It's Work, Not Magic
What it actually took to build one boring AI workflow for my team
I just wanted to build a workflow automation. I had a simple use case in mind: we’re kicking off a new project, and I wanted to create a meeting repository in which my team could leverage generative AI to answer project questions and surface insights as we move along. This should include the latest meeting notes, project contracts, and estimates. Since LLMs excel at synthesizing data, this felt like a strong use case. Whatever the solution, it should be resilient, automated, and leverage tools we already use, if possible.
My initial architecture assumed I could integrate two tools I’m already using: Granola for meeting transcription and a Claude project I shared with my team. Add a dash of Zapier and voila, magic! I honestly thought it would be this simple.
Spoiler alert: it wasn’t. I hit roadblocks almost immediately. Turns out, in a shared project in Claude, I can’t connect to a shared Google Drive. Even if that team in Claude has permission to that specific folder in Google Drive. Permissions are complicated! Yes, I could manually add the notes to a Claude project context, but that would require me to actually do that after every meeting, which would defeat the purpose. This is supposed to be easier for me.
So, back to the drawing board. I did a quick survey of all the tools we use or have access to, including Granola, Slack, Claude, Google Drive, Gemini, Trello, and GitHub. Yes, we all use Slack, and the notes are published there, so they are searchable, but we aren’t using Slack’s AI right now.
We could use Granola, broadly, but not everyone on the team has a license, and that would include only the project notes, not other important documents. I could create a GitHub repo, but I ruled that out because not everyone uses Claude code, so interrogating the documents felt a little more difficult than I wanted it to be. Knocking these options down, one by one.
I ended up settling on a different combination of tools than I expected:
Granola for transcription and distilling of meeting notes
Zapier to save new notes to a shared doc on Google Drive
A shared notebook in NotebookLM as the project repository
NotebookLM is one of my favorite, often overlooked, generative AI tools. It works really well for this context. I can pull in resources from so many relevant sources and distill important information really quickly. Win!
There are many ways I could’ve created this solution, and I guess that’s my point. This isn’t a super complicated example, and the roadblocks I hit weren’t dealbreakers. But if you read the headlines and listen to the loud, evangelist crowd, you might think these useful, simple things are fully solved. On the frontier, this is boring and unsexy work.
But this is where the real work is happening right now. Organizations and teams are struggling with mandates requiring them to rethink their processes, survey their existing tools, determine how data will be governed and protected, and still achieve efficiencies from new (and increasingly expensive) tech that is changing by the moment. Case in point, the day I finalized this workflow, Claude’s new Slack Tag feature dropped.
This work is still hard. It’s still worth doing. But it’s work, not magic.