So the team is sold on the pitch: "build an AI agent workforce" to power your real-time recommendations. Relevance AI looks slick, and the idea of dragging boxes around instead of writing another microservice is tempting. But you've got 15 engineers who presumably know how to code. Before you let the vendor lock-in genie out of the bottle, let's talk brass tacks.
The real question isn't whether you can build a recommendation flow in their studio. You can. It's whether you can afford it at scale, and what happens when you need to do something they didn't put in a box. Their pricing, like most in this space, is a delightful black box of "Workflow Units." You'll need to run some serious projections on your expected inference calls, data processing steps, and user count. My back-of-the-napkin math for a moderately busy app quickly approached "just run your own inference on spot instances" territory.
And then there's the escape hatch. Once you've wired your data pipelines, logic, and models into their canvas, how do you leave? You're not just migrating a service; you're reverse-engineering a proprietary workflow engine. Multi-cloud? You're now on the Relevance AI cloud. Enjoy the view.
I'd be curious to hear from teams who actually pushed it to production. Not for a demo, but for something serving thousands of recs per second. Did the cost model hold? When you needed a custom scoring function or a novel data join, did you have to beg for a feature or build a grotesque workaround?
Beware of free tiers