Platforms Don't Give You Tenant Isolation for Free
Twilio limits concurrent active calls and calls-per-second (CPS) per account, including subaccounts. The default is 1 CPS per account, and self-service console requests can raise Programmable Voice up to 30 CPS and SIP Trunking up to 15 CPS. Subaccount CPS and concurrency don't roll up into the parent account or other subaccounts, which makes a subaccount boundary the cleanest tenant split available.
LiveKit Cloud works the other way: limits apply per project. The Build plan caps concurrent agent sessions at 5, Ship at 20, and LLM calls get separate requests-per-minute (RPM) and tokens-per-minute (TPM) ceilings. Run several customers' callbots on one LiveKit project, and that ceiling applies to the whole project, not per tenant.
The Noisy-Neighbor Failure: One Tenant's Burst Starves Another
Share a platform's CPS pool without subaccounts, and one tenant's outbound campaign queues behind another tenant's urgent call in the same FIFO. Twilio's own docs say a 1-CPS account's queue can hold calls for up to 24 hours — as many as 86,400 a day — and nothing inside that queue tags which tenant a call belongs to.
The same pattern repeats at the LLM layer. Without a per-tenant RPM budget, one tenant's bulk dialing can occupy the shared LLM request queue and push up response latency for every other tenant's live call.
From Design to Ops: A Multi-Tenant Callbot Loop Isolation Checklist
Define tenant boundaries first. A new multi-tenant callbot platform hardcodes per-tenant CPS, concurrency, and LLM RPM quotas as a share of the contracted ceiling. If the platform-wide contract is 30 CPS, cap any single tenant at 40% of that — 12 CPS — so three tenants spiking at once can't let one of them exhaust the whole pool.
On Twilio, issue each tenant its own subaccount to separate CPS, concurrency, and caller ID. On LiveKit, put a token bucket keyed by tenant ID in your application layer to re-split the project-wide ceiling. In front of the shared LLM API, add a per-tenant semaphore that reserves 70% of contracted RPM for the callbot loop and leaves the rest for text chatbots and batch jobs.
When one tenant exceeds its cap, reject new call creation immediately or demote it to a low-priority queue — but keep that queue tenant-specific. A single shared queue lets a bursting tenant's calls pile up first, so another tenant's urgent call doesn't dial out for hours.
When a tenant's LLM semaphore is exhausted, fail only that tenant over to a scripted fallback response and keep every other tenant's requests on the normal path. Without isolation, one tenant's outage cascades into a platform-wide fallback switch, making the blast radius impossible to scope.
Before shipping, deliberately load one tenant to 150% of its cap and verify other tenants' latency and failure rates don't move. Tag every call log and LLM call log with a tenant_id field, and mask PII such as phone numbers and recordings regardless of tenant. Ship only when cross-tenant impact is zero and the isolation-test pass rate hits 100%.
Every week, roll up per-tenant CPS, concurrency, and LLM RPM utilization. If one tenant repeatedly tops 80%, that's the trigger to renegotiate its quota or split it onto a dedicated subaccount. Quota changes touch contract terms, so base the decision on at least two weeks of utilization trend, not a single spike.
A Checklist You Can Use Today
In a multi-tenant callbot loop, the limits carriers and LLM APIs enforce apply at the account or project level, not automatically per tenant. Hardcode each tenant's quota as a share of the contracted ceiling, split queues and semaphores by tenant, and require a 100% isolation-test pass rate before shipping — that's what keeps one tenant's burst from eating into every other tenant's call quality.
References
Twilio — Understanding CPS and Concurrent Call Limits in Twilio
LiveKit Cloud — Quotas and limits
Ask AI about this article
The assistant has read this article. Ask anything — it answers from the text and says so when something isn't in it.
Loading the chat…