I have a use case with an AI column that I hope someone in the community can help with.
I have a Text column and four AI columns that extract details from it, such as name, phone number, email, and other basic information.
What I want to achieve is simple: if an AI column already contains a value, I do not want the AI to run again for that cell. Ideally, the AI should only process rows where the relevant output field is still empty.
So far, I have tried:
- Triggering the AI column with a button, but AI column refresh does not seem to work well with
If() conditions.
- Using Disable If on the AI column, but it cannot reference the column itself.
- Using a checkbox controlled by a formula to determine whether the AI should run, but then the AI column gets stuck and never returns a result.
Any other ideas?
Of course, one way would be to add a button for each text column, disabling it if the column isn’t empty. However, for my planned table with 100+ AI columns, that would mean adding 100+ buttons… So not really an option.
what about this: use AntiGravity or Claude via the MCP to extract data and to skipp rows that contain already the data like a phone number.
cheers, christiaan
Coda > Claude > Coda to avoid processing the columns that are not empty 
Would you mind guiding me a bit more, please?
sorry to be unclear
i mainly use antigravity and claude and the skill files I wrote to instruct them.
in my skill files i have instructions on how to execute a taks
thus your the harness of your choice instead of the ai columns.
in your scenario, instead of using the ai column logic, you instruct the harness of your choice to fill out the rows and you tell the LLM (via your instructions) to skipp the rows that already contain something.
do you already use the coda mcp? claude and chat gtp have a native connector, rather easy to set up.
Thanks, Christiaan!
That makes sense. I have limited experience with Coda MCP - I’ve tried a few use cases, and the only one that has stuck so far is submitting website data to a Coda table.
Am I right in assuming that the setup would run within the context of the user uploading the content, using that user’s permissions?
well here you touch a tricky subject. for the moment makers have full access to the coda docs via the MCP, but editors not so much.
I wrote about it in a recent blog. and more specically here: 1. Permissions — the wall with no workaround
Cheers, Christiaan
Yeah, this is exactly what I ran into in my previous test with Coda MCP. Most of my docs are used by users—they want a simple 1-2-3 click, and that’s the most I can get from them. In fact, I can’t even expect typing; they need to dictate. So, yeah… the material world is not like a digital creator’s dream.