I've looked at the Relevance AI docs and the pricing page. The explanation is, frankly, a mess of marketing terms that obscures the actual cost drivers. They talk about "AI units" and "workflows," but I need to map this to real usage and a real invoice.
From what I can piece together:
* The core charge seems to be for **"AI actions."** These are individual steps inside a workflow that call an LLM (like GPT-4) or use a tool (like a web search).
* A "workflow run" is the execution of a chain of these actions. So one run could cost multiple AI action units.
* Different models and tools cost a different number of units per action. A simple GPT-3.5 call might be 1 unit, while a complex GPT-4 with a large context could be 10+.
My specific questions are:
* Is there a base platform fee plus the AI unit consumption, or is it purely consumption-based?
* What is the actual price per AI unit? The site only shows monthly bundles.
* Are there separate charges for data storage, workflow editing, or user seats?
* Can you provide a concrete example? If I have a workflow with three AI actions (a classification, a summarization, and a web search) and I run it 1,000 times in a month, what is my bill?
I'm evaluating this against building with raw API calls or using other orchestration tools. The abstraction is only valuable if the pricing is transparent and the ROI is clear. Right now, it feels like I'm being asked to buy "units" of something without knowing the factory cost.
show me the numbers
Yeah that's exactly where I got stuck too. The marketing terms make it really hard to budget.
So if I understand your breakdown, each time a workflow triggers, you're paying for all the little steps inside it? That could add up so fast. I'm looking at a simple customer support bot idea, and now I'm worried.
Do you know if there's a free tier or a trial to test this before committing? I can't even guess my monthly units without that.
Spot on about the forecasting being critical. That's the bit that'll catch you out in production.
I ran a proof-of-concept last quarter. The overage fees, like you hinted, are brutal. You think you're buying a buffer, but the moment you tip over your bundle, your effective cost per unit jumps. It turns a predictable linear cost model into something with a nasty little step function.
My advice? Don't rely on their dashboard estimates. Build your own meter by logging every action type and its unit cost in your own system for a week. Their "AI units" are just a proxy for your actual LLM API costs, marked up. You need to reverse-engineer it.
That breakdown makes a lot of sense, thanks for writing it out. So basically, if my workflow has three steps, I pay for three actions every single run. That feels obvious now, but I totally missed it at first.
You mentioned mapping this to a real invoice. Is the unit cost for each action type published somewhere clear, like a table? Or do you have to guess based on the bundled plans? I'm trying to do the math for a simple two-action workflow.
Exactly right - each step that uses an LLM or a tool is an action, and you're billed for each one per run.
The unit cost per action type is buried, but it's there. You'll find it under "Action Pricing" in the docs, not on the main pricing page. It's usually a table showing things like:
- `llm/gpt-4o` = 6 units per 1k tokens
- `llm/gpt-3.5-turbo` = 1 unit per 1k tokens
- `tools/web_search` = 4 units per call
So for your two-action workflow, you'd need to know which models/tools you're calling and estimate your token usage. The math gets real, real quick. 😅
What's the workflow you're sketching out? Maybe we can ballpark it.
Build fast. Fail fast. Fix fast.
Absolutely, that opaque unit-to-dollar conversion is where the real budgeting pain starts. Your example math is spot on - it forces you into a weird game of reverse-engineering your own effective rate.
And the moment you cross that bundle threshold, the cost structure changes entirely. It's not just a higher per-unit cost for overages; your *entire* usage for that period often gets charged at the overage rate, not just the excess. That step function can double your bill overnight if you're not watching daily consumption like a hawk.
I've resorted to building a simple external monitor that pings their usage API and alerts me at 80% of my bundle. Without that, you're flying blind until the invoice lands.
pipeline all the things