So, another "AI" platform that needs you to bring your own data. Color me surprised.
Here's the gist: Relevance AI for the workflows, Airtable as the "vector database." Because who needs a real database when you have a spreadsheet API, right?
* Customer data lives in Airtable (support tickets, basic profile info).
* Relevance AI pulls it via their Airtable "integration" to answer questions, do some light classification.
* Set up a couple workflows to tag incoming user queries based on past tickets.
It works. Sort of. The latency is noticeable, and you're at the mercy of Airtable's limits. But the real kicker? You're now paying for two services to do the job of one (somewhat) expensive one.
It's a decent stopgap if you're already buried in Airtable. But let's not pretend this is a scalable architecture. It's duct tape and hope.
logs don't lie
logs don't lie
Yeah, duct tape and hope is the default architecture for most of these "AI" stacks now. The double billing is the real joke though.
I tried a similar setup with Google Sheets and Bubble last month. Latency killed it after about 50 records. Airtable's a step up, but you're still building on sand.
You get what you pay for, I guess. Or in this case, you pay twice for half of what you need.
The double billing is the real joke though.
I've seen this pattern three times in the last quarter alone. A team pays for a core CRM, then pays again for Airtable/SmartSheets to "flexibly" extend it, then pays a third time for an AI layer to make sense of the mess they've created. The cost per sales-accepted lead becomes a hilarious horror story you tell at conferences.
The latency isn't just a performance issue, it's a data freshness problem. Your workflow tags a query based on a ticket that was updated 10 seconds ago in Airtable, but the AI pull is on a 90-second cycle. Now you've scored a lead with yesterday's intent. Good luck with that attribution.
MQLs are a vanity metric.
You're right about the duct tape, but I think you're underselling how many places this actually works. The real problem isn't Airtable as a makeshift vector store, it's that Relevance AI's integration is probably just a scheduled pull with zero real-time capability.
I've had to build similar garbage for marketing teams who refuse to leave their spreadsheets. The key is to stop thinking of it as a database and treat it as a slow cache. You pre-compute embeddings on a cron job and dump them into a column, then your workflow just does a stupid similarity search against that. It's still bad, but the latency becomes predictable.
The double billing is the tax you pay for not having an engineer who can stand up a Postgres with pgvector in an afternoon. That's the actual trade-off.
Automate everything. Twice.
Latency isn't the only issue. You're now locked into Airtable's API schema changes, row limits, and rate limiting. One "improvement" from them and your workflows break.
The real cost is the complexity debt, not just the double billing.
Simplicity is the ultimate sophistication