Every company has a spreadsheet that runs the place. Finance pretends it is temporary. Ops knows it is load-bearing. Somebody’s name is in a formula comment from 2019.
You could spend a quarter replacing it with a Proper System. Or you could pair an agent with the sheet you already live in and buy back ten hours a week without a migration sermon.
Prefer the second option when the sheet is good enough and the pain is repetitive.
Why spreadsheets still win
They are:
- visible
- editable by non-engineers
- easy to sort, filter, and argue about in meetings
- already embedded in habits
Agents are great at transforming text and structured rows. Spreadsheets are great at being the shared surface. Together they are a blue-collar automation stack.
Good jobs for a sheet-plus-agent combo
- classify inbound rows (lead type, urgency, topic)
- draft blurbs from a row’s fields (meta descriptions, ad variants, FAQ answers)
- clean messy free text into normalized columns
- summarize weekly tabs into a manager note
- flag anomalies (missing phone, weird CPL, duplicate names)
Notice these are propose-and-write-cells jobs, not “agent secretly emails customers from a hidden macro.”
Bad jobs (please do not)
- letting the model overwrite financial source numbers without a review column
- putting API keys in a cell and sharing the workbook with the company
- fully autonomous outbound from a CSV because it felt efficient
- one giant prompt that “figures out the whole workbook”
If a wrong cell could move money, care, or compliance, you want a human checkpoint or a harder system than a sheet.
The pattern you use
Input columns humans or integrations fill.
Helper columns the agent fills: drafts, labels, confidence, rationale.
Final columns humans confirm or edit.
Status column drives the queue: new, drafted, needs review, done, error.
The agent never touches done rows unless asked. That alone prevents a lot of chaos.
Baserow, Airtable, Google Sheets, Excel
This is not religious. Use what the team opens every morning.
Structured tools (Baserow / Airtable) make AI fields and linked records easier. Plain Sheets work fine with an Apps Script or a small worker that reads rows and writes back.
The important part is row-level identity and clear column contracts, not the logo.
Briefing the agent per row
Send compact JSON, not the whole tab.
Example fields: keyword, page_type, brand_voice, banned_claims, source_url.
Instruction: “Return only these keys: meta_title, meta_description, confidence, notes. Meta title under 60 characters. No claims absent from source.”
Structured output saves you from scraping poetry out of paragraphs.
Human QA without killing speed
Review in batches. Filter needs review and confidence below threshold. Approve with a keystroke or a checkbox.
For low-risk drafts, auto-accept above a confidence line and spot-check 10 percent. For regulated or customer-facing language, keep the human gate longer than your impatience wants.
Speed is worthless if you publish garbage at scale. Garbage at scale is just a new job title: cleanup.
Keep an audit trail
Write:
- model / prompt version
- timestamp
- input hash or source row id
- who approved
When someone asks “why does this meta title exist,” you need more than a shrug and a myth about “the AI.”
A starter project for this week
Pick one dull column your team fills manually for 20 plus minutes a day. Wire drafts into a helper column. Measure time saved and edit rate for seven days. If edit rate is high, improve the brief or the source fields. If edit rate is low and quality holds, expand.
Do not start by automating the entire workbook. Earn trust on one column. Spreadsheets remember trauma.
Closing
The future of work is not always a shiny agent canvas. Sometimes it is a familiar grid with one smarter column and a review habit that respects the humans who still own the outcome.
Pair the agent with the spreadsheet you already live in. Constrain what it can write. Log what it did. Keep the boring surface your team trusts.
Bottom line
That is use you can ship before lunch, which beats a platform RFP that starts in Q4 and ends in regret.
