Hire the provider that proves safe resolution on representative calls, not the one with the most human-sounding demo. Define permitted actions, test accents, noise, interruptions, and exceptions, require a reachable human handoff, review consent and data flows, and compare total cost per correctly resolved case. A scripted demo without your telephony, CRM, and policies is not an operations test.
Separate inbound service from outbound calling
Start inbound with narrow, reversible intents such as order status or appointment changes, with human service always reachable. Outbound campaigns need a separate consent, identification, opt-out, and calling-rule review. The FCC has clarified that covered artificial or prerecorded voice restrictions apply to AI-generated voices; this does not mean every AI voice call is categorically forbidden. Obtain counsel's analysis for each use and jurisdiction.
Voice Service Proof-7: seven procurement gates
- 1. Contact rights and disclosure — who is called, why, what is disclosed about AI or recording, and how a person opts out.
- 2. Scope and authority — which answers are informational, which actions change an account, and which require verified identity or human approval.
- 3. Speech understanding — recognition across accents, noise, pace, names, numbers, and code-switching without invented confirmation.
- 4. Dialogue and accessibility — interruption, repetition, adjustable pace, error recovery, and reachable human backup.
- 5. Context-preserving handoff — a live transfer and useful summary without forcing the customer to repeat sensitive details.
- 6. Integration and security — telephony, CRM, least-privilege access, audio/transcript retention, and subcontractor dependencies.
- 7. Outcome and economics — correct resolution, repeat contact, complaints, transfers, residual human effort, and fully loaded cost.
Mark each gate absent, asserted, demonstrated in a controlled environment, or proven on a representative sample. Never average away a failure in consent, identity, human access, or data handling. This MAKINAI evaluation tool is an illustrative procurement method, not a certification or market benchmark.
Run a live-call evidence test
Use an approved, anonymized sample: common requests, regional accents, noisy audio, customer interruption, unclear speech, out-of-scope requests, a fraud attempt, and an unavailable CRM. Replay identical scenarios for finalists. Independently check correct task completion; measure elapsed time including transfer, critical errors, abandonment, and repeat contact in a defined window. Do not expose production customer data without approved purpose, instructions, and safeguards.
Build a cost-per-resolution card
Divide fully loaded period cost by cases correctly resolved without avoidable repeat contact. Include carrier and minutes, speech recognition and synthesis, models, integrations, human review, monitoring, support, and rework. Report containment, resolution, and satisfaction separately: a system can contain calls by making escape difficult while failing customers. Ask for your own baseline rather than vendor-selected industry figures.
Four gates from pilot to contract
- Pilot — one reversible intent, authorized call cohort, human or IVR baseline, and a hard stop for critical errors.
- Launch — task-level acceptance by language and adverse condition plus privacy, security, and accessibility review.
- Operate — named owners for prompts, integrations and evaluations, incident detection, rollback, and retesting after model changes.
- Exit — portable workflows and data, verified deletion of recordings/transcripts, and continuity of phone numbers and human queues.
US diligence is use-specific
Have counsel assess TCPA and applicable state calling/recording rules, sector obligations, consent evidence, and vendor roles. The FTC describes voice-cloning risks including impersonation and fraud. The W3C voice-accessibility document cited here is an early research draft rather than an endorsed binding standard; use it as a source of user-needs test cases, including an accessible human fallback.
Related decisions and next step
For enterprise integration use https://makinai.co/insights/en/how-to-choose-ai-integration-partner-enterprise-systems; for service operations use https://makinai.co/insights/en/ai-service-levels-support-incident-response-provider; for architecture review use https://makinai.co/insights/en/evaluate-ai-solution-architecture-proposal-before-hiring; and for ongoing vendor accountability use https://makinai.co/insights/en/ai-provider-governance-performance-management-before-hiring. MAKINAI can structure the priority call flow and comparative test before vendor award: https://makinai.co/services/en/ai-agents-automation-development.