In short:
- CareBill Ally uses automated software to read bills, check numbers and make calls. On calls, it may use an AI model (from Anthropic, OpenAI or Google) to understand representatives, under strict rules.
- Dollar amounts are checked with fixed math rules. The software does not guess.
- Every finding links to the page it came from, so you can check it.
- You confirm key numbers, and you approve letters and big decisions.
- A person on our team reviews hard questions, and you can ask for a person at any time.
- Our automated phone assistant always says it is automated and never pretends to be you.
1. Why we publish this notice
Roy Digital Ventures LLC, which runs CareBill Ally, wants you to know exactly which parts of the service are automated, what they can and cannot do, and how people stay in control.
2. The AI we use
CareBill Ally uses these artificial intelligence technologies:
- Speech recognition on automated phone calls: our phone provider, Twilio, turns spoken answers into text in real time.
- A computer-generated (synthesized) voice that speaks the assistant's words. It does not copy any real person's voice.
- One generative AI model, only if we have turned it on: Anthropic's Claude, OpenAI's GPT models or Google's Gemini (we use one at a time and keep the same rules for all of them). It helps the phone assistant understand what a representative says and choose a natural reply. See "Rules the AI must follow on calls" below.
- Microsoft's document reading service, only if we have turned it on, which turns photos and scans into text (section 3).
None of them creates voiceprints. Everything else, including checking the numbers and deciding what to request and what may be shared, is rule-based software that follows fixed, tested instructions.
Rules the AI must follow on calls
When an AI model is used, our software checks every turn before and after the model, and the fixed script always handles the important parts:
- The opening announcement, the permission and recording questions, and the answer to "Are you a person or a robot?" always come from the fixed script. The assistant always says it is an automated AI assistant and never pretends to be you or a person.
- The AI model never receives your name, date of birth, or account or claim numbers. When the representative needs one of them, the fixed script says it, and only if your Case Assistance Authorization allows it.
- Anything the model writes that contains an identifier or value we did not allow is blocked and replaced with: "I'm not able to provide that information."
- The assistant cannot agree to payments, settlements or payment plans, sign anything, or give card or bank details. It says: "I'm not able to agree to payments, settlements or payment plans, sign anything, or give card or bank details on the patient's behalf. Could you send the options in writing so the patient can decide?"
- It does not give medical or legal advice.
- If the model is slow or unavailable, the fixed script answers instead.
- The AI provider (Anthropic, OpenAI or Google) may not use call information to train its models. We keep a text transcript of each call with the patient's case (see the Recording and Transcription Notice); it is never sent to an AI provider after the call or used to train models.
If we change how we use AI in a way that matters to you, we will update this notice before the change is used with your information.
3. What is automated
| Function | How it works | What you or our team do |
|---|---|---|
| Sorting documents | Rules look for key words to decide whether a document is a bill, an insurance statement, a Medicare Summary Notice, a denial letter or an assistance policy | You can correct the type |
| Reading key details | Digital PDFs are read on our servers. Fixed patterns find fields such as dates, account numbers and amounts, and record the page and text each came from with a confidence level. For photos and scans, Microsoft's document reading service may turn the image into text, only if we have turned it on; otherwise you type the numbers in | You confirm the values before they are used. Unclear values are highlighted |
| Checking the numbers | Deterministic, versioned rules compare the bill with insurance statements, using only values you confirmed. Amounts use exact decimal math. Missing values stay missing; they are never treated as zero | You review findings |
| Findings and explanations | Fixed templates explain each finding as a question to ask, not an accusation. A language check blocks words like "fraud" or "overcharged" | A trained person reviews possible fraud, medical or legal concerns before anything is sent |
| Financial assistance screening | Fixed program rules and the federal poverty guideline math, using program records reviewed by two of our policy editors | Only the program decides whether you qualify |
| Letters | Fixed templates filled in with your confirmed information | You review and approve each letter |
| Planning next steps | Rules suggest routine requests, such as asking for an itemized bill | A policy check approves or blocks each step; consequential steps need your specific approval |
| Guided intake assistant | A step-by-step questionnaire that asks one question at a time | You answer and can go back |
| Automated phone assistant | Calls verified business lines, presses keys to move through phone menus, and talks with representatives using a fixed script, or an AI model within the rules in section 2. Twilio's speech recognition turns spoken answers into text in real time | Hands off to a person when needed |
| Support line assistant | Answers calls to (844) 553-2559, says it is automated and that a text transcript is kept, understands the reason for the call from words or keys using fixed rules (no AI model), confirms identity with a one-time code before sending account information, and connects the caller to a person during business hours | A person on our team, during business hours |
| Planning tools | Fixed math for payment-plan schedules and cost estimates | Estimates only; not offers |
| Safety checks | Rules look for possible signs of financial exploitation | Trained staff review privately |
| Read-aloud | Your device's own speech feature | You control it |
4. Evidence and limits
- Every finding links to its source: the document, version, page and the text it came from, or the rule that produced it, so you can check it yourself.
- Each output records the version of the rules used, so results can be reproduced and reviewed.
- The software can be wrong. Text may be misread, especially from poor scans. Unusual bill formats may not match our patterns. Program rules may change. That is why you confirm key numbers and why findings are written as questions.
- A difference is not proof of an error, and never proof of fraud.
- We separate kinds of results: a provider reducing a charge, an insurer correcting a payment, a confirmed reduction in what you owe, a refund, an assistance award, and a delayed payment are reported separately. Only confirmed reductions in what you owe count as savings.
- Verbal is not written. What a representative says on a call is recorded as "verbal" until confirmed in writing.
- The software does not decide your balance, your eligibility, whether care was needed, legal rights, deadlines or what an organization agreed to. It shows deadlines as they appear on your papers, with their source.
- Instructions inside documents are ignored. Text in uploaded documents and incoming emails is treated as information only, never as a command.
5. Human review and your control
- You confirm the key details read from your documents.
- You approve letters, applications and every consequential action, such as submissions, appeals, signatures, settlements and payment plans.
- A trained person reviews possible fraud concerns, medical or coding questions, legal issues, urgent appeals, unclear or close deadlines, and anything the software flags as uncertain, before anything is sent.
- You can ask for a person at any time through Help, by email at support@feedalto.com, or by phone at (844) 553-2559. Support hours are shown on the Help page. We do not offer 24-hour human support.
- You can stop outside activity on a case at any time with "Stop working on this case."
- Emergency stop. Our team can pause all outgoing calls, emails and texts at once if something goes wrong.
6. Protections that apply to all automation
- Automated steps can only propose actions. A separate, rule-based policy check, using your Case Assistance Authorization, decides whether an action is allowed, and checks again immediately before it happens.
- Nothing automated can create your consent or approval.
- The phone assistant always says it is automated, never claims to be you, does not copy anyone's voice, and does not use voice biometrics.
- It never gets around security checks or CAPTCHAs.
- Your information is not used to train AI models.
7. If automation is unavailable
If an automated function or vendor is unavailable, you can still see your documents, tasks and deadlines whenever the core service is running. Outgoing calls and messages pause until it is safe. When they resume, we check what already happened before sending anything again, so nothing is sent twice by mistake. Deadline reminders do not depend on the phone assistant.
8. Changes and testing
Changes to rules and templates are tested against review cases before release, and rule versions are recorded. We will update this notice before introducing any new kind of automation that affects you.
9. Questions or concerns
If you think an automated result is wrong, tell us in the app (Help) or contact support@feedalto.com or (844) 553-2559. A person will review it. Reports about wrong results are handled first.