The unit of value is a worker, not a seat.
A virtual worker is evaluated against what a competent person in that role produces, which is why the objective ladder, the outcome contract and the improvement loop matter more than model quality. It also means the number that matters in a business case is projected cost per worker per month, and the platform is built to tell you that number rather than leave you to estimate it.
Pricing is quoted per engagement, because worker count, channel mix, language coverage and conversation volume vary too widely for a list price to be honest.
| Layer | What you get | Basis |
|---|---|---|
| Hub platform | The runtime, the channel fabric, the calibration subsystem, the control plane, the studio, and the governance and compliance regime | Platform licence, per installation |
| Worker packs | A commissioned worker: definition, agent template, ladders, flows, guardrails, outcome contract, scoring rubric and evaluation suite | Per worker, per seat equivalent, or per outcome |
| Templates, voices and skills | Curated personality, per language voice mappings and licensed skill packs beyond the standard library | Included at base, premium and bespoke chargeable |
| Domain content | The licensed knowledge that makes a worker competent in a named domain | Subscription, on the Knowledge Substrate model |
| Channel, model and usage | Telephony, messaging and model consumption, priced per task class | Pass through plus margin, against an auditable meter |
| Services | Commissioning, knowledge and asset onboarding, methodology capture, calibration tuning | Fixed scope engagements |
| Billing and tax | Handled by Cleverbridge as merchant of record: payment, tax calculation and remittance, invoicing, dunning and refunds | No per jurisdiction tax registration on your side |
Each is an entitlement configuration rather than a manual arrangement, which is what allows a pilot to be provisioned quickly and converted without a rebuild.
A bounded proof against your own accounts, sized so the result is meaningful and the exposure is not.
A commissioned worker running against your live pipeline, supervised by your team, improving on evidence.
Several worker types across teams and markets, with supervision, entitlement management and internal facing workers.
Two meters, both yours to see. Voice minutes, because that is the unit you already budget in and the one carriers bill. Credits, a single currency covering every other metered activity.
What each activity costs in credits is set by the publisher of the worker pack and ships inside the pack itself. What a credit costs in money is set by whoever sells you the deployment. That separation is deliberate: it is what lets a white label, a marketplace listing and a Corvair-operated tenant price differently without anyone forking the engine.
Because the schedule ships in the pack and versions with it, your rate cannot move underneath you. A repricing arrives as an upgrade you approve, itemised against your current schedule, rather than as a change of terms you discover in an invoice.
The schedule is a design instrument, not just a price list. In the property pack an enquiry callback is priced below a general conversation, deliberately, because speed of first callback is the metric that moves performance most and the schedule should not discourage the ten-minute callback. An appraisal lead call is priced the same way, because winning the instruction is the most valuable outcome the agency produces.
The three ceilings behave differently, and the differences matter more than the numbers. Minutes exhausted: conversations in progress finish, no new outbound dial, and text channels continue if credits remain. In property, inbound enquiry callbacks are the last thing to stop, because an unanswered enquiry is a lost viewing. Credits exhausted: discretionary consumption stops first, being simulation, research refreshes, enrichment and match runs, while scheduled work is honoured and the queue holds rather than drops. Licence ceiling: a further worker cannot be deployed, existing workers are unaffected, and the ceiling is visible before provisioning rather than after.
What you never see on your invoice is our cost of goods. Model tokens and audio seconds are our problem, and a platform that shows you its margin instead of your bill has confused two different products.
Entitlement suspends; it does not degrade. A worker at zero credits stops and says so. A worker that quietly gets worse as the balance falls is the least debuggable failure we could ship you.
A worker definition is data, not code. A worker pack is that data made portable, versioned and signed, and it is what ships in every model below.
The worker definition, the template and application prompt layers, the agent template and voice mappings, ladders, flows, guardrails, the outcome contract schema, the scoring rubric, the skill and tool manifests, the knowledge scope declaration, the connector requirements and the entitlement shape.
The kernel prompt layer, which is engine code and is what makes every worker in the family behave like one. And customer data, in any form, ever.
| Deployment mode | What ships | Whose brand | Who operates it | Who is accountable for compliance |
|---|---|---|---|---|
| Direct | A tenancy on the Corvair platform | Corvair | You | You |
| Corvair operated | A configured worker we run on your behalf | Either | Corvair | You, with Corvair as processor, stated in the contract |
| White label | The engine in your own cloud project, under your brand | Yours | You | The distributor |
| Embedded | A single worker pack running inside your product | Your product | You | The distributor, unless the host passed it on in writing |
| Marketplace | Packs, bundled with domain content, data or API entitlements | Publisher branded in a catalogue | Whoever installs | The tenant |
Not a roadmap in order. Five shapes the same engine has to fit. White label turns brand into configuration and needs a stated version support boundary, because a partner runs a version rather than the current one. Embedded workers frequently use the host product's own interface as their channel, which costs nothing architecturally because a channel is an adapter. A marketplace sells certification, not packs: schema and guardrail conformance, the pack's own evaluation suite executed with published results, a compliance review of its declared purpose and data classes, and an authority profile checked for proportionality. A lead qualification worker requesting compensation data does not pass.
The market has lead generation tools and it has call automation tools. It does not have a governed runtime for role holding workers. A buyer does not ask which model you use. They ask what it achieved, how you know, and what happens when it gets something wrong. The Hub has an answer to all three, and each answer is an artefact rather than an assurance.
Every conversation your workers hold makes them measurably better, and the improvement is expressed in versioned artefacts you own and can inspect: prompts, knowledge, assets and skills. That is a switching cost built from delivered value rather than from lock-in, and it is why the calibration loop is priced as part of the platform rather than as an add-on.
The most useful first session is a working demonstration against a scenario from your own pipeline, followed by an honest scoping of what a pilot would involve.