The Task
Every Monday, I go through the vendor spreadsheet.
It's 60 rows. Columns include vendor name, order number, promised delivery date, current status, last contact date, and a short notes field. Most rows don't need anything from me. They're on track.
But about 10 rows need a follow-up. Maybe the delivery is late. Maybe the status hasn't updated in a week. Maybe the vendor promised something and went quiet.
My job is to find those 10 rows and turn them into a follow-up list for the week. Who to contact, about what, and by when.
My baseline for this task is 45 minutes. That's the manual version: scrolling the sheet, sorting by status, reading the notes column, and writing the list by hand.
This week, I tried an AI-assisted version.
The Setup
Tool: One general-purpose AI assistant (free tier)
Date: October 15, 2026
Task: Extract follow-up items from a vendor spreadsheet and build a prioritized list
Input I provided: A redacted version of the spreadsheet. Vendor names replaced with [VENDOR 1] through [VENDOR 12]. Order numbers replaced with [ORDER]. All dates and statuses kept as-is.
What I wanted: A follow-up list with three columns: priority, vendor, and suggested action.
My baseline: 45 minutes
The Prompt
Here's the exact prompt I used.
"Below is a spreadsheet of vendor orders. I need a follow-up list for this week. For each row that needs a follow-up, give me: (1) priority (high, medium, low), (2) vendor, (3) order, (4) reason for follow-up, and (5) suggested action. Sort by priority. Only include rows that actually need a follow-up. Do not include rows that are on track. If a row is unclear, include it and mark it as 'unclear.' Keep the list under 15 rows. Here is the data: [pasted redacted spreadsheet]."
The Result
The AI returned a list in about 12 seconds.
It identified 12 rows that needed follow-up. My manual count was usually around 10. So the AI found two extra.
Here's what it got right:
It correctly flagged 8 of the 10 rows I would have flagged myself.
It correctly identified the two highest-priority items (a late shipment and a vendor who hadn't responded in nine days).
It provided reasonable suggested actions for most rows.
The format was clean and easy to read.
Here's what it got wrong:
It flagged one row that was already marked "delivered" in the status column. That was a false positive.
It flagged another row that was marked "cancelled." Another false positive.
It missed one row where the status was "partially shipped" and the notes column said "follow up on remaining units." That was a false negative.
It misordered two priorities. A vendor who was only one day late was listed as high priority, while a vendor who hadn't responded in a week was listed as medium.
The Corrections
I spent 13 minutes fixing the list.
2 minutes: removing the two false positives
3 minutes: adding the missed row
4 minutes: reordering priorities
2 minutes: adjusting two suggested actions that were too vague
2 minutes: reading the final list out loud and confirming against the sheet
Total time from start to finish: about 28 minutes.
My baseline was 45. So the AI version saved about 17 minutes. That's a 38% reduction.
What I Learned
A few things stood out.
The AI reads statuses literally. It treated "delivered" as "delivered." It didn't second-guess. That's usually good. But in this case, two rows had statuses that were probably outdated. The AI didn't catch that.
The notes column is where the real signal is. The row the AI missed had a status of "partially shipped" but a note that said "follow up." The AI prioritized status over notes. I do the opposite. Notes are usually where the real action items live.
Priority sorting is the weakest part. The AI got most priorities right but mixed up two. Priority requires context the AI doesn't have—like which customer is waiting, which vendor is reliable, and which delay will cause the most downstream problems.
The time saved was real. 17 minutes on a weekly task is about 1.2 hours a month. Not huge, but not nothing. And the output was easier to work from than a blank page.
The redaction step added time. I spent about 5 minutes redacting the spreadsheet before pasting it. That time is included in the 28 minutes. Without redaction, the AI version would have been faster, but I'm not willing to paste real vendor names into a tool.
What Still Needed My Attention
Even after corrections, I still did four things manually.
I confirmed every row against the sheet. The AI can't be trusted to catch outdated statuses. I have to read the original.
I made the priority calls. Priority is a judgment call. I know which customers are waiting and which vendors are reliable. The AI doesn't.
I adjusted the suggested actions. Some were too generic. "Contact vendor" is not an action. "Ask for updated delivery date by Wednesday" is.
I decided what to leave off. The AI included a row that was technically follow-up-worthy but not urgent. I removed it. Not everything that can be followed up should be followed up this week.
The Verdict
Kept, with a specific use case. I'll keep using AI for the first pass on the vendor sheet. It saves real time and gives me a starting point. But I'll always do the priority sort and the notes-column check myself.
Baseline note: My baseline was 45 minutes. The AI version took 28. That's a 17-minute saving, or about 38%. Over a month, that's roughly 1.2 hours. I'll track this again over the next four weeks to see if the pattern holds.
The Limitation
This test used one spreadsheet on one day.
A different spreadsheet with different columns might produce different results. A spreadsheet with cleaner data might produce a better AI output. A spreadsheet with more ambiguous statuses might produce a worse one.
Also, I redacted the data before pasting it. That added time and reduced the AI's context. If I were using a paid tool with better privacy controls, I might paste more data and get a better result. But I'm not willing to do that with vendor names.
And the AI still needs a human to make priority calls. That's not going away.
Test it in real life.
No comments yet.