Just got bit by this. We signed with OpenClaw last quarter for their "unlimited AI agent runs" tier. Sounded perfect for our spiky, unpredictable inference workloads. The sales deck was all rainbows and "scale without limits."
Fast forward to last night's load test. Everything's humming, then... HTTP 429s. Throughput flatlines. Check the dashboard: "Account usage normal operations threshold exceeded."
The hell is that? Buried in Appendix C, Section 8.2 of the MSA:
> "...'Unlimited' usage is provided under the condition that it remains within the bounds of **normal operations** as determined by OpenClaw. Normal operations are defined as usage patterns consistent with the Customer's historical consumption and industry-standard benchmarks for similar-sized organizations."
So "unlimited" means "unlimited, as long as you don't use it in a way we didn't expect." No hard numbers, just a vibes-based quota. Their support basically said we were "operating at a scale atypical for a company of our employee count."
**Red Flags I should've caught:**
* The word "reasonable" or "normal" attached to any "unlimited" claim.
* No explicit, document-defined throttling limits in the SLA.
* The "industry-standard benchmarks" clause—means they can move the goalpost.
**What I'm doing now:**
1. Adding a clause to any future contract: " 'Unlimited' shall be defined as the maximum throughput the technical system can provide, irrespective of customer profile or 'industry benchmarks.' "
2. Logging everything. If they claim "abnormal," I want a 90-day traffic graph from a "normal" customer for comparison.
3. Architecting for instant bail-out. All calls to OpenClaw now go through a proxy that can switch to a fallback provider on specific error codes.
The lesson? "Unlimited" is a marketing word. Your contract defines the reality.
Pager duty survivor.
NightOps