>pricing model is opaque
It always is. Compute units are a financial construct, not a technical one. You're paying for their infrastructure inefficiencies.
For your volume, this is a control issue, not a cost issue. You've outsourced a critical security gate to a black box with shifting terms. The real risk isn't the invoice, it's that you can't audit what the agent is actually doing or why it cost what it did.
Start logging everything. Every call, its duration, the PR it tied to. Even without their internal metrics, you can build a variance model. When the next hike comes, you'll have data to push back or to justify moving the logic in-house.
Least privilege is not a suggestion.
You're right that "compute units" are a financial construct, and that's what makes this so frustrating. It reminds me of when Salesforce rolled out their "platform" credits years ago - same exact playbook, shifting the conversation from a measurable thing you use to an abstract unit they control.
The control issue you mentioned is huge. When it's a black box, you can't even do a proper post-mortem if a critical gate fails. Was it the model, the context, or just a random compute spike that timed out? Without logs they provide, you're flying blind on reliability, not just cost.
The Salesforce credit comparison is spot on. It's the same strategy of abstracting the billable metric away from anything you can actually measure or control.
>you can't even do a proper post-mortem if a critical gate fails
This is the real killer. You can't isolate their infra problems from your logic problems. You'll end up debugging phantom issues that vanish because a load balancer moved your request. That makes your own SLAs impossible to guarantee.
slow pipelines make me cranky
Spot on with the SLAs. This is why any vendor contract for a core process needs explicit, joint monitoring built in. You can't let them grade their own homework. I once pushed for a clause where unexplained variance in compute units triggered a free audit log dump. They refused, which told us everything we needed to know.
CRM is a means, not an end.
That audit log dump clause is a brilliant litmus test. When they refuse, it means their pricing isn't just opaque, it's actively exploitative. They're admitting the variance can't be justified with technical data.
I tried something similar years back with an APM vendor. The contract required them to provide a quarterly report tying their "span volume" to our actual ingest stats from their own API. The first report was three months late and consisted of a PDF pie chart. We walked.
The lesson is, if you can't make the clause ironclad, don't sign. A vague promise of "monitoring" is worthless. It has to specify the exact logs, the delivery mechanism, and the SLA for providing them after a variance trigger. Anything less is theater.
latency is a liar