> ## Documentation Index > Fetch the complete documentation index at: https://docs.synthpop.ai/llms.txt > Use this file to discover all available pages before exploring further. # Voice: trust & compliance > Why you can rely on what the voice agent says and does — determinism, auditability, and PHI safety. The [voice agent](/guides/engagement/voice-agent) runs real conversations with real patients, so "trust me" isn't good enough. This page is the case for why you can rely on what a call says and does: the call runs a **reviewed conversation flow**, every call leaves a record built for audit, patient data stays out of the model's context, and payment handling is isolated by construction. ## A reviewed flow, not an improvising prompt A voice call is not a model improvising over a system prompt. It runs a **reviewed conversation flow**: the sequence of steps a call can take, the tools it can use, and the boundaries between them are fixed and reviewed **before the call is ever placed**. Within that fixed structure, the model still does real work — it chooses the **words it speaks and the values it fills in** (a date of birth it heard, an order quantity the patient confirmed) — but it does not choose **which step comes next or which tools exist**. Those are decided ahead of time, not improvised turn by turn. A question your reviewer can answer before a single call rings is "what can this call do, and what can it not do" — because the flow, not the model's judgment in the moment, is what draws that boundary. This page describes the guarantees behind a call's behavior, not how a flow is authored. If you're evaluating the voice agent for a compliance review, the sections below map to the questions that review typically asks.

Every call is auditable

Every call produces an audit record — not a summary reconstructed after the fact, but a record built as the call happens. It includes: * **A hash of the exact flow version** that ran the call, so you can point to precisely what was reviewed and confirm it's what actually ran. * **Confirmation that no undeclared step-transition or tool use occurred** — the call stayed inside the boundaries the flow declared. * **An independent, automated call-auditor verdict** — a separate automated review reads the transcript against what the flow promised and records whether the conversation held to it. * **The full transcript** of the call, attached to the record. Taken together, that's what "auditable by construction" means here: the record isn't assembled on request when someone asks about a call — it's a byproduct of the call itself, produced whether or not anyone ever asks. Ask your Synthpop contact what's available for a given call. Audit-record contents are part of the platform's compliance posture, not a per-call self-serve export.

The model never sees backend PHI

A call can look up or act on data in your organization's systems through a per-org integration — a patient's record, an order status, an eligibility result. **The model never sees raw backend PHI.** The model decides *to* look something up; the lookup itself runs outside the model's context, and only a non-PHI result — a match, a status, a boolean — comes back into the conversation for the model to act on. That boundary is why a call can reflect live data from your systems without that data ever becoming something the model reads, reasons over, or could repeat verbatim.

Payment security

The voice agent can collect an end-of-call co-pay over the phone. Card data is captured through secure DTMF entry — the touch-tone digits a patient enters on their own keypad — and **never reaches the model or the flow engine**. Neither ever holds, sees, or processes the card number. The autopay gate that authorizes the charge **fails closed**: if it can't confirm the charge should proceed, the call transfers to a human instead of risking a wrong charge. That trade favors a human handling the edge case over any chance of a double charge or a charge against the wrong order. ## Contact limits & call windows Outbound dialing is bounded by guardrails on **EVA's own dialing**: at most once per day, at most three times within a rolling seven-day window, with an optional local-time calling window an organization can configure. A call that would exceed the cap is refused rather than placed — see the [voice agent](/guides/engagement/voice-agent) page for how that shows up in the API. These are guardrails on how often Synthpop's voice agent will call a given patient, not a claim of blanket regulatory compliance for every jurisdiction your patients are in. ## Verbatim disclosures Where a disclosure is legally required, the voice agent speaks the exact wording it was given for that language, **verbatim** — not paraphrased, not summarized, not reworded call to call. The same disclosure reads identically every time it's spoken in a given locale, so what a patient heard is exactly the text your organization approved. ## Where to go next The full catalog of what a call can do, with the guardrail for each capability stated alongside it. The API contract for creating and reading voice Tasks. HIPAA and SOC 2 posture, access control, and data handling for the platform as a whole. How engagement fits alongside intake and coverage.