Hi everyone,
I’ve been evaluating Relevance AI for a potential team rollout, and I’m at the stage where I need to really dig into the cost implications. Their per-query pricing model is transparent, which I appreciate, but I’m trying to project costs for a steady, high-volume workflow.
For those of you running dozens or hundreds of automations or agent workflows daily, how has the per-query cost held up? I'm particularly interested in cases where you’re chaining multiple AI steps (like a research agent that does several web searches and summarizations per single user request). Does the cost compound in a way that made you reconsider your architecture?
Also, have you found the tiered usage discounts to be meaningful once you hit the higher brackets? Or did you end up needing a custom enterprise plan sooner than expected?
I’d love to hear about real numbers if you’re comfortable sharing—not just “it got expensive,” but what scale triggered a cost review and how you managed it. This kind of feedback is so valuable for the rest of us trying to plan responsibly. 😊
~nancy
Per-step pricing is the killer here, especially for chained agents. If a single user request triggers 5 AI steps, you're paying for 5 queries. At high volume, that multiplier destroys any perceived transparency.
Tier discounts rarely help when your cost driver is complexity, not raw query count. You hit a higher bracket faster, but your per-request cost remains linearly tied to steps.
We hit a review at ~50k multi-step workflows/month. The bill was scaling directly with feature use, which is terrible for forecasting. We had to rewrite to batch logic outside their platform and call their API only for unavoidable steps.
cost per transaction is the only metric
You've highlighted the core challenge with per-step pricing models. The cost multiplier isn't just theoretical, it becomes a primary architectural constraint.
In our TCO analysis for a client running customer support triage agents, each ticket triggered an average of 3.7 AI steps. At 10,000 tickets monthly, that's 37,000 queries. The tier discount at that volume was a 15% reduction, but the total cost was still 3.7x the base "per request" expectation. The financial model became untenable because cost scaled directly with process complexity, not business output.
We managed it by implementing a pre-processing layer to collapse multiple logical steps into a single, more complex prompt where possible, reducing the average steps to 2.1. This is the necessary workaround user170 hinted at. The tier discounts only help on the margin, they don't change the fundamental cost driver.
Trust but verify.