Prodigal

Nothing survives removal of the models. The company describes running fine tuned large language models in production against proprietary industry specific data, and every application in the suite is an expression of the same engine: the voice agent, the copilot that prompts a human collector mid call, the automated note taking, the conversation analytics, the quality scoring and the training workflows.

The intelligence engine underneath them is itself described as the product, trained on a stated half a billion consumer finance interactions and existing to carry context between applications. There is no dialler, no case management system and no services practice beneath the modelling that a customer could still buy.

A named mechanism whose internals are not described, which is the middle rung of this index's ladder. The stated approach is to programmatically encode each customer's own regulatory requirements into guardrails that govern all operations, with the software said to operate strictly within those parameters. Encoding per customer rather than per statute is a real design decision and a defensible one, since a credit union and a healthcare receivables operator answer to different rules.

What is absent is the specification: no rule is named as enforced, no threshold, contact cap or calling window appears, no confidence level or escalation trigger is stated, and no human approval point is described anywhere in a product line whose stated ambition is complete automation of collections operations at scale. A peer in this same pocket publishes its controls with the statute each one implements, which is what separates that grade from this one.

More is said about the models than most vendors offer and none of it is validation material. The company states that its engine is trained on half a billion consumer finance interactions and that it runs fine tuned large language models in production against proprietary data, which describes the shape of the system.

What an institution would need to defend that system internally is missing entirely: no accuracy, containment or escalation rate, no false positive figure for the compliance monitoring that customers rely on for quality assurance, no description of how the models are evaluated or how often they are retrained, no artificial intelligence management system certification, and no validation documentation. The published 98 percent reduction in compliance mistakes is an outcome measured on the customer's error rate, not evidence about how the models behave.

The strongest evidence available in this pocket, because the customer said it rather than the vendor. Consumer Portfolio Services, a lender listed on Nasdaq, announced its selection of the platform for collections and servicing through its own investor relations channel, which puts the reference inside a disclosure route the customer is accountable for.

Quantified outcomes are published and specific: up to 27 percent more digital engagement, 16 percent more payments, 30 percent greater agent effectiveness, productivity gains stated as high as 30 percent, and one customer reducing a quality assurance team by a third. Scale figures are consistent across independent trade coverage, an investor's published investment thesis and the company's own material. The residual weakness is that most outcome figures still attach to unidentified customers, so the named reference and the measured results are not the same engagement.

This vendor presents the pooled corpus question in its sharpest form anywhere in the index, and never addresses it. The engine serving every customer is stated to be trained on half a billion consumer finance interactions accumulated across a customer base of more than 100 financial companies, and the company states that nearly one in five United States borrowers has engaged with its systems.

So a single private model holds conversational history on roughly a fifth of American borrowers, assembled from creditors who compete with one another. Nothing states whether one lender's calls train models that serve a rival lender, whether a borrower's history at one creditor informs how an agent treats them at another, whether a customer can exclude its data from the shared corpus, or what happens to that contribution when a customer leaves. No safety architecture, red teaming practice or output boundary is described either.

No data processing agreement, subprocessor list, retention schedule or consumer facing privacy statement was located, and the sensitivity here is at the top of the range for this index. The platform holds recorded conversations in which identified borrowers discuss hardship, income and inability to pay, alongside account balances and payment histories, obtained from a creditor rather than from the consumer.

The trust centre is the natural place for this material and named only governance policies in what was retrieved. Nothing states how long recordings and transcripts are retained, whether a borrower can reach or delete their own record, or what happens to that data when a creditor stops using the platform.

A dedicated trust centre is operated at its own subdomain, publishing named governance documents including an acceptable usage policy and an access control policy, and framing the purpose as letting customers see the security practices and responsibilities before choosing the provider. Running a standing trust surface rather than answering security questionnaires ad hoc is a real structural commitment and most of this index does not do it.

Stated plainly and recorded as a limit on this grade rather than glossed: no certification was confirmed from the material retrieved. Whether a service organisation control report or an information security management system certificate sits behind the trust centre's request gate was not established, so the credential half of this axis is unverified and worth a targeted check on a later pass.

The company is a technology supplier to regulated creditors and collectors and holds no licence of its own, which is the correct posture and attracts no penalty. Its compliance work is real and is credited on the autonomy axis instead, under the standing rule that encoding a customer's regulatory requirements is a product capability rather than supervisory standing.

No regulator engagement, sandbox participation or supervised assessment was located, and no position is stated on the pending federal rulemaking on artificial intelligence generated calls, which bears directly on whether its autonomous voice agent may keep operating as designed.

No fairness testing, differential outcome monitoring or bias assessment is published, and the product design makes the exposure specific rather than generic. The agents are described as personalising strategy per consumer, generating propensity to pay scores that drive segmentation, and deciding when to show understanding during hardship and when to press directly for payment.

Those are differential treatment decisions about people in financial distress, made at scale on a population the company states includes nearly one in five United States borrowers. Nothing addresses whether propensity scoring or conversational tone varies systematically across demographic, geographic or linguistic groups, and no per cohort figures are published for any of it.

No liability position, error rate, remediation commitment or correction path is published. The headline compliance claim points directly at the gap: eliminating a stated 98 percent of compliance mistakes is a claim about the remaining errors as much as the removed ones, and nothing describes what happens with them.

A borrower given incorrect balance or settlement information by an autonomous agent, misidentified during right party verification, or scored into a collection strategy on faulty data has no described route to have that reviewed or corrected, and nothing states whether the creditor or the vendor answers for an automated violation.

Partial disclosure of an unusual kind and none of the part that matters. The company states plainly that it fine tunes large language models on its own proprietary industry data and operates its own intelligence engine, which tells a buyer that the modelling is done in house rather than resold, and that is more than most vendors here disclose.

What is never stated is which base models are fine tuned, whose they are, what version, where inference runs, or whether recorded borrower conversations pass through infrastructure the company does not control on their way to being processed.

The integration architecture is a stated design principle rather than a list. The intelligence engine is presented as connecting to a customer's systems and standardising the data across them, and during a live call the agent is described as searching the customer relationship system, the loan management system, policy documentation and product specific guidelines simultaneously rather than working from a script.

The payment product is stated to work alongside an existing payment portal or interactive voice response system rather than replacing it. Those are the right categories of system to name for this workload. Held off the top grade because not one of them is named as a specific supported product, no integration count is published, and no loan servicing or collections platform is identified by name.

No hosting regions, residency commitments, tenancy model or deployment options are published. The product is described as cloud based and nothing further is stated about where borrower conversations are processed or stored. The company operates a global team with engineering presence outside North America while its customers and their consumers are North American, which makes the question of where recorded consumer financial conversations are handled a concrete one for a creditor's vendor risk assessment rather than a theoretical gap.

No pricing, tier structure or billing basis is published for any of the eight named products, and the route to a figure is a sales conversation. The suite structure makes the omission more consequential than usual, because a buyer cannot tell whether the voice agent, the copilot, the analytics product and the payment portal are separately licensed or bundled, nor whether pricing runs per seat, per account, per minute or on recovered value. Competitors in this pocket are reported to publish per minute models, so the silence is a choice within the category rather than a category norm.

Both dimensions of this axis are satisfied, which is uncommon. The buyer types are genuinely distinct rather than variations on one shape: collection agencies, direct lenders, credit unions, auto finance companies and healthcare revenue cycle providers, with banking named alongside them.

Scale is stated consistently across the company's own material, an investor's published thesis and a customer's regulatory announcement at more than 100 financial companies across North America, upwards of half a billion interactions and more than 15 million loan accounts analysed. A named customer anchors the claim: Consumer Portfolio Services, a listed auto lender. The bound worth recording is geographic, since the coverage and the compliance design are North American and nothing addresses operating under another jurisdiction's collection regime.