A voice AI agency margin is decided by one thing: who owns the cost floor. On a metered platform, the floor is the vendor per-minute rate and your margin is a markup on it. When you operate the platform on your own infrastructure, the floor becomes your own GPU capacity, and the margin above it is yours to set. Everything else, pricing, branding, packaging, sits downstream of that single fact.
This guide explains why metered resale compresses agency margin over time, the three levers that change the math, and what PeachDesk does and does not ship for agencies today. For the solution overview, see white-label voice AI for agencies.
Why does metered resale compress agency margin?
On a metered platform, your cost floor is the vendor per-minute rate, on every call, forever. You can mark it up, but you cannot lower it, and you cannot move inference to cheaper or local models because the platform owns the model plane. As volume grows, two things happen: clients benchmark your price against the same published headline rate you pay, and your markup becomes the only negotiable line. Margin compresses toward the vendor floor.
The floors are published, and they sit above zero on every tier. Vapi advertises $0.05 per minute for hosting alone, excluding model costs (vapi.ai pricing page, accessed 2026-07-30; see our Vapi alternative comparison). Retell AI advertises a $0.07 blend floor before realistic configuration costs (retellai.com/pricing, accessed 2026-07-30; see our Retell AI alternative comparison). Synthflow starts at about $0.08 to $0.09 per minute, vendor-cited, with effective rates reported higher (secondary reporting, accessed 2026-07-30; see our Synthflow alternative comparison). Reselling any of these means inheriting that floor as your cost of goods.
Fairness notes. Retell AI publishes a full component rate card, the most transparent pricing surface in this set. Vapi supports BYOK, which can zero the model portion of its bill while the hosting floor still applies. Synthflow ships a full white-label programme aimed at agencies, with deeper agency tooling than PeachDesk offers today (synthflow.ai, accessed 2026-07-30). The compression argument is structural, not a claim that these platforms are bad products.
What are the three margin levers an agency controls?
Agency margin on voice AI moves on three levers: the cost floor, the attribution, and the retail price. Lower the floor by controlling where inference runs, prove the margin with per-call cost and price records, and set retail freely per client. A metered-resale platform gives you only the third lever, and weakly. Operating the platform gives you all three.
- Cost floor, via self-hosting. When you run the platform on your own infrastructure, your floor is your infrastructure plus GPU capacity, not a vendor rate. In Local mode there is no per-minute provider charge on self-hosted stages; you pay for GPU capacity, so the inference cost of each additional minute falls toward zero as utilization rises. Frontier mode stays metered for the stages where you want leading cloud providers, and telephony carriers always meter.
- Attribution, via per-call economics. Every metered PeachDesk call records cost of goods, retail price, margin, and sourcing mode. Unit economics stay visible per call and per client organisation, so margin is a recorded number, not a monthly invoice you reverse-engineer. This is the proof mechanism that makes the other two levers auditable.
- Retail freedom, via per-organisation pricing. You set the retail price each client organisation pays, over the true cost the platform records. On the managed Bibha Plan your floor is known in advance, $0.06 per minute, as low as $0.03 per minute at volume (see PeachDesk pricing), so you can price client packages with the margin already decided.
How does self-hosting change the cost floor?
Self-hosting replaces a metered vendor rate with capacity you operate. Inference for locally served stages stops being a per-minute charge and becomes GPU capacity you pay for whether or not a call is running, so average cost per minute falls as utilization rises. Billing is off by default in the self-hosted PeachDesk stack; you still pay your own providers and infrastructure. What disappears is the vendor in the middle of every minute you sell.
The trade is real: you take on operating responsibility for the GPU capacity and the deployment. Two costs never move: telephony stays metered by the carrier in every configuration, and Frontier-mode stages stay metered per use. For the component-level breakdown, see our explainer on voice AI cost per minute.
What does PeachDesk not ship for agencies today?
PeachDesk does not ship sub-accounts, delegated billing, per-client branding, or a partner programme today. Multi-organisation workspaces from one login are an operating convenience, not a reseller programme: each client gets an isolated organisation with its own agents, telephony configurations, and usage records, and tenant isolation is enforced in code. If branded client dashboards are your deciding feature, a platform with a shipped white-label programme, such as Synthflow, is the honest recommendation.
What PeachDesk ships instead is the economic layer underneath: a platform you can self-host and operate, per-call cost and margin attribution, and per-organisation retail pricing. The trade is operational responsibility for economic ownership.
Run the margin math on your own call volumes
Bring your client structure and your minutes to a demo. We will walk through cost of goods, retail price, and margin per call on the model you actually plan to run, self-hosted or managed.