How to Audit a Voice AI Platform Before Committing
Audit a voice AI platform by testing whether your agency can price, launch, support, and retain clients on it without surprise costs or avoidable handoffs. The 10 checks below cover operating model, total cost, compliance, reliability, latency, support, branding, APIs, trial terms, and company stability, with a practical test for each.
A polished demo proves that one call can work. It does not show what happens when a healthcare prospect asks for a BAA, ten clients need separate workspaces, or a failed booking reaches support. Use this checklist before moving a client, signing a long commitment, or standardising your agency on a platform. For the broader buying framework, see the white-label voice AI platform guide for agencies.
1. Operating Model: Who Owns the Client Outcome?
Choose an operating model your agency can support profitably: count the systems, accounts, bills, credentials, and support teams required to launch and run one client. Architecture matters because it determines those day-to-day handoffs, not because a credible vendor must build every carrier, speech model, language model, and cloud service itself.
Three questions separate the main models: who owns the agency-facing product, what platform runs the core voice agents, and whether the agency must maintain another vendor account or bill. Vapify, Voicerr, and ChatDash document a classic wrapper model in which the agency also connects or funds an upstream voice platform. VoiceAIWrapper is now a hybrid: as of October 2026 its built-in engine runs on one VoiceAIWrapper invoice with no upstream account, while its optional bring-your-own-key path connects Vapi, Retell, or ElevenLabs and is billed by that provider. GoHighLevel is different: its public subprocessor list names Retell, but the product, API, billing, and support path appear to stay inside HighLevel. ConvoCore documents both a Vapi-connected route and a separately managed voice stack, so its answer depends on the route you choose.
A first-party platform can still use contracted telephony, speech, model, and infrastructure providers. What matters operationally is whether the vendor owns and supports the application you sell, and whether it can resolve a client issue without sending your agency to a separately contracted voice-platform provider.
Green flags: The vendor maps the full responsibility chain, identifies any account your agency must open elsewhere, explains who bills each usage component, and owns the support escalation through resolution.
Red flags: A second provider account appears only after signup, the quoted price excludes a required upstream voice engine, credentials must be copied between vendors without a documented support boundary, or each vendor can send you to the other when a call fails.
What to check: Ask for a one-page diagram of the commercial and support path for one live client. Then create a test agent, trace every account and bill it needs, and compare incident history with the voice AI wrapper versus native platform guide.
2. Pricing Transparency
A platform's pricing page should tell you exactly what you will pay at 5, 10, 20, and 50 clients without requiring a sales call. The most common pricing traps in voice AI are: hidden per-seat fees that scale with client count, tiered plans that force expensive upgrades at arbitrary thresholds, and separate charges for compliance features that should be included.
Calculate your total cost at three scales: current client count, 2x current, and 5x current. Include base plan fee, per-minute usage, per-seat or per-sub-account fees, phone number costs, compliance add-ons, and integration fees. If the pricing page does not give you enough information to complete this calculation, that is itself a red flag.
Green flags: Published platform, usage, telephony, phone-number, workflow, and workspace terms. Clear feature gates between plans. Regulated-use requirements and any contract-specific costs are stated before a client is sold.
Red flags: A self-serve plan cannot be priced from the public page. Per-seat fees compound without warning, required provider charges are omitted, or usage rates and regulated-use terms change between tiers without clear documentation. Quote-based enterprise pricing is not a red flag by itself when the deployment is genuinely bespoke.
What to check: Build a spreadsheet with three scenarios. If you cannot fill in every cell from the pricing page alone, the platform fails the transparency test. Compare your results against published white-label pricing breakdowns to verify your numbers are in the right range.
3. Compliance Certifications
Compliance must match the client's data, call type, jurisdiction, and contract. There is no general "HIPAA certification": a healthcare workflow involving protected health information normally requires an executed BAA and an agreed covered scope, while SOC 2 Type II and ISO 27001 provide different forms of security assurance. GDPR documentation, call-consent rules, TCPA controls, retention, residency, and audit requirements must be evaluated separately.
The difference between "we are working toward compliance" and "here is our SOC 2 Type II audit report" is the difference between a promise and a fact. Ask for documentation, not claims.
Green flags: Current certifications and audit scope can be verified. The vendor can provide its DPA and, for an eligible healthcare deployment, explain the BAA process and plan requirements. Contract-specific controls such as data residency, private cloud, SSO, retention, or an SLA are written into the applicable order form or agreement.
Red flags: The vendor uses a badge as a substitute for contractual scope, calls itself "HIPAA certified," cannot produce the promised DPA or audit evidence, or implies that every plan may process regulated data without explaining the required agreement. Enterprise-only controls are not a red flag if they are disclosed before the sale.
What to check: Write down the client's data types and jurisdictions, then request the relevant audit evidence, DPA, BAA process, subprocessor list, retention terms, and contract language. Do not infer that a document does not exist from a slow sales response; verify the evidence before promising the capability to a client. The white-label compliance guide explains how these documents fit together.
4. Uptime SLA
An uptime claim protects your client only when the contract defines what is measured, which exclusions apply, how incidents are reported, and what remedy follows a breach. Set the target from the client's business impact: an after-hours receptionist and a hospital contact centre may need very different commitments.
Green flags: Financially backed SLA with defined service credits. Public status page with historical uptime data. Uptime measured at the call-completion level, not just server availability. Planned maintenance windows disclosed in advance.
Red flags: "99.9% uptime" on the marketing page with no contractual backing. No public status page. Uptime measured by server ping rather than successful call completion. History of extended outages visible on community forums but not acknowledged on the status page.
What to check: Visit the vendor's status page and review the last 90 days. Count incidents, duration, and resolution time. Ask for the SLA document (not the marketing page) and look for the service credit table.
5. Latency Benchmarks
Voice AI latency is the time between when a caller finishes speaking and when the AI begins responding. Anything under 2 seconds feels natural in a phone conversation. Anything over 3 seconds creates awkward pauses that make callers hang up or lose confidence in the interaction.
Latency has two components: AI processing time (how fast the model generates a response) and telephony overhead (the time it takes for audio to travel through the phone network). Vendors that quote only AI processing latency without telephony overhead are giving you an incomplete number. End-to-end latency, measured from the caller's last word to the first word of the AI response, is what matters.
Green flags: Published end-to-end latency numbers (not just AI inference time). Latency benchmarks under 2 seconds. Latency data measured under load, not just single-call testing. The ability to test latency yourself during a trial period.
Red flags: Latency claims that reference only "AI response time" without telephony. End-to-end numbers that are implausibly low for the full telephony path (a complete round trip includes network and telephony overhead, so treat suspiciously low all-in figures with scrutiny rather than as an automatic disqualifier). No latency data published at all. Demo calls that sound fast but production calls that lag.
What to check: During your trial, make 20 test calls at different times of day (including peak hours) and time the response delays yourself. Compare your measurements against published latency benchmarks across platforms.
6. Support Channels
Support quality determines how fast you can resolve client-facing issues. An agency with 10 clients running voice AI agents cannot afford to wait 48 hours for an email response when a client's agent stops booking appointments correctly. The support structure should match the urgency of voice AI failures, which are immediately visible to your clients' customers.
Green flags: Dedicated support channel (Slack, direct line) for agency plans. Response times measured in hours, not days. Access to engineering staff, not just tier-1 support reading scripts. Community resources (playbooks, templates, live Q&A) that reduce support dependency.
Red flags: Email-only support on all plans. Discord server as the primary "support" channel (community help is not vendor support). "Priority support" that still takes 24-48 hours. No escalation path to engineering. Support staff cannot reproduce or diagnose technical issues.
What to check: Submit a technical support ticket during your trial and measure the response time, resolution quality, and whether the responder understood the issue. Check community forums or Discord for unresolved complaints older than 7 days.
7. White-Label Depth
White-labeling is not binary. Some platforms let you swap a logo. Others give you a fully branded client experience under your own domain with custom emails, branded dashboards, and minute markup controls. The depth of white-labeling directly affects whether clients perceive your agency as the provider or as a reseller of someone else's product.
Green flags: Custom domain support (clients visit youragency.com, not platform.com). Custom branded emails from your domain. Branded client dashboards with your logo and colors. Custom minute markup (you set your own per-minute price). Clients never see the underlying platform brand anywhere.
Red flags: Logo swap only, no custom domain. Platform branding visible in emails, invoices, or dashboard footers. No client-facing dashboard (clients must contact you for everything). White-label features locked to the highest pricing tier. Branding customization requires CSS or code changes rather than a settings panel.
What to check: Sign up for a trial and walk through the full client experience: onboarding, dashboard login, email notifications, call reports, and invoices. Note every place you see the platform's brand instead of yours and confirm that each required branding control is included in the plan you will buy.
8. API and Integration Quality
API quality determines whether you can automate client onboarding, build custom integrations, and scale beyond what the platform's dashboard supports. A platform without an API locks you into manual workflows for every new client. A platform with a poorly documented API wastes engineering time on trial-and-error integration.
Green flags: A documented API, event webhooks, scoped authentication, usable examples, and published rate limits. The platform either has the integrations your clients need or documents how to build and support them without pretending every API connection is a native connector.
Red flags: "API coming soon." Documentation that is incomplete, outdated, or auto-generated without examples. No webhook support (you must poll for updates). API access locked to Enterprise plans. No sandbox or test environment. Zapier listed as the primary integration method with no native API.
What to check: Open the API documentation during your evaluation. Try to answer three questions from the docs alone: How do I create a new sub-account? How do I deploy an agent? How do I retrieve call data? If any of these require a support ticket to answer, the API documentation is insufficient.
9. Trial, Cancellation, and Exit Terms
A useful evaluation period lets you test a realistic client configuration before the first paid renewal or long commitment. A money-back guarantee is one model; a card-required free trial with clear limits is another. Judge the access, cancellation deadline, data export, number ownership, and migration path rather than rewarding one label.
Green flags: The trial or pilot exposes the features you intend to sell, its credits and restrictions are clear, cancellation can be completed before billing, and the vendor documents how to export configurations and call data or move phone routing.
Red flags: The trial hides key agency features, phone numbers become chargeable without a clear notice, cancellation requires an undisclosed process, or the vendor controls a client's number or data without a documented exit route. A short trial is a constraint to plan around, not automatic evidence of a weak product.
What to check: Read the terms and trial limits, schedule cancellation reminders, and test one full client workflow before billing begins. Document what you would need to export or reroute if the platform fails the audit. Also check how much notice the vendor gives before a price change: Trillet, for example, bills Studio and Agency monthly with no contracts and gives at least 30 days' notice of price changes unless a signed agreement states otherwise.
10. Company Stability
Voice AI is a young market, and recent history shows how fast a platform can vanish from under the agencies that built on it. PlayAI's voice platform was wound down after Meta acquired its team in 2025 and was fully retired on December 31, 2025. Air.ai is now defunct after the FTC's March 2026 settlement permanently banned the company and its named owners. Pricing can move just as abruptly: Voicerr's agency pricing has been volatile, so check the live page and governing terms before building a client offer around it. An unstable platform can force your agency to rebuild every client deployment at once.
Green flags: Consistent product releases, transparent incident communication, named customer evidence, clear legal ownership, and commercial terms that do not depend on an obviously temporary discount. Funding can support continuity, but it does not substitute for delivery history or a migration plan.
Red flags: Material price or entitlement changes without notice, repeated unfulfilled launch claims, unclear ownership, no usable customer evidence, or a quiet changelog while support complaints accumulate. A young company is not automatically unstable, but it needs stronger operational evidence.
What to check: Search for the company name plus "shutdown," "lawsuit," "price increase," and "outage" to surface issues the marketing site will not show you. Check LinkedIn for team growth or contraction. Review the company's blog or changelog for consistent product updates over the past 12 months.
The Audit Scoring Framework
Score each area from 1 to 3: 1 means a material gap, 2 means acceptable with a documented constraint, and 3 means the vendor has supplied evidence and your test passed. Do not treat the total as a certification. Set dealbreakers for the client types you plan to serve, then compare platforms on the operating outcomes that affect your margin and retention.
| Audit Area | Weight | What a Score of 3 Looks Like |
|---|---|---|
| Operating model | High | Clear owner, accounts, bills, and escalation path |
| Pricing Transparency | High | Full cost calculable from public page |
| Compliance | High | Required evidence and contract path are documented |
| Uptime SLA | Medium | Workload-appropriate commitment and incident history |
| Latency | Medium | Sub-2s end-to-end, published benchmarks |
| Support | Medium | Tested response and a defined escalation owner |
| White-Label Depth | Medium | Client journey matches the agency's plan |
| API Quality | Medium | Complete docs, webhooks, required integrations |
| Trial and Exit Terms | Low | Realistic test access and a documented exit path |
| Company Stability | High | Operating history and consistent product delivery |
Treat a score of 1 in operating model, pricing transparency, required compliance, or stability as a stop until the vendor resolves it. A platform can be suitable for a local-service receptionist and still be unsuitable for a regulated workflow. Once you are live, apply the same discipline to your delivery by tracking the client success metrics that show voice AI ROI.
Run this audit on Trillet too. As of September 2026, Studio is $99 per month with 1,000 included minutes, three phone numbers, branding on a Trillet-hosted address, and a hard cap of three workspaces. Agency is $299 with 3,000 included minutes, ten phone numbers, unlimited workspaces, a custom domain, branded emails, custom minute markup, and priority support. Both plans have a seven-day card-required trial and charge $0.12 per AI minute after the allowance; telephony is separate. Trillet displays SOC 2 Type II and ISO 27001 certifications, while BAAs are available on Agency and Enterprise through the standard process. On the standard hosted service, the agency works with Trillet and Trillet manages its subprocessors. Compare the current tiers on the white-label pricing page and test the Trillet white-label platform against every check above.
Frequently Asked Questions
How long should I evaluate a voice AI platform before committing?
Use the full trial or pilot window and complete at least three realistic client configurations, using simulated data unless a live client has approved the test. Test different call conditions, support response, branding, billing assumptions, and one integration. A demo call alone is not an agency-level evaluation.
What is the most important factor when choosing a voice AI platform for an agency?
The operating model is the most important factor: your agency needs to know which company owns the product, which platform runs the calls, how many accounts and bills are required, and who resolves a failure. Then test total cost and the compliance requirements of the clients you intend to serve.
Should I choose the cheapest voice AI platform?
Not necessarily. The cheapest per-month platform often has hidden costs that exceed more expensive alternatives at scale: per-seat fees, compliance add-ons, limited sub-accounts requiring plan upgrades, and support gaps that cost you client relationships. Calculate total cost of ownership at your target client count, not just the base monthly fee.
How do I verify a vendor's compliance claims?
Start with the client's data and jurisdiction, then request the evidence that applies: an audit report or trust-centre evidence, a DPA, a subprocessor list, retention and residency terms, and the BAA process for an eligible HIPAA workflow. Confirm that the plan and contract you will buy actually cover the proposed use case.
What happens if my voice AI platform shuts down?
Migration time depends on the number of agents, integrations, phone-number ownership, data-export options, and client approvals. Reduce the disruption by documenting configurations, keeping exportable prompt and knowledge assets, knowing how numbers are routed, and testing a fallback before you need it.
Updated for October 2026: reclassified VoiceAIWrapper as a hybrid with a built-in engine and an optional bring-your-own-key path, and based the GoHighLevel Retell reference on its public subprocessor list.
Updated for September 2026: Reframed architecture around agency accountability, added embedded and hybrid platform models, corrected compliance terminology, and aligned Trillet's plans, trial, and feature gates with the current white-label offer. Also added a price-change-notice check to the trial and exit terms section.
