Keep coming back to a pattern that sounds too humble for the current AI fashion show:
A spreadsheet as source of truth. An agent that fills and flags rows. A human who decides.
It is underrated ops. It ships. It audits. It does not require a six-month platform saga before you get value on Tuesday.
Why the sheet still wins
Spreadsheets are:
- visible
- sortable
- easy to diff socially (“why is this cell red?”)
- friendly to non-engineers
- good enough for many workflows that are basically queues with metadata
A lot of internal ops is queues with metadata. We overcomplicate that fact because queues with metadata do not sound like a keynote.
What the agent should do
In my setups, the agent:
- drafts values for empty fields from approved sources
- proposes classifications
- flags anomalies (missing required data, conflicting values, stale rows)
- writes notes in a dedicated column, not by silently overwriting history
- never becomes the only place truth lives
The sheet remains canonical. The agent is a power user with a leash.
What the agent should not do
- delete rows without a human confirm path
- invent identifiers
- overwrite human-approved cells without a change log
- hide uncertainty as confidence
- bypass the columns you use for audit
If you cannot explain an agent’s edit in a weekly review, the design is wrong.
A concrete workflow shape
Columns like:
- status
- owner
- source link
- agent draft
- human final
- flag reason
- last reviewed at
The agent fills agent draft and flag reason. Humans promote to human final. Downstream systems read only finals.
That one separation prevents a surprising amount of chaos.
When to graduate from the sheet
Graduate when:
- row counts and permissions outgrow sanity
- you need strict relational integrity
- workflow branching is genuinely complex
- audit requirements demand stronger controls than a careful sheet process
Until then, a disciplined sheet plus agent beats a half-built app that nobody opens.
Checklist
- One canonical tab (or clearly versioned tabs)
- Agent writes to draft columns only
- Flags are visible without opening a separate tool
- Humans have an owner column that means something
- Weekly cleanup of stuck rows
Story
I watched a team wait on internal tool development while a sheet-plus-agent process could have cleared their backlog the same week. The tool eventually mattered. The waiting did not.
Use the humble system to learn the workflow. Build the fancy system after the workflow stops changing every Thursday.
Closing
Spreadsheet plus agent is still underrated ops because it respects how organizations actually trust data: they want to see it, sort it, argue about a cell, and move on.
Keep the sheet as source of truth. Let the agent fill and flag. Let people decide.
It is not glamorous. It is operationally adult.
Permissions and copies
A sheet that everyone can edit is not a source of truth. It is a group chat with cells.
Restrict edit access. Give view access liberally. Keep an archive tab for old finals. When the agent proposes bulk changes, do it on a staging tab first. Promotion is a verb for a reason.
Closing
Humble tools plus clear ownership beat glamorous tools plus fog. Keep the sheet. Leash the agent. Move the rows.
Start this week
Pick one queue your team already tracks in a sheet. Add draft and flag columns. Wire an agent to propose, not finalize. Review daily for five days. You will learn more than another platform demo will teach.
Metrics from the sheet
Because the sheet is visible, you can count flags, cycle time, and stuck rows without a BI project. Those numbers tell you whether the agent helps or merely rearranges dirt. If flags never shrink, fix the brief before you hire another tool.
Closing note
Bottom line
If a founder pitch mocks spreadsheets, check whether they have ever cleared an ops backlog on a deadline. Humility scales. Glamour stalls.
