Classify 10,000 rows with one AI formula

Classification is the most common AI-in-spreadsheets workload. It's also where most people overthink it — sentiment as positive/negative/neutral, support tickets as billing/technical/feature, leads as hot/warm/cold. The hard part isn't the model; it's writing a prompt that holds up across 10,000 unpredictable rows.

This post is the playbook: where classification beats hand-rolled rules, the prompt patterns that don't drift, and the cost math.

The one-line version

=AI_CLASSIFY(A2, "positive, negative, neutral")

That's it. A2 is the text. The second argument is the category list. The formula returns one category, exactly as written, no commentary. Drag down to 10,000 rows; come back when it's done.

When AI classification beats rules

• Free-text in, structured label out. "I love it!" → positive. Regex won't handle paraphrasing; AI does.
• Domain knowledge embedded. Classify SaaS plans as starter/growth/enterprise just by reading the customer's company description.
• Noisy / multilingual data. Tickets that arrive in 7 languages with typos. Rules can't span that; LLMs can.
• Iteration speed. Adding a new category means editing one string. With rules it means writing a new regex bundle and re-running the whole pipeline.

When rules do beat AI: exact string matching, high-volume hot paths where 50ms latency matters, regulated domains where the auditor wants deterministic logic. For the other 80% of business workloads, =AI_CLASSIFY wins.

Five prompt patterns that hold up

1. Single-category, clean list

=AI_CLASSIFY(A2, "billing, technical, feature_request, account, other")

The shortest pattern. Always include an "other" bucket — without it, the model will hallucinate borderline rows into the most-frequent category.

2. Single-category with instructions

=AI_CLASSIFY(A2,
  "hot, warm, cold",
  "Hot = explicitly asking for a demo or pricing. Warm = engaged but no buying signal. Cold = generic question or info-only.")

The third argument is a free-form instruction string. Use it to nail down boundary cases before the model has to guess.

3. Multi-tag (tagging instead of classification)

=AI_TAG(A2,
  "bug, ui-issue, performance, mobile-only, dataloss, security",
  "Apply up to 3 tags. Be conservative — don't tag if not clearly present.",
  3)

Use AI_TAG when items can have multiple labels. Returns comma-separated.

4. Open-ended classification (no fixed list)

=AI(
  "In 1-3 words, what category does this customer complaint fall into? Examples: shipping delay, broken product, billing dispute.",
  A2)

For exploration. The output isn't constrained — you'll get drift across rows. Useful as a first pass before settling on a fixed taxonomy.

5. Hierarchical classification (two passes)

=AI_CLASSIFY(A2, "support, sales, billing, spam")  // → B2
=IF(B2="support",
   AI_CLASSIFY(A2, "bug, how-to, feature_request, account"),
   "")

Two-stage classification cuts cost in half because the narrower second-tier prompt only runs for the relevant first-tier bucket.

The cost math

Average classification prompt: ~150 input tokens (category list + instructions + the row text), ~5 output tokens (one category label).

Classification is the cheapest LLM workload there is. Don't optimize prematurely — use Gemini Flash by default, swap up if accuracy lags.

Three real-world examples

Customer support triage

Pipe Zendesk ticket subjects + first message into column A. Classify into {billing, technical, feature_request, account, spam} in column B. Route to the right team via Zapier on column B's value. Setup time: 20 minutes. Replaces a 2-hour-per-day human triage step.

Lead scoring from form responses

"Tell us about your company" free-text field → classify as {enterprise, mid_market, smb, individual, junk}. Sort the leads list by that column. Sales talks to the top of the list first.

Survey theme extraction

Open-ended NPS responses → tag with {pricing, performance, feature_gap, support_quality, ux, integrations} via AI_TAG. Pivot table tells you what people actually complain about — without 4 hours of manual reading.

The thing nobody tells you about LLM classification

The model isn't picking from your list — it's predicting the next tokens. If your list overlaps semantically, it'll pick the wrong one consistently.

If you classify with "happy, satisfied, content", you'll get noise. Those aren't different categories — they're synonyms. Either collapse them or differentiate clearly: "happy_about_product, happy_about_price, happy_about_support, generic_positive".

Same trap: temporal labels ("this_week, this_month, this_quarter") where rows span multiple time windows. The model has no clock. Pre-compute the time bucket in another column and classify on the bucket.

Running 10,000 rows: the Bulk runner

Dragging =AI_CLASSIFY down 10,000 rows works, but Sheets will recalculate the entire column whenever anything in the workbook changes. That's expensive.

Open the sidebar → Bulk tab. Set input range A2:A10001, prompt template "Classify {value} as positive, negative, or neutral", output starting cell B2. Hit Run. Results are written as static values — no recalc storms.

You can stop and resume; the partial output stays. Ten thousand rows finish in ~25 minutes against Gemini Flash at default parallelism.

Combining classification with the chat agent

Once you have a classified column, the agent gets useful:

• "Of the rows tagged technical, what are the most common keywords?"
• "Plot the daily count of each category over the last 90 days."
• "Show me 10 example rows from each cluster so I can sanity-check the labels."

The agent reads your sheet, picks the right chart, and writes it inline. No formula gymnastics.