AI Processing

How processing works

When you click Process, every row and every AI column in the grid is worked through in sequence. For each cell, it runs through these steps:

1. Reads your column prompt — the instruction you wrote for this column
2. Resolves @mentions — replaces @Column Name placeholders with the actual cell values from that row
3. Loads documents — if your prompt @mentions a document, it is read and parsed (PDF, Word, PowerPoint, and images are all supported); see Documents & Files
4. Searches the web — if Web Browsing is enabled for this column, a live web search runs for relevant current information; see Web Browsing
5. Generates the answer — using all of the above context, it writes the cell content
6. Applies the format type — structures the output according to the column's format type

Results stream in as they're ready. You'll see cells fill in one by one — you don't have to wait for everything before you start reviewing.

Column processing order

When columns depend on each other (e.g. Column B uses @Column A in its prompt), the order in which they run matters. Grid Settings → Column processing order lets you control this:

• Auto — Order is derived from @mentions. If B mentions A, A runs first, then B. Columns with no dependencies run in one batch afterward.
• Manual — Drag columns in the dropdown to set your own order. Useful when you want a specific sequence regardless of mentions.
• None — All columns run in one concurrent batch. Use when there are no dependencies.

Columns that participate in @mention dependencies show an @ icon in the list. If you've set a manual order and want to go back to automatic, choose Auto in the dropdown (when mentions exist) or None (when they don't).

Stopping processing

While cells are being generated, you can click Stop to cancel active jobs. Stop appears in the toolbar (cancels all active jobs) or on a column header (cancels only that column's jobs). Cancelled cells show "Cancelled by user" and can be reprocessed by clicking Process again.

Batch processing vs. individual regeneration

Batch processing (Process button): Runs AI generation for all cells that are empty or have not been manually edited. Manually-edited cells are skipped. This is the primary way to process a grid.

Individual regeneration: Right-click any single cell and choose Regenerate to rerun AI for just that one cell. This is useful when you want to get a second attempt at a specific result, or when you've changed a column prompt and want to test it on one row before reprocessing everything.

For cell states (Empty, Processing, AI-generated, Manually edited, Error), see Cells.

Error cells

When a cell fails to generate, it enters an Error state. Click the cell to see the error details. Common causes:

• The prompt asked for information that doesn't exist (e.g., finding a URL for a person with a very common name)
• The web search didn't return useful results
• A document @mention pointed to a file that couldn't be parsed

You can always regenerate an error cell, manually enter a value, or leave it empty.

Confidence scores

Every AI-generated cell has a confidence score — a value between 0% and 100% that reflects how certain the answer is. You can see the score by hovering over a cell or clicking into it.

• Above ~80% — Confident result. Spot-check these, but they're usually reliable.
• 50–80% — Moderate confidence. Worth a quick review.
• Below ~50% — Low confidence. Manually verify or regenerate.

Low confidence often means the information wasn't clearly available. Refining your prompt or enabling web browsing can help.

Sources

When Web Browsing is enabled, sources are tracked automatically for each answer. Click on an AI-generated cell to see the URLs that were consulted and a brief summary of what was found at each one. Click any source URL to open it in a new tab and verify the information yourself.

Chain-of-thought reasoning

A reasoning process is used to work through complex prompts — not just a direct guess. When you open a cell's detail view, you can often see the reasoning steps taken before arriving at the answer. This is especially useful for understanding why a particular result was given or why confidence is low.

Writing good prompts

The quality of your results depends almost entirely on how clearly you write your prompt. A few core principles:

Be specific about what you want

• Bad: "employees"
• Good: "How many full-time employees does @Company Name have? Return a number only."

Specify the format in the prompt too

Return the founding year as a 4-digit number only, nothing else.

Use @mentions to provide context

Based on @Company Description, write a one-sentence tagline.

Set scope boundaries

• Bad: "What are the risks?"
• Good: "List the top 3 financial risks for @Company Name, each in one sentence."

Give perspective when relevant

From the perspective of a Series A investor, rate the attractiveness of @Company Name on a scale of 1–10.

For binary questions, specify the expected answer format

Does @Company Name have a free tier? Answer "Yes" or "No" only.

How manual edits are protected

Once you manually type into a cell, it is marked as user-owned. From that point on:

• Clicking Process will skip that cell
• Only an explicit Regenerate with confirmation will overwrite it

This means you can safely mix AI-generated content and manual corrections in the same grid without worrying about losing your work.

Giving feedback on results

Every AI-generated cell has a thumbs up and thumbs down button (visible when you click into the cell). Use these to signal whether the result was useful. You can add a short comment to explain what was wrong.