Cedar

Cedar Intelligence and the Kora voice agent are genuine model driven capability, and Kora is trained on the company's own billing interaction data rather than a general model, which is the substantive differentiation claim. Held back from A because the underlying business is a patient payments and billing platform that predates the AI and functions without it; Kora and the personalization engine are layers on a payments product, which the company's own framing as a performance engine for financial experience reflects.

The escalation design is stated with the detail this axis rewards: Kora escalates to a live agent with full context whenever human judgment is required, so the patient does not restart the conversation, and the vendor positions the agent explicitly as first level triage that lets call center staff work on complex interactions rather than as a replacement for them.

Sentiment and tone detection is a meaningful control in this specific context, since a billing call is often an emotionally charged interaction and a distressed caller is precisely who should reach a person quickly.

One genuine disclosure and very little else.

The disclosure worth crediting is provenance. Kora is trained on Cedar's proprietary healthcare billing data rather than adapted from a general purpose model, and the company states this plainly and repeatedly. In a category where most voice products are a general model with a prompt in front of it, saying what the model was actually built on is meaningful, and the platform's stated foundation of more than a billion patient interactions gives the claim substance.

Everything else is absent. No architecture, no accuracy figures, no containment or first contact resolution rate defined, no escalation threshold governing when the agent hands to a human, and no evaluation methodology.

The published numbers describe volume and cost rather than performance: nearly 400,000 patient calls handled across ten provider organisations in the first year, and a reported 30 percent reduction in patient billing calls. A deflection figure measures how many calls did not reach a person. It does not measure how many patients got a correct answer, and for a product explaining what someone owes and why, those are different questions.

The absent number that matters most is how often a call the agent resolved generated a second call about the same issue.

Cedar states that its voice agent is trained on the company's own healthcare billing data rather than adapted from a general purpose model, and that claim is worth a specific warning, because it is the most commonly misread sentence in this whole category. A statement about the training corpus is not a statement about the model layer.

Training on proprietary data is entirely compatible with fine tuning somebody else's base model, so the sentence forecloses nothing about who is in the chain, while reading to a buyer like a declaration that nobody is. Treat it as a provenance claim about data, and ask the separate question about the model. What is disclosed is the hosting tier, across multiple Amazon availability zones, and the record notes a named telephony partner.

What is not disclosed is the part that matters most for this particular product. A voice agent needs speech recognition and speech synthesis, those layers are usually supplied by third parties, and they are the components through which recorded conversations about a named person's medical debt and inability to pay would pass. Neither is named.

No subprocessor list is published, so a buyer cannot enumerate which parties touch that content, and the open retention question already recorded on the stewardship axis compounds it: without knowing who is in the chain, asking how long the audio is kept has nobody to ask it of. Two passes over the vendor's public material located no equivalent of the subprocessor page that peers in this index do publish.

Deployment evidence is specific, dated, and includes a named site with volume: nearly 400,000 calls handled in Kora's first year across ten provider organizations, with Gastro Health accounting for more than 60,000 calls since September 2025 across 120 plus locations. The original target, automating 30 percent of inbound billing calls, was published in advance and is therefore checkable, which is rarer than it should be. Held back from A because actual containment against that target is not reported, and resolution quality, as distinct from call volume handled, is unpublished.

The data here is protected health information joined to payment card data and financial hardship, which is an unusual combination and a sensitive one. Cedar claims HIPAA, HITRUST and PCI compliance, maintains a public security page describing continuous monitoring and disaster recovery, and hosts across multiple Amazon availability zones.

One control deserves specific credit because it is the right one for this product. Kora authenticates callers before discussing an account. For a voice agent that will explain what a named individual owes and for which services, verifying who is on the line is the primary safety control, and it is built in rather than bolted on.

The central gap follows from the product itself and is unaddressed. An artificial voice agent handling hundreds of thousands of patient calls generates a large corpus of recorded conversations in which people discuss their medical bills, their insurance problems and their inability to pay. Nothing published states whether those conversations are recorded, how long any recording or transcript is retained, who can access it, or whether that audio is used to train or improve the models. For every other data type in this index the retention question is important; for a recorded conversation about someone's financial distress it is the whole question.

No subprocessor list is published, which compounds it given the named telephony partner.

Cedar is a business associate. It handles patient billing data on behalf of hospitals, health systems and physician groups, so the status follows from the work.

What is published is thin relative to the sensitivity. A statement of HIPAA compliance, a claim of HITRUST compliance without certification language or level, and a description of Kora as designed to be compliant from the ground up including HIPAA privacy and security safeguards. Designed to be compliant describes an intention in the build rather than an attestation about the result.

No business associate agreement or its terms are published, the company does not state its role in those words, and no review cadence is given, so none of the three routes to a higher grade in this index is taken.

One scoping point a buyer should raise directly. Much of what flows through this product is financial rather than clinical, but it is not therefore outside HIPAA, because the fact that a named person received a specific billed service at a specific provider is protected health information regardless of whether a diagnosis appears alongside it. Establish in the agreement how billing communications, call recordings and transcripts are treated, since those are the artefacts this product creates that a conventional billing vendor does not.

Three frameworks are named and none is stated precisely enough to verify, which is the recurring failure of this category appearing here in its fullest form.

The company states that it is HIPAA, HITRUST and PCI compliant. Compliant is not certified. The HITRUST level is not given, and the difference between the entry tier and the risk based tier is material. The PCI service provider level is not given either, which matters for a business processing patient card payments at volume. No SOC 2 report is claimed at all, in a category where it is the baseline expectation. No auditor is named and no report is offered.

What is published is a security page describing continuous monitoring, disaster recovery, multiple Amazon availability zones and a stated approach of targeting the top security risks facing healthcare providers. Those are real practices and they are credited.

The grade reflects the gap between the practices described and the assurance offered. Naming the HITRUST level, the PCI service provider level and the SOC 2 type would cost three short phrases and would move this record substantially, because the underlying programme appears to exist.

The FDA has no jurisdiction, and the regime that does govern this product is different from the rest of the revenue cycle category. Cedar's counterparty is the patient as a payer rather than the payer as a customer, which brings consumer financial regulation into range: the Fair Debt Collection Practices Act where a vendor assists a provider in collecting amounts owed by patients, state debt collection rules, PCI DSS for the card payments it processes, and hospital price transparency obligations.

One exposure is specific, current, and unaddressed anywhere public. Kora is an artificial voice agent, and where it places outbound calls it sits inside the Telephone Consumer Protection Act. The FCC's Declaratory Ruling of 8 February 2024 confirmed that AI generated voices are artificial voices under the statute, effective immediately, meaning such calls require the prior express consent of the called party and must identify the entity responsible for initiating them. A September 2024 proposed rule would add mandatory disclosure that a call is AI generated, both at the point of consent and at the start of the call.

The distinction that matters is direction. Inbound calls, where the patient rings the billing line, raise none of this. Cedar also markets proactive outreach to recover payments, and that is the regulated case.

PCI compliance is claimed, which is the right regime for the payments side. Nothing published addresses the Telephone Consumer Protection Act or consent capture for artificial voice outreach. A provider deploying outbound Kora carries that exposure and should establish where consent comes from and who holds it.

Unusual for this index: the company publishes research examining who its own category underserves, reporting that billing experiences built for financially stable patients are failing those under greater financial strain, and that nearly 40 percent of collectible dollars on its platform now come from uninsured patients, up 54 percent in three years. Naming a population its own product historically served poorly is a genuine equity disclosure.

Held back from A because no evaluation of the AI itself for differential performance across patient groups was retrieved, and a personalization engine deciding how aggressively to pursue payment from whom is exactly where that analysis belongs.

The defining feature here is that the legal exposure created by the product does not sit with the company that built it. No device regulator governs revenue cycle software, so the regimes in range are consumer financial ones, and they attach to the caller rather than the supplier.

Where an artificial voice agent places outbound calls to patients about balances owed, the Telephone Consumer Protection Act applies, carrying statutory damages calculated per contact, a private right of action and an active plaintiffs bar. Debt collection rules attach where a vendor assists a provider in collecting patient balances.

In both cases the provider deploying the agent is the party exposed, and nothing published by the vendor addresses consent capture, revocation handling, calling window controls or scrubbing against do not call registers, which is precisely the material a provider would need to discharge the obligation it has taken on. The patient on the receiving end has a private right of action, and it runs against the provider, not against the vendor whose agent placed the call.

That is the same shape recorded on other vendors in this index and it is sharper here, because the harm is not a wrong answer but an unwanted contact about a medical debt. What does exist and deserves credit is the in product route. The agent escalates to a live representative with full context, so a patient disputing what they owe reaches a person without restarting the conversation, and tone and sentiment detection routes a distressed caller toward a human faster.

For a billing dispute that is a real and well designed correction path, and it is the reason this sits at C rather than lower. No published terms of service, indemnity, accuracy commitment or remediation obligation was located. Ask specifically who holds the consent record for outbound calling, what evidence of consent the vendor surfaces to the provider, and what the vendor commits to when an agent gives a patient an incorrect balance.

Deeper than the marketing suggests, because the product cannot work otherwise. A voice agent that explains a specific charge, states the current balance, describes insurance coverage applied and offers payment options has to reach live billing data at the moment of the call. Cedar describes real time data integrations supporting exactly that, and reports deployment across provider organisations running both Epic and Cerner, alongside physician staffing groups and large specialty groups.

That is a meaningful integration claim. Answering a live financial question correctly requires current adjudication status, not a nightly extract.

Held at B because none of it is enumerated. No named integration list, no partner directory, no standards or interfaces described, and no statement of what happens when the underlying data is stale or a claim is mid adjudication. The breadth is also far narrower than the platform vendors in this category, which is appropriate to a focused product rather than a deficiency, but it should not be read as equivalent.

A buyer should establish which systems are supported natively, what the refresh latency is, and how the agent behaves when it cannot resolve a balance with confidence.

Partially answered. Cedar names Amazon Web Services and states it runs across multiple availability zones, with disaster recovery and continuous monitoring described on a public security page. Naming the cloud provider puts it ahead of most peers in this category.

What is missing is everything downstream of that. No region or residency statement, no processing location for the data, and no subprocessor list.

The subprocessor gap is not academic here, because the company names a build partner itself. Kora was developed in collaboration with Twilio, which means patient telephone conversations about medical bills traverse a third party communications platform. That is an ordinary and sensible architecture, but it puts a named third party into the path of both protected health information and payment discussion, and nothing published describes the boundary, what is retained on either side, or what the contractual position is.

A buyer should ask for the full subprocessor list rather than the hosting summary, and should ask specifically what the telephony partner receives and retains.

No published pricing and no published pricing mechanism.

One specific description exists and its provenance should travel with it. A research publication reports Cedar selling as enterprise software on a monthly fee scaled to seats and integrations, with custom contracts on the payments product charging between 1 and 4 percent of collections. That is attributed to a single unnamed former employee, not to the company, and it is not corroborated. It is more concrete than most third party speculation in this category and it is still not a disclosure.

If a percentage of collections applies, the incentive question is sharper here than anywhere else in this lane, and it is worth setting out because the company's own published material sits on both sides of it.

Cedar writes publicly about propensity to pay modelling, the practice of identifying which people are most likely to pay and allocating effort accordingly, and notes that healthcare is adopting an approach long used in consumer finance. Where a vendor is paid a share of what it collects, a model that sorts patients by expected recovery directly determines whose bill gets pursued, through which channel, and how persistently. The person being sorted is not the customer and cannot see the assessment.

The counterweight is real and belongs beside it. Cedar publishes research finding that 40 percent of collectible dollars on its platform now come from uninsured patients, up 54 percent in three years, and argues that billing experiences designed for financially stable people are failing those under the greatest strain. That is a vendor publishing something inconvenient about its own market. Ask which of those two instincts the pricing structure actually rewards.

Tightly bounded to patient financial engagement across health systems, physician groups, and clinician staffing organizations, with stated coverage of Epic and Cerner environments and athenaOne practices. No clinical claims are made, which is correct for a billing product.