Collectly

Billie is a genuine conversational agent handling billing questions and taking payment across chat, text, email, and voice. Held back from A because a substantial part of the platform is payments infrastructure, statement delivery, and plan management that functions without AI. The company's own framing is accurate: the EHR is the system of record, and this is the system of action layered on top.

Stated design is that the agent handles the majority of cases and staff step in for exceptions, which is the correct shape. Held back from A because no escalation criteria or confidence threshold is published, and the consequence surface deserves it: a patient asking why they owe an amount is receiving an explanation with financial consequence, and an incorrect answer can produce either an unwarranted payment or an avoidable escalation to collections.

The agent is described by its channels and its coverage, never by its performance.

What is published: an artificial intelligence agent available around the clock across chat, email, text and voice, answering billing questions, explaining what is owed and taking payment, with the company stating it handles the majority of cases while staff step in for exceptions. A separate agent handles eligibility and benefits verification.

The majority claim is the one that needs a number and does not have one. Containment rate, meaning the share of interactions resolved without a person, is the central operating metric for a conversational agent and it determines both the economics and the patient experience. A majority could be fifty one percent or ninety five percent and those describe entirely different products.

Also absent: any model or architecture description, accuracy on explaining a balance correctly, escalation criteria defining what constitutes an exception, error or complaint rates, and any versioning policy.

The accuracy question is not cosmetic in this category. A cited industry finding notes that forty five percent of insured adults received a bill they believed should have been covered. An agent explaining a disputed balance is operating in exactly that territory, where being confidently wrong has a direct financial consequence for the patient.

Ask for the containment rate, the escalation criteria, and how often patients dispute after an agent interaction.

The holding is layered and only one layer has published terms. Four categories meet here that are usually separate: patient identity and contact details; billing information, which necessarily carries clinical implication because what a person owes reveals what they were treated for; payment card data; and recordings or transcripts of patients discussing medical bills, which is a category of health adjacent content that barely existed before conversational agents and which no standard was written for.

The card layer is genuinely covered, by the payment card industry standard, which is the correct and prescriptive instrument for that data with defined requirements for storage, transmission and retention. Nothing equivalent is published for the other three.

No retention period for billing records, call recordings or transcripts, no statement of whether patient interactions train the agent across a reported presence in more than three thousand facilities, no de identification step, and no description of what is returned or deleted when a provider leaves.

One detail is worth crediting as a small signal of care and this index has credited the same choice elsewhere: the company's own product demonstration uses a simulated patient scenario rather than a real interaction, explicitly to avoid exposing real information. Ask whether calls are recorded, for how long, whether interactions train the agent, and what governs the transcript layer.

Scale is substantial: more than 3,000 healthcare facilities and over $1 billion in patient payments processed, which is a hard number rather than an engagement metric. Held back from A because performance claims, collections increases of 75 to 300 percent and days to collect as low as 12.6, are vendor and customer reported without baselines, and a range that wide suggests results depend heavily on what the organization was doing before.

A layered holding, one layer of it well protected, and no published terms for the rest.

What the platform handles spans three categories that are usually separate. Patient identity and contact details. Billing information that necessarily carries clinical implication, because what a person owes reveals what they were treated for. And payment card data. Voice interaction adds a fourth: recordings or transcripts of patients discussing medical bills, which is a category of health adjacent content that barely existed before conversational agents.

The card layer is genuinely covered. PCI DSS compliance is held and it is the correct and prescriptive standard for that data, with defined requirements for storage, transmission and retention.

Nothing equivalent is published for the rest. No retention period for billing records, call recordings or transcripts. No statement of whether patient interactions train or improve the agent, which matters for a product reported to handle the majority of cases across more than three thousand facilities. No de identification step, and no description of what is returned or deleted when a provider leaves.

One detail is worth crediting as a small signal of care: the company's own product demonstration uses a simulated patient scenario rather than a real interaction, explicitly to avoid exposing real information.

Ask whether calls are recorded, for how long, and whether interactions train the agent.

A genuine compliance stack, correctly named, with the instrument itself unpublished.

What is held is documented and healthcare specific: HIPAA compliance, SOC 2 Type 2 with the type stated, HITRUST i1 certification, and PCI DSS. That combination is well matched to what the platform actually does, since HITRUST addresses the health information side and the payment card standard addresses the transaction side. Naming the SOC 2 type and the HITRUST level rather than saying certified and leaving it there puts this record ahead of most of the revenue cycle lane, where this index has repeatedly found both details omitted.

One calibration point for a buyer, consistent with how this index reads the framework elsewhere: i1 is the fixed control set certified for one year, while r2 is the tailored risk based two year certification. Both are real. A buyer told only that a vendor is HITRUST certified would reasonably assume the higher tier, and naming the level avoids that.

The role is straightforward. The provider organisation is the covered entity and this platform processes patient information on its behalf as a business associate.

Held at B because no business associate agreement, addendum, subcontractor flow down, breach notification timetable or review cadence is published. The route to an A here is publishing the instrument, and it is not taken.

Ask for the agreement, and for the current certification dates.

The most complete certification stack among the patient financial vendors in this index: HIPAA, SOC 2 Type 2, PCI DSS, and HITRUST i1. PCI DSS matters specifically here and is often missing from comparable claims, since this product takes card payments directly, and card data handling is a separate risk surface from protected health information.

No device pathway applies and none is claimed. What governs is consumer financial regulation, and the voice agent puts this record squarely inside a regime the company does not address.

The scoping call for patient billing is settled in this index: the Fair Debt Collection Practices Act where it applies, the Telephone Consumer Protection Act, state debt collection rules, payment card standards and price transparency. The coding and claims exposure that governs the payer facing side of the revenue cycle does not transfer here.

One credential is directly on point and is credited. PCI DSS compliance is the right standard for a platform taking card payments, and holding it is substantive rather than decorative.

The gap is the voice agent. The February 2024 declaratory ruling confirmed that artificial and prerecorded voice restrictions cover voices generated by artificial intelligence, with no exemption for systems that sound human. Direction is the distinction this index draws and it matters here: the product is presented primarily as answering patients who call in, and inbound contact raises none of this. Outreach is a different matter.

Two further layers are worth separating rather than merging. Calls about a debt owed to the caller are informational rather than marketing, which changes the consent analysis. Separately, where an account has gone to collection, the debt collection statute imposes its own conduct rules on disclosure, calling hours and frequency that apply regardless of consent.

Nothing published states a position on any of it. Ask who bears consent and do not call obligations under the contract.

No governance framework, bias evaluation, monitoring commitment or third party review was retrieved, and two specific risks apply.

The first is segmentation. The platform offers personalised outreach, payment plans and intelligent statements, which means something decides which patient receives which treatment, in what sequence and on what terms. This index has an established thread on exactly this in adjacent records, where models sort patients by expected recovery and that sorting determines whose bill gets pursued, through which channel and how hard. Nothing published states whether outreach intensity or payment plan terms vary by any predicted characteristic, and nothing states they do not. The question is open and it is the one a buyer should ask, because the person being sorted is the patient, who did not choose the vendor and cannot see the assessment.

Write it in both directions, as this index does elsewhere. Identifying someone who needs a payment plan and offering it earlier is a genuine good, and the same prediction pointed the other way allocates collection effort.

The second is the voice agent itself. Speech recognition performance varies across accent, dialect and language, and a patient who cannot be understood by the agent that explains their bill is worse off than one who reaches a person. No language coverage or performance breakdown is published.

Ask what drives outreach segmentation, and which languages the agent supports.

Two passes located no model or architecture description, no accuracy figure for explaining a balance correctly, no escalation criteria, no error or complaint rate and no warranty, indemnity or remediation commitment. The claim that needs a number and does not have one is the coverage claim: the company states the agent handles the majority of cases while staff step in for exceptions, and a majority could be fifty one per cent or ninety five per cent, which describe entirely different products with entirely different staffing consequences and entirely different patient experiences.

Containment rate is the central operating metric for a conversational agent and it is absent. The domain makes the missing accuracy figure consequential rather than merely unsatisfying. A cited industry finding notes that a large share of insured adults have received a bill they believed should have been covered, so an agent explaining a disputed balance is operating in exactly that territory, where being confidently wrong has a direct financial consequence for the patient and where the patient has less information than the system does.

A wrong explanation that sounds authoritative may end a dispute the patient would have won. Nothing describes what the agent does when it cannot reconcile a balance, or what escalation criteria define an exception. Ask for the containment rate, accuracy on balance explanation, the escalation criteria, and the rate at which patients dispute after an agent interaction.

Integration breadth is the strongest element: more than 20 leading EHRs via API or FHIR, spanning ambulatory and enterprise systems, plus a bidirectional Epic integration listed in Connection Hub on the Epic Showroom as of July 2026, meaning live Epic customers and pre built integration rather than a claimed capability. EHR agnostic deployment matters here because patient billing sits across every system a group runs.

Integration is documented in detail and hosting is not documented at all.

The connection side is strong and is graded separately: more than twenty record systems via application programming interface or the standard clinical data exchange protocol, and a bidirectional integration with the dominant record vendor listed in its own connection directory. That tells a buyer how data moves in and out.

Where it rests is unpublished. No cloud provider, no region, no tenancy model, no statement of whether customer environments are separated, no subprocessor list, no backup or recovery posture and no retention schedule.

The subprocessor question is more pointed here than for most records because of the voice channel. Conversational voice agents are typically assembled from components, telephony, speech recognition, a language model and speech synthesis, and those are frequently third party services. If any of them sits outside the vendor's own environment, patient billing conversations are traversing suppliers the provider has never assessed. Nothing published names a single one.

Payment processing raises the same question in a regulated form, though the card standard the company holds imposes its own requirements on that path.

Ask where the platform is hosted, which suppliers sit in the voice path, and whether tenancy is isolated.

No rate card, but the model is disclosed as monthly subscription rather than a percentage of collections. That distinction matters more than the number: contingency pricing on patient collections aligns a vendor with collecting more from patients, while subscription pricing does not, and this is the same structural question raised for Lyric on the payer side.

Precisely bounded, and the company states the boundary better than most.

The scope is the patient responsibility portion of provider revenue: statements, outreach, payment plans, card on file, autopay, refunds, and eligibility and benefits checking. It is deliberately not clinical and not the payer side of the revenue cycle, and it does not claim to be.

The positioning statement is unusually clear about where the product sits relative to what a provider already runs. The electronic health record is framed as the system of record and this platform as the system of action on top of it. That is a scoping claim rather than a marketing line, because it tells a buyer what the product does not intend to replace, and this index consistently rewards a vendor that describes its own limits.

Specialty coverage is broad by construction rather than by claim. Patient billing behaves similarly across specialties, so a reported footprint of more than three thousand facilities reflects a horizontal function rather than an assertion of clinical breadth.

The market framing is also quantified rather than gestured at: patient responsibility rising from roughly five percent of provider revenue in 2010 to about twenty percent, projected toward thirty percent. Naming the trend that justifies the category, with figures, is better practice than asserting a large addressable market.