Kompato AI

The removal test leaves a conventional call centre, which is exactly the cost structure the company argues against. Named generative agents conduct unscripted voice conversations with empathy and payment plan negotiation at a stated capacity of up to ten million interactions monthly, while models score accounts by repayment likelihood, predict payment behaviour and select channel, timing and strategy per debtor by behaviour, segment and geography. The parent company's consumer credit data science background underpins the scoring layer.

The split is published rather than implied, with agents resolving 96.4 percent of queries and 3.6 percent of live voice calls escalated, and the rationale for the boundary is stated sensibly: humans take the high-risk accounts requiring careful judgement, empathy or escalation. Compliance is described as human-in-the-loop.

Held at B because automation covers the overwhelming majority of contacts with a financially distressed population, and no criterion is published for what triggers escalation, nor any monitoring of whether the threshold is set correctly.

No validation method, accuracy measure or evaluation approach is published for models that predict repayment likelihood and drive contact strategy. The performance claims that do exist are inconsistent with one another and lack stated baselines or comparison periods, which is the opposite of what a model risk function needs, and a cited third-party figure putting repayment prediction accuracy at 80 percent refers to the technique generally rather than to this vendor's models.

Performance figures are plentiful but no customer is named anywhere, and two competitors independently observe the absence of public enterprise references or case studies, which matches what is findable. The figures also do not reconcile: a 62 percent liquidation uplift, liquidation rates 20 percent above peer agencies within three months, and up to twice the rate of traditional approaches are three different claims on the same measure with no stated baselines. Scaling from ten thousand to over a million accounts in 45 days is a concrete operational fact and is self-reported.

No boundary statement was located, and the corporate structure makes the question pointed. The company is a subsidiary of a consumer credit data science business, and its own platform accumulates payment behaviour, hardship disclosures and negotiation outcomes across lenders and debt buyers. Nothing states whether that behavioural data informs the parent's scoring work, is isolated per client, or is retained once a portfolio is resolved.

No data protection agreement, retention schedule or subprocessor list was located. Payment card standard certification indicates disciplined handling of payment data specifically, and the agents perform identity verification during calls, which is itself a personal data operation. What is absent is any statement about how debtor records, call recordings and behavioural profiles are retained or shared, which matters where a portfolio moves between creditors and debt buyers.

Three certifications are held rather than targeted, covering the service organisation control standard, the payment card data security standard and the international information security management standard, which together address process controls, payment handling and the security programme itself.

That combination is appropriate for a business that takes payments from consumers over the phone and holds portfolio data for regulated creditors, and it is the most complete certification set in the index alongside one other vendor.

The company holds an actual collection agency licence, which is a regulatory permission rather than an alignment claim, and the rules it operates under are named individually and embedded in the product as what it calls compliance-as-code: the federal debt collection practices statute, the telephone consumer protection statute, Regulation F including its seven-calls-in-seven-days frequency limit, and state-level requirements, with the system automatically controlling contact frequency and timing. Naming the specific rule rather than gesturing at compliance is what distinguishes this from most claims in the index.

Governance is claimed as infrastructure rather than addition, covering built-in explainability, bias monitoring and compliance auditing described as standard rather than afterthoughts, which is the right framing for a system that decides which debtors to contact, how often and with what settlement terms.

Held at B because none of it is evidenced: no bias metric, monitoring result or audit output is published, and differential treatment by segment and geography is an explicit product feature whose distributional effects are not examined.

No guarantee, indemnity or correction process was located. The consumer's position improves structurally, since frequency and timing limits enforced in code protect them from the over-contacting that training alone fails to prevent, and no route is described for a debtor who disputes the debt, believes they were misidentified, or wants to reach a person rather than an agent, beyond an escalation threshold set by the vendor and not disclosed.

No base model, provider, speech technology vendor or hosting arrangement is identified for agents conducting unscripted consumer conversations at scale. The parent company is named, which discloses the corporate dependency, and the model dependency behind the voice agents is not, so a creditor cannot document whose models are speaking to its customers.

Integration targets the three system types collections actually runs on, described as full integration with existing customer systems, dialers and payment systems, and dialer integration in particular is what allows contact frequency rules to be enforced across all channels rather than only within the platform. Held at B because no individual system is named and no developer documentation was located, so the depth of those connections rests on assertion.

No hosting provider, region selection, residency commitment or private deployment option was published. Participation in a major cloud provider's generative AI accelerator implies where workloads run without stating it, and for a platform recording voice conversations with consumers in financial distress, the location and handling of those recordings is not addressed.

The commercial model is stated plainly and is unusual: as a licensed agency it charges on an outcome basis, with no fee where nothing is collected, which places recovery risk on the vendor rather than the client. An independent review adds structure for platform deployments, describing five-figure monthly costs scaling with call volume, channel mix and analytics depth, plus a one-time setup fee covering portfolio analysis, channel integration and compliance rule configuration. Held at B because the vendor publishes no rates of its own and exact figures are negotiated case by case.

Coverage spans both sides of the charge-off line, serving first-party lenders recovering delinquent accounts before write-off and debt buyers pursuing recovery after it, which are different economics and different regulatory footing. Channels cover voice, message, email and digital with automated routing between them. Held at B because operations are confined to one country and the company has existed only since 2024, so segment breadth is not matched by institutional depth.