Domu AI

The conversation is the deliverable, so the removal test leaves nothing standing. Generative agents hold the servicing call, text and email themselves, and a behavioural model decides which account to approach, at what hour and through which channel before the agent opens its mouth. Strip the models and there is no residual workflow product, no dialler and no case management system underneath, only an integration into a lender's core system with nothing to send through it.

This is the assistant native shape rather than a communications platform with models bolted to it, which is the distinction that separates the top tier of this category from channel platforms that added intelligence to an existing spine.

The oversight architecture is split across three named modules sitting at three distinct points in the execution path, and each one names the law it is checking against. Before deployment, Alex stress tests conversation flows against Fair Debt Collection Practices Act and Telephone Consumer Protection Act boundaries in a synthetic environment and confines the agent to an approved repository of data.

During the conversation, required disclosures are delivered automatically, consent state is tracked in real time, and inappropriate language, threats and policy deviations are flagged as they occur rather than found afterwards. After deployment, Jordan audits live conversations against unfair, deceptive or abusive acts and practices standards and state specific collection law, flags policy drift and assembles evidence an examiner can read.

A human supervisor receives any conversation that moves beyond preset parameters and sign off is required on high stakes decisions. Naming the statute at each stage and placing a staffed human layer in the path is what separates this from a general assurance that guardrails exist. It sits below the strongest example in this pocket only because the preset parameters themselves are never given values: no calling window, no contact frequency cap and no confidence threshold is published.

There is a real validation mechanism and it runs at both ends of deployment, which is more than most of this pocket offers. Conversation flows are adversarially tested in a synthetic environment before the agent reaches a customer, and live behaviour is audited continuously afterwards, with the stated output being audit ready evidence an institution can put in front of an examiner. That structure hands the buyer something to inspect rather than asking it to accept a published claim.

Held off the top grade because none of it is quantified: no accuracy, containment or escalation rate is published, no false positive rate is given for the compliance screening the institution would be relying on, no retraining cadence or drift measurement appears, no artificial intelligence management system certification is held, and no model documentation or validation pack is offered.

Named institutions appear next to specific numbers, which is the bar and is rare in this pocket. SBS Insurance is credited with a three percent liquidation improvement inside the first two months of use, Skandia with one million dollars recovered, and BNP Paribas with a forty percent increase in right party contact rate, alongside a forty six percent reduction in cost per account resolved and a thirty percent fall in complaints per hundred calls at an unnamed top five fintech.

Nu, DigniFi and Alorica are named as customers separately. A 25 million dollar Series A led by Standard Capital puts an independent party with money at stake behind the diligence. The residual weakness is that every figure is company reported, with no customer published announcement and no third party verification, so the reference base is named but the measurement is not independent.

A safety architecture is named and one component of it is a genuine boundary rather than a description of good intent: the agent is confined to an approved repository of data, so what it can draw on in a live conversation is bounded in advance, and it is adversarially tested against threats and inappropriate language in a synthetic environment before reaching anyone. Every supported conversation is then monitored and evaluated.

Held off the top grade on the boundary question this pocket keeps raising and nobody answers. The behavioural models learn from institutions' historical servicing data across a customer base the company describes as including eight of the twenty largest banks and insurers in the Americas, which means competing creditors, and nothing states whether one creditor's conversations inform the models serving another, whether a consumer's behavioural profile follows them when a different creditor places their account, or what happens to a client's contributed data when the relationship ends.

No privacy policy detail, processing agreement, subprocessor list or retention schedule was located. What is at stake is unusually sensitive even by the standards of this index: recorded conversations in which consumers discuss why they cannot pay, disclose hardship circumstances and hand over payment credentials, gathered from people who are behind on an obligation and who did not choose this vendor.

The company also reaches across the Americas, with named customers in Brazil and Colombia as well as the United States, so the same recordings cross data protection regimes that each impose their own obligations, and nothing addresses which regime governs a given conversation or how long the audio and transcripts are kept.

Two audited credentials are claimed, SOC 2 Type II and PCI, and both are appropriate to a platform that negotiates and takes payment on live consumer calls. The presentation is the problem and it is the second instance of a pattern already recorded in this pocket: the same row lists the Consumer Financial Protection Bureau and the Telephone Consumer Protection Act as though they were certifications alongside them.

A federal supervisor and a federal statute are not schemes anyone can certify against, so a buyer reading the row cannot tell which items were audited by a third party and which are assertions of compliance with law. Held at this grade because two audited certifications remain two audited certifications, and off the top grade for no scope statement, no audit period, no named auditor, no assessed payment card level and no trust centre or evidence portal located.

The company presents itself as a platform sold to regulated institutions and holds no licence in its own name in anything located, which is the correct and unpenalised posture for a software vendor. One claim is flagged rather than credited: a third party application directory describes the company as a licensed collections agency as well as a platform, and if that were established in the company's own material it would place it inside the regulatory perimeter and change this grade substantially, on the precedent already set for licensed digital first agencies in this same category.

The claim appears in no company published source reviewed, names no state or jurisdiction and no licence number, so on the standing rule that a credential must appear in the vendor's own material and that an ambiguous claim earns nothing even when the cost falls this way, it earns nothing. Worth one targeted check.

The published governance work is conduct screening rather than fairness testing, and the difference matters here. Checking a conversation against unfair, deceptive or abusive acts and practices standards catches what an agent said to one person; it does not test whether the system treats comparable people differently.

The exposure sits one layer earlier, in the behavioural model that decides who is contacted, how often, through which channel and with what tone, which is a differential treatment decision made about people in financial distress and made before any conversation a compliance module could review.

No fairness evaluation, differential outcome analysis or cohort level reporting appears anywhere, and nothing states whether contact intensity or channel assignment distributes evenly across geography, product or borrower type.

The architecture is preventive from end to end and there is no route afterwards, which is the structural position every software vendor in this pocket occupies. The creditor remains the regulated party and absorbs the statutory consequence of a call it did not script, made by a model it did not build and cannot validate, while the platform operates the agent. Evidence assembled for examiners is a defence for the institution, not a remedy for the consumer.

No liability position, error rate or remediation commitment is published, and a consumer wrongly contacted, misidentified as the right party or held to a payment commitment they did not understand has no described way to have the interaction reviewed by anyone.

No model provider, family, version or hosting arrangement is named. The material describes generative agents and language models trained to sound like high performing collectors and stops there. The voice specific version of this gap is the sharper one: speech synthesis and real time transcription are components a company of this size almost always licenses rather than builds, so an undisclosed supply chain on a voice product means a consumer's spoken account of their financial difficulty passes through at least one third party the buying institution was never told about, and the institution cannot enumerate its own fourth party exposure when its examiner asks.

The company is unusually candid that this is not a plug and play voice bot and that it will not work without connection to the institution's core systems. Buyers are told to expect application interface setup, data mapping and agent training before the platform can handle live conversations, and the behavioural layer is described as depending on the lender's own historical servicing data, which makes the integration load a stated prerequisite rather than an optional extra.

Held off the top grade for the reason that recurs across this pocket: no core banking, loan management or servicing platform is identified by name, no integration count or partner directory is published, and no public application interface documentation was located, so the depth is asserted through the size of the implementation rather than shown.

The platform is cloud delivered and integration led, and nothing beyond that is published about where it runs. No hosting region, residency commitment, tenancy model or single tenant option appears. Residency is a concrete question rather than a formality here, because a San Francisco company is handling recorded consumer collection conversations for institutions in Brazil, Colombia and the United States, and each of those markets treats the movement of that data differently. Nothing states where audio, transcripts or the behavioural profiles derived from them are stored or processed, or whether a Latin American client's borrower conversations remain in region.

No price, tier, billing basis or starting figure is published for any part of the platform. The gap is conspicuous because the company publishes extensively on how institutions should evaluate the total cost of ownership of platforms in this category, including the costs of integration work and compliance remediation, while disclosing none of its own.

A buyer cannot tell whether the model is subscription, per account, per connected minute or a share of amounts recovered, and in collections those imply materially different incentives, since a vendor paid on recovery has an interest in contact intensity that a vendor paid a licence fee does not. The stated requirement for substantial integration work also implies an implementation cost that is never sized.

Six distinct buyer types are addressed and the named references land in several of them: banks, fintech lenders, buy now pay later providers, insurers, loan servicers and business process outsourcers, with Nu, DigniFi and Alorica named as customers and SBS Insurance, Skandia and BNP Paribas named alongside published outcomes.

That set spans a digital bank, a specialist finance lender, an outsourcer, an insurer and a European banking group, which is genuine breadth rather than one buyer described several ways. The company states it works with eight of the twenty largest banks and insurance companies in the Americas. Coverage is regionally bounded to the Americas and the material offers nothing for Europe or Asia, which is the honest limit of the claim rather than a gap in the evidence.