Advanced · Lesson 18 of 22

Coda AI: Summarize, Classify, Generate

Coda's AI is woven into the doc — not a separate chatbot. It reads your tables, runs prompts per row, classifies at scale, and answers questions about your actual data. Here's how to use it without burning through your credits.

⏱ ~26 min 🤖 AI & Automation ✅ Prerequisite: Lesson 10
01 — AI Overview 02 — AI Columns 03 — Summarize 04 — Classify & Tag 05 — Extract & Generate 06 — AI Block & Chat 07 — Credits & Best Practices Practice
01 — AI Overview

Three AI interfaces, one connected doc.

Coda AI is not a general-purpose chatbot you paste content into. It's built into the doc and has direct access to your tables, pages, and formulas. When you write an AI prompt that references a column, Coda passes the actual cell values into the AI call for you — no copy-paste required.

There are three surfaces where Coda AI operates:

📊

AI Columns

A column type that runs a prompt for each row automatically. Reference any other column with @[Column Name]. Results update when source data changes.

📄

AI Blocks (Canvas)

An AI-powered block on a page. Insert with /ai. Runs a prompt against the current page's entire content — tables, text, and embedded docs included.

💬

AI Chat Panel

A chat interface (brain icon, top-right) that answers questions about your entire doc. Ask in natural language — it queries your actual live tables and pages.

Context-aware, not copy-paste AI Tools like ChatGPT require you to copy data out, paste in, copy results back. Coda AI operates on live data — columns reference other columns by name, AI results flow into formulas, and everything stays in the doc.
02 — AI Columns

A prompt that runs for every row.

An AI column is a regular column with the type set to AI. You write a prompt once, and Coda evaluates it for every row — passing that row's data into the prompt. It's like a formula column, but the computation happens in a language model rather than a formula engine.

To add an AI column: click + on any table → column type → AI. The prompt configuration panel opens. Write your instruction. Reference other columns by typing @[Column Name] — Coda substitutes the actual cell value for that row when it evaluates.

AI Column Configuration Panel
AI Column: "Meeting Summary"
Prompt
Summarize @[Meeting Notes] in 3 bullet points focusing on decisions and open questions.
Auto-run for new rows: ON
Re-run on source change: Manual

After writing the prompt, click Run all to process all existing rows. New rows added after that will auto-run the AI column unless you've paused it. Coda offers four preset task types — Summarize, Find action items, Find key insights, Classify — plus a Custom prompt option for anything else.

03 — Summarize

Compress long text to the essentials.

The Summarize preset is ideal for any column containing long-form text: meeting notes, customer feedback, article text, ticket descriptions. Coda generates a condensed version for each row — no manual reading required.

Summarize Prompt Examples
// Meeting notes → 3 bullet points
Summarize @[Meeting Notes] in 3 bullet points.
Focus on decisions made and open questions.

// Customer email → situation summary for a busy reader
Summarize @[Email Body] in 2 sentences
for a busy executive. Focus on key decisions.

// Research article → TL;DR
Write a one-sentence TL;DR of this article:
@[Article Text]
📋

Meeting notes → action summary

100 rows of raw meeting notes become 100 concise summaries. Filter and group them in a view to track outstanding decisions across all meetings.

📧

Customer emails → situation summary

A support inbox with hundreds of emails — one AI column gives every row a two-sentence summary that takes 3 seconds to scan instead of 3 minutes to read.

Be specific about format "Summarize this" produces inconsistent results. "Summarize in exactly 2 bullet points, each under 15 words" produces predictable, usable output. Format specificity is the single biggest quality lever.
04 — Classify & Tag

Automated tagging at table scale.

The Classify task instructs the AI to assign each row to one of a set of categories you define. The output is a tag — consistent, automatically applied, and filterable. Add a Select column type for the AI column so values snap to allowed options and prevent typos.

Classify Prompt Examples
// Sentiment classification
Classify @[Feedback] as exactly one of:
Positive, Negative, Neutral.
Respond with only the category label.

// Support ticket routing
Categorize this ticket as one of:
Bug, Feature Request, Question, Other.
Ticket: "@[Subject]@[Description]"
Respond with only the category label.

// Lead scoring
Classify this lead as: Hot, Warm, or Cold.
Company size: @[Employees].
Industry: @[Industry].
Notes: @[Notes]. Respond with one word only.

Once your AI column assigns labels, the table becomes filterable and groupable by AI-assigned category. Group feedback by sentiment, route tickets by type, or sort leads by temperature — all without a single manual tag.

Constrain classify outputs Always list the exact categories and tell the AI to respond with only the label. Unconstrained classify prompts produce inconsistent labels that break your filters. "Respond with only the category label" is required boilerplate for every Classify prompt.
05 — Extract & Generate

Structured data and personalized content at scale.

Extract: pull structure from messy text

The Extract pattern finds and returns a specific piece of information from a larger block of text. This is useful when data arrives unstructured — emails, notes, form submissions — but you need it in a column for filtering and sorting.

Extract Prompt Examples
// Extract company name from an email body
Extract the company name from the email below.
Return only the company name, nothing else.
Email: @[Email Body]

// Extract a budget amount
Extract the budget amount from: "@[Proposal Text]"
Return only the number (e.g., 15000). No currency symbol.

// Extract attendee names
List the people mentioned in these meeting notes,
comma-separated. Notes: @[Meeting Notes]

Generate: one prompt, hundreds of personalized outputs

The Generate pattern creates new content based on data from the row. Write the template once as a prompt — Coda generates personalized content for every row. One AI column can draft 500 follow-up emails as fast as you add rows.

Generate Prompt Examples
// Draft a follow-up email per contact
Write a friendly 2-sentence follow-up email to
@[Contact Name] about our meeting on @[Meeting Topic].
Reference the outcome: @[Meeting Summary]

// Personalized outreach subject lines
Write a compelling email subject line (under 8 words)
for outreach to @[Company Name] in @[Industry].

// Product descriptions from specs
Write a 50-word product description for "@[Product Name]"
with these specs: @[Specs]. Tone: professional, benefit-focused.
06 — AI Block & AI Chat

Page-level insight and conversational queries.

AI Block: run AI against a whole page

An AI block is a canvas element (not a column) that runs a prompt against the entire page's content. Type /ai on any canvas to insert one. Coda analyzes all the text, tables, and embedded content on that page — then produces the output inline.

📝

Built-in AI block tasks

Summarize page · Find action items · Find key themes · Extract decisions. Click any preset and the AI block runs immediately against the page.

🔄

Refreshes as content changes

When the page content changes, click "Refresh" on the AI block to re-run the prompt. Useful on meeting pages where notes accumulate throughout the day.

AI block vs AI column AI columns process a table row by row — each row gets its own AI output. An AI block processes one page holistically — it synthesizes everything on the page into one output. Use columns for per-row classification. Use blocks to summarize a week's worth of activity on a single page.

AI Chat: ask questions about your live data

The AI Chat panel — click the brain icon in the top-right of any Coda doc — is a conversational interface that queries your actual doc data. Unlike a general chatbot, it has context about your tables, pages, and their values. It returns answers with citations so you can verify which rows or pages it used.

AI Chat Panel — Example Queries
You asked
"Which tasks are overdue and who owns them?"
You asked
"Who owns the most open deals?"
You asked
"What were the key decisions in last week's meeting?"
You asked
"What's the total estimated hours for the Website Redesign project?"
07 — Credits & Best Practices

Reliable results without wasting credits.

Coda AI operations consume AI credits. Each workspace has a monthly allocation based on its plan. Each row processed by an AI column = one operation. Running an AI column on 500 rows = 500 operations. Check usage in Workspace Settings → AI.

⏸️

Pause AI columns while building

While you're iterating on a prompt, click "Pause" to stop auto-evaluation. Paused columns don't consume credits. Resume when the prompt is right.

🧪

Test on 5–10 rows first

Manually trigger the AI column on a small sample before running it on the full table. Check output quality before committing 500 operations.

Best practices checklist

Feedback Table — AI Sentiment + AI Summary Columns
Feedback Text AI Sentiment AI AI Summary AI
The onboarding process was confusing. I spent an hour trying to set up my first table and couldn't find documentation. Negative Onboarding is confusing; documentation hard to find. User spent excess time on basic setup.
I love the formula system — once I got it, it's incredibly powerful. Way better than Notion for data work. Positive User praises formula system; finds it more powerful than Notion for data-heavy use cases.
The mobile app is okay. Missing some features from desktop but works for quick edits on the go. Neutral Mobile app functional but limited vs desktop. Suitable for quick edits only.
Automations are a game changer for our workflow. Saved us hours every week on status updates alone. Positive Automations significantly reduce manual workflow overhead. High positive impact on team productivity.
Both AI columns run automatically for new rows. Sentiment uses a Classify prompt; Summary uses a Summarize prompt — each referencing @[Feedback].
Practice

Test your knowledge.

Lesson 18 Quiz

5 Questions
Question 1 of 5
How do you reference another column's value inside an AI column prompt?
✓ @[Column Name] is the reference syntax for AI column prompts. Coda replaces it with the actual cell value for each row when the AI runs.
Question 2 of 5
What is the best practice for Classify prompts to avoid inconsistent output?
✓ Constrain classify prompts by listing exact categories and instructing the AI to return only the label. Without this, outputs vary in phrasing and break your filters.
Question 3 of 5
What is an AI block (canvas), and how do you insert one?
✓ An AI block is a canvas element inserted with /ai. It runs a prompt against the page's full content — text, tables, embeds — and refreshes when content changes.
Question 4 of 5
What does the AI Chat panel (brain icon) operate on?
✓ AI Chat queries your actual doc data across all tables and pages. It returns answers with citations so you can verify which rows or pages it used.
Question 5 of 5
What is the best way to manage AI credits when building and iterating on an AI column prompt?
✓ Pause the AI column while editing the prompt — paused columns consume no credits. Once the output looks right on a small sample, unpause and run on the full table.
← Previous Building Forms & Intake Workflows Lesson 17