The Short Answer
A hospital health system can run outbound voice AI agents entirely on infrastructure it owns: the agent runtime sits on a hospital-controlled server, the call transcript lands in the hospital's own analytics, and no vendor cloud stores protected health information. The runtime itself costs roughly $4/month on a 2 vCPU VPS. Calls are billed separately at published telephony rates — about $0.013–$0.014 per US outbound minute — plus model tokens.
Self-hosting does not eliminate every Business Associate Agreement. It reduces the BAA surface to the components you actually outsource: the telephony carrier, and a hosted model endpoint if you use one. Terminate your own SIP and run the model in your own VPC or on-premise, and that surface reaches zero.
What does a self-hosted voice AI agent for a hospital actually cost?
A self-hosted voice AI agent has three separable costs, and vendor marketing tends to collapse them into one number. The runtime is the cheap part: a 2 vCPU / 4 GB VPS runs about $4/month, and that figure does not change whether the agent places 10 calls or 10,000.
The calls themselves are metered. At published US outbound rates of roughly $0.013–$0.014 per minute, a four-minute discharge follow-up call costs about $0.055 in telephony. Model tokens for a short structured conversation add a fraction of a cent to a few cents depending on the model.
Here is the arithmetic for a realistic monthly volume — 1,000 post-discharge follow-up calls averaging four minutes:
| Component | Basis | Monthly @ 1,000 calls |
|---|---|---|
| Agent runtime (VPS) | 2 vCPU / 4 GB, flat | $4 |
| Telephony | 4,000 min @ ~$0.014/min | ~$56 |
| Model tokens | Varies by model and call length | Metered |
| Total infrastructure + telephony | Owned stack | ~$60 + tokens |
| Vendor voice AI, per-seat or per-minute | Priced per agent seat or per call minute | Scales with panel size |
The structural point is not the size of the bill. It is the shape of it. Per-seat and per-minute vendor pricing multiplies with the size of your patient panel; infrastructure you own does not. A health system that doubles its follow-up program doubles a $56 telephony line, not a per-seat contract.
Do you still need a BAA if the voice agent runs on your own servers?
Yes — for any component that still touches protected health information, and self-hosting does not automatically make that list empty. This is where most self-hosted voice AI marketing overstates the case, so it is worth being precise.
The telephony carrier carries PHI. If a patient says their name, date of birth, or symptoms on a call routed through a carrier, that carrier is handling PHI and needs a Business Associate Agreement. Major carriers offer them; you still have to execute one.
A hosted model endpoint carries PHI. If the agent's reasoning runs against a third-party API, the conversation text goes to that provider, and that provider needs a BAA. Running the model in your own VPC or on-premise removes that hop entirely.
The platform vendor does not, when you self-host. When the agent runtime and transcript store run on hardware you control, the software vendor is not a data processor. There is no PHI in the vendor's systems to cover.
What self-hosting buys you is a shorter list you actually chose, rather than a mandatory chain of processors. And as we have argued in HIPAA-Compliant AI: Why a BAA Alone Is Not the Answer, a signed BAA is necessary but not sufficient — it allocates liability after a breach; it does not change where your data physically lives.
Where does PHI actually travel during an AI phone call?
During an AI phone call at a hospital health system, PHI travels along four distinct hops, and the deployment model determines who holds it at each one. Mapping the hops is the fastest way to evaluate any voice AI proposal.
| Hop | Vendor-hosted voice AI | Self-hosted on your infrastructure |
|---|---|---|
| Audio path | Vendor's carrier account | Your carrier account, or your own SIP trunk |
| Reasoning / model | Vendor's model choice, vendor's account | Any model — hosted API, your VPC, or on-premise |
| Transcript at rest | Vendor database, vendor retention policy | Your analytics store, your retention policy |
| Audit trail | Vendor dashboard, vendor export | Your logs, queryable directly |
The transcript row is the one that matters most and gets the least attention in vendor evaluations. A call is not an ephemeral event — it becomes a durable record of a clinical conversation. On the ibl.ai platform, that record lands in Analytics → Transcripts alongside every chat conversation, with topic detection, sentiment scoring, and per-conversation cost attribution, on infrastructure the health system controls.
What can a self-hosted voice agent do for a health system?
A self-hosted voice agent at a health system handles the structured, high-volume outbound calling that staff rarely have capacity for. Three patterns come up most often, and all three share one property: the conversation content is PHI, which is exactly why the deployment model matters.
Post-discharge follow-up. AHRQ's Re-Engineered Discharge (RED) toolkit treats structured post-discharge contact as a core component of safe transitions, and CMS's Hospital Readmissions Reduction Program attaches real financial consequence to 30-day readmissions. An agent can place the 48-hour call, ask a fixed set of questions about medication adherence and symptom changes, and escalate to a nurse on any concerning answer.
Appointment reminders and scheduling. Reminder calls are repetitive, scripted, and enormous in volume — the ideal shape for an agent. Self-hosting keeps patient name, appointment time, provider, and location inside the perimeter rather than in a scheduling vendor's database.
Cross-facility care coordination. Multi-site systems lose referrals in the gaps between facilities. An agent can confirm referral appointments and log status updates into the system's own records, so the coordination trail stays internal.
In every case the guardrails matter more than the capability. On this architecture the agent's operating rules are written in plain English in its own configuration — confirm the destination number before dialing, never call a number the user did not name, escalate rather than improvise — and edited by the people who own the clinical process, not by an engineer in a release cycle.
How long does it take to deploy a voice agent on your own server?
Standing up an outbound voice agent on your own server takes about 30 minutes, most of it spent waiting on package installation. We published the complete walkthrough as a tutorial: Give an Agent a Phone Number, with every command in the ibl.ai API repository.
The shape of the work is deliberately small. A VPS with TLS runs the agent runtime. The telephony integration is two files — one Markdown skill describing when to place a call, one shell script that places it. The agent's identity and operating doctrine are four more Markdown files. Everything else is configuration.
Real-time transcription is already part of the platform's call stack: both sides of the conversation stream as text while the call is happening, then persist to Analytics when it ends. That is documented here, along with call style, spoken language, and provider selection.
What should a health system CIO ask before buying voice AI?
Before signing a voice AI contract, a health system CIO should ask five questions that vendor demos rarely cover — each one maps directly to a hop in the PHI path above.
- Where does the transcript live, and who can read it? If the answer is the vendor's database, your retention policy is now the vendor's retention policy.
- Which model runs the conversation, and can we change it? A platform locked to one provider's model means a model deprecation becomes your migration project.
- How does pricing scale with our patient panel? Per-seat and per-minute pricing multiply with volume; owned infrastructure does not.
- Which entities in this architecture require a BAA? A vendor claiming none is either running everything on your hardware or has not thought carefully about the carrier.
- Can we read and modify the agent's operating rules? Clinical escalation logic belongs to the clinicians, in a file they can edit.
ibl.ai is family-owned and operated from New York, NY, and organizations deploy the platform on their own infrastructure — cloud, VPC, on-premise, or air-gapped — with full source-code and data ownership, and the freedom to run any model. For health systems evaluating the broader deployment picture, see Self-Hosted AI for Hospitals and Health Systems and our healthcare solutions overview.