Create tables with AI Studio

A table is useful when every row answers the same questions. If one row is a vendor, each vendor should have the same fields. If one row is an action, each action should follow the same status, owner, and due-date rules. Without that schema, AI produces a decorative grid that still needs to be rebuilt.

Choose Table in AI Studio when rows and columns are the deliverable: a comparison, checklist, scorecard, extraction, inventory, or execution plan you want to edit as a spreadsheet-style canvas object.

The short version

1. Open the editable canvas where the table should live.
2. Select the source objects you want AI Studio to use.
3. Open AI Studio and choose Table.
4. Add completed attachments or public source URLs if needed.
5. Define the column headers, what one row represents, how to order rows, and what to write when a value is missing.
6. Select Submit.
7. Check every row against the source, especially numbers, units, scores, owners, and blank cells.

Start with a prompt such as:

Create a comparison table from the attached proposals. First row: Vendor, Price, Contract term, Included support, Main risk, Source. One vendor per row. Preserve currencies and units. If a value is not stated, write "Not stated". Do not infer scores or recommendations.

This walkthrough shows AI Studio turning canvas context into an editable table:

The placeholder appears first, then the finished spreadsheet-style object replaces it on the canvas. Review the cells before treating the grid as source data.

Choose Table only when the grid is the artifact

Table is intentionally a more specific choice than Doc. Use it when consistent fields matter more than narrative.

Asking for “a table” inside a Doc creates a document with prose and an in-document table. Choosing Table creates a standalone spreadsheet-style object whose cells are the main result.

If your request needs a long explanation of tradeoffs, choose Doc. If every option can be evaluated against the same fields, choose Table.

Start with source context and a schema

On desktop, open the canvas and use the AI Studio panel. On supported phones and narrow screens, open AI Studio from the canvas toolbar. You need canvas edit access and usable workspace AI credits.

AI Studio can build the table from selected canvas objects, prompt text, completed attachments, and pasted public URLs. A draft can include up to 10 attachments, with up to 20 MB read for context from each file. You can select up to 20 canvas objects.

Selection is source context. AI Studio creates a new table near the visible canvas area instead of changing the selected objects.

Before you submit, define four things:

1. Headers: Name the columns in the order you want them.
2. Row rule: Say what one row represents, such as one vendor, one requirement, one task, or one source record.
3. Ordering: Say how to sort the rows, such as source order, highest priority first, or chronological order.
4. Missing-value rule: Choose a visible marker such as Not stated, Unknown, Not applicable, or Needs review.

AI Studio tables support up to 20 rows and 6 columns. The first row counts toward the 20-row limit, so ask for a header row and up to 19 data rows. If your source needs more, narrow the scope, split it into several tables, or extract the highest-priority rows first.

Match the prompt to the table job

Extract structured facts

Use extraction when the source already contains facts and the table should preserve them.

Extract the contract terms from the attached files. First row: File, Effective date, Renewal, Notice period, Fee, Source page. One contract per row. Copy dates and currency exactly. Write "Not stated" for missing values.

Do not ask AI to “fill the gaps” unless inference is genuinely part of the job. A complete-looking cell can be more dangerous than an honest missing value.

Compare options

Define neutral axes before asking for a recommendation.

Compare the selected tools. First row: Tool, Primary use, Setup effort, Collaboration, Export options, Evidence. One tool per row. Keep claims neutral and source-backed. Write "Needs verification" when the source is unclear.

If you need a reasoned recommendation after the grid, create a Doc instead and ask for a comparison table inside it.

Create a checklist

Make the completion rule explicit.

Create a launch-readiness checklist. First row: Check, Pass condition, Owner, Status, Evidence, Notes. One check per row. Use only the selected plan. Leave Owner as "Unspecified" unless a person or team is named. Use "Needs review" for status unless the source states otherwise.

AI should not turn an unstated status into Complete or invent the person responsible.

Build a scoring matrix

Define the scale and whether AI is allowed to score.

Create a vendor scoring matrix. First row: Vendor, Security, Reliability, Support, Price, Total. Use a 1-5 scale. Score only when the attached evaluation explicitly provides enough evidence. Otherwise write "Not scored". Do not calculate a Total when any required score is missing.

If you want AI judgment rather than extraction, state the rubric for each score and verify every score yourself. A number generated without a rubric only looks objective.

Make an execution plan

Keep plans grounded in stated commitments.

Turn the selected project notes into an execution table. First row: Work item, Owner, Due date, Dependency, Status, Source. One work item per row. Preserve the wording of dates. Use "Unspecified" for missing owners or due dates. Do not invent dependencies or status.

For a proposal that explains why the work matters, choose Doc. For the rows people will update during execution, choose Table.

Keep rows and columns clean

The output is most usable when every row has the same number and meaning of cells.

• Put headers in the first row and keep them short.
• Use one entity or action per row.
• Keep one kind of value in each column.
• Include units in the header, such as Cost (USD) or Duration (days).
• Keep dates in one format if the source permits it.
• Do not combine owner, status, and due date in one cell.
• Use a consistent missing-value marker instead of switching between blanks, dashes, and guesses.
• Ask for a Source or Evidence column when factual verification matters.

Avoid a catch-all Details column if the details can become two clearer fields. But do not create six columns merely because six are available. Every column should support a real decision or follow-up action.

Verify the result cell by cell

AI Studio returns a standalone spreadsheet-style object sized to its content. That makes it easy to scan and edit, but the neat geometry does not validate the data.

Review:

When accuracy matters, compare each row with the source and keep a Source or Evidence column. If a cell cannot be verified, label it visibly instead of letting the table imply certainty.

After review, edit cells directly on the canvas. For a targeted AI follow-up on the existing table, open its AI version badge or Object Chat and say exactly what may change:

Check the Price and Contract term columns against the attached proposals. Change only those cells. Use "Not stated" when the source has no value.

Spend credits on the schema, not repeated extraction

Table generation uses workspace AI credits. Deleting a completed table does not return credits.

Before selecting Submit:

• Start with the smallest useful source set.
• Decide the headers and row rule yourself.
• Limit extraction to the records you actually need.
• State the missing-value marker.
• Define a scoring rubric before asking for scores.
• Ask for a Source or Evidence column when verification matters.
• Split work that would exceed 20 rows or 6 columns.

Manual cell edits are usually cheaper than regenerating a nearly correct table. If the schema is wrong across every row, revise the prompt and regenerate once. If three cells are wrong, fix those three cells or request one bounded Object Chat edit.

When table generation does not behave as expected

Deleting a processing table placeholder cancels the task from the canvas. If you contact support, include the canvas link, approximate request time, prompt summary, intended headers and row count, attachment types, whether the placeholder appeared, and whether credits were blocked or consumed. Do not send private source files or customer data unless support specifically needs them and you are allowed to share them.