Choose the work you can reliably own after launch. A public Vapi forward-deployed engineering role explicitly includes production integrations, telephony/SIP, webhooks, debugging and deployment handoff. That is evidence of implementation work around the platform—not evidence of any particular customer’s cost or delivery time. (jobs.ashbyhq.com)
Use this proposed responsibility comparison in the statement of work:
| Responsibility | Component-based build | Configured platform | Managed application/service |
|---|---|---|---|
| Business policy and permitted actions | Buyer must own policy | Buyer must own policy | Buyer must approve policy |
| Speech and call orchestration | Assign to internal engineers and infrastructure suppliers | Confirm the platform boundary | Confirm the service boundary |
| CRM/PMS/dispatch integration | Buyer engineering or appointed integrator | Buyer, partner or supplier—name one | Require an explicit included scope |
| Outcome verification | Buyer-defined evidence and reconciliation | Do not assume the dashboard proves completion | Require access to underlying evidence |
| Change testing | Buyer-owned release process | Agree responsibilities for model and workflow changes | Contract for regression tests and approval rules |
| Human fallback | Buyer-operated or separately contracted | Buyer-operated or separately contracted | Verify whether staffing is actually included |
| Incidents and recovery | Buyer coordinates dependencies | Name the incident lead and escalation chain | Require one accountable service lead and dependency disclosure |
| Exit | Preserve code, configuration and accounts | Secure export and migration rights | Secure records, numbers and transition assistance |
A managed service deserves a premium only if the contracted work is valuable to you. A platform deserves a lower apparent price only if you account for the work retained internally.
Does speech-to-speech beat a transcription-based pipeline?
Do not make architecture a substitute for an outcome trial. Twilio documents a text-mediated interface in which caller speech becomes messages and application responses become speech. Amazon’s Nova Sonic materials describe speech capabilities and external tool use. Those are different technical routes, not proof of a universal quality winner. (twilio.com)
| Architecture choice | What to investigate | Buyer acceptance test |
|---|---|---|
| Speech recognition → text reasoning → speech synthesis | Component configuration, intermediate records and behaviour across the full pipeline | Names, numbers, interruptions, tool accuracy and end-to-end response timing |
| Speech-to-speech model | Audio behaviour, available logs and control over tool execution | The same outcome and interruption tests, plus evidence sufficient to investigate failures |
| Constrained workflow around either architecture | Explicit states, permitted actions and server-side validation | Attempts to skip required checks, change identity or induce an unauthorised action |
Our recommendation for booking and account-changing workflows is to enforce authorisation and business constraints in the application or integration layer. A prompt should not be the only barrier between a caller’s request and a consequential system update.