows | Re-export the table; sum in the sheet first |
| No vendor flagged, ever | Assistant too conservative | Tighten the prompt threshold to 10% |
| Memo reads like boilerplate | Context block too weak | Add 2 lines about what this week matters for |
| Data policy reluctance | Sharing concern | Run assistant with local-only mode, or redact vendor names |
What can go wrong¶
- Numbers can be wrong. Assistants are not calculators; always independently sum the table before trusting any total.
- Stale data. If the export is cut before the week closes, the review is silently incomplete; check the date range each week.
- Context bleed. Pasted notes from other weeks can leak into the draft; start each week with a fresh session.
- Tooling changes. Assistant behaviors change between releases; re-verify the workflow after major assistant updates.
Run it¶
- Time: ~30 minutes per week (
10 minprep,5 minAI drafting,10 mincross-check,5 minpublish) - Repeat: every week
- Store: final memo in your reporting tool; assistant exchange and data table archived alongside
Related¶
- Editorial standards: How PartlyGood works
- Taxonomy: fields and content types
- Management playbook: Drafting meeting minutes with AI