A team uses SH for operational tasks and HappyScribe for meeting transcripts and summaries.
During operations meetings, the team discusses a mixture of new action items, updates to existing tasks, and general discussion. The goal is to use SH AI (Docs AI or AI Columns) to process the meeting output, identify the actionable items, update the relevant existing tasks, and create only genuinely new ones.
Docs AI can create tasks when they are provided individually. However, when the team pastes an entire transcript or summary, it struggles to reliably:
- Distinguish new tasks from updates to existing tasks
- Identify which existing task should be updated
- Ignore non-actionable discussion
- Avoid duplicates or incorrect assumptions
At the moment, manually creating or updating each task is often faster. AI would only add value if it could process the full meeting output through a simple, reliable workflow.
Ideally, a team member could paste or send the HappyScribe summary/transcript into Coda, run an action button (or else) once, and have the relevant task table updated safely.
Has anyone built a practical setup for this without a complex external workflow or requiring the team to manage multiple tools?
Hi Stefan
I tried to do something similar sometime ago, in this case using Google Gemini transcripts as a starting point.
I attach here the Document where you can see the workflow. It is not 100% accurate but I suck at prompting and it was before the improvements in the AI columns models, so maybe with some tweaks it can suit your workflow.
Sorry, but the whole document is in italian…
Hi @Diego_Delfino ,
Your input is highly appreciated! I’ve already identified some promising ideas to explore in one of my attempts. Are you still using this setup?
I have it in my todo list to complete it when I have time. I still need to force the adoption of minutes of meeting inside Coda, as currently the rest of the office still uses Google Docs …
I understand exactly what you mean by user adoption of new processes. Over the past three months, I’ve realized that words like ‘cool’ and ‘new’ only resonate with about 10% of users. Most people actually prefer terms like ‘easy’ and ‘stable.’ Unless we ensure the process works flawlessly without requiring their input, it simply won’t gain traction at scale.
So true!
Anyway, i’m Interest to know where this testings will lead you! 
I’ll keep you posted if I have any luck. So far, I’ve managed to run out of credits within a couple of hours, and the AI Column still seems highly unreliable, even with very specific prompts.
One interesting finding is that HappyScribe’s extraction of action items is significantly better than anything we’ve been able to achieve with our prompts so far. I also applied Max’s approach of breaking the prompts down into smaller pieces to make the process more deterministic. That helped with consistency to some extent, but, as Coda Support confirmed today, chaining AI Columns—where one is triggered after another finishes—is practically unusable for longer documents. In one test, it took 5+ minutes to process a single cell for one column.
The most promising approach so far has been to:
-
Copy the action items extracted by HappyScribe into a separate column. They’re already fairly well-defined, just not in the context of the rest of the data.
-
Add the full script purely as context to the AI prompt.
-
Ask the AI Column to review the action-item list, identify anything that’s missing based on the full script, and match each item to the relevant table references using a formula that points to the data tables.
This approach was actually looking pretty promising. Until I ran out of credits about 20 minutes later 
One of the reasons I still use Coda instead of tools like ClickUp, Monday, or Notion is precisely because it doesn’t lock me into a specific ecosystem. That said, I’m not hyped from such moves. Instead, I would’ve preferred polishing the existing product and making it more useful and bug-free.