Skit.ai

The removal test leaves nothing. What a collections operation buys is an agent that holds a conversation with a consumer, understands what was said and produces the next line, and every part of that is model work. There is no dialler business, no data product and no services practice underneath it that would survive if the models were taken out; what would remain is the human calling team the product exists to replace.

The compliance layer wrapped around the conversation is a control on the models rather than a separate product, and the vendor positions the whole system as a voice layer sitting over collections software it does not itself sell.

The most completely specified oversight architecture encountered in this index, and the reason is that each control names the rule it implements rather than describing a general intention. Before a number is dialled, eligibility, consent and timing are resolved, accounts are filtered against do not call registries, bankruptcy filings, statute of limitations and known litigators, the calling window is enforced at 8am to 9pm in the consumer's own local time under Regulation F, and a seven contacts in seven days cap is applied per account under the same rule.

During the call, every drafted line is screened in real time before the consumer hears it, with disclosure logic and escalation rules encoded into the agent. After the call, the record is scored, reconciled and preserved. Deployment itself carries a gate: a staged rollout requiring client sign off on scripts, constraints and consent flows. Controls that cite the statute they enforce are auditable in a way that guardrail language never is.

The validation mechanism is real and unusual, and the model documentation behind it is absent. Every engagement is stated to open with a live production grade pilot designed to prove performance, compliance and return under real conditions, which hands the institution the ability to measure the system on its own accounts before committing rather than accepting published figures.

Compliance checking is described as running before, during and after each call, with conversations reviewed against the federal regimes and state rules. What is missing is everything about the models themselves: no accuracy, containment or escalation rate, no false positive figure for the compliance screening, no description of how the agent is trained or evaluated, no artificial intelligence management system certification and no validation documentation supplied to the institution that would have to defend the system to its own regulator.

Quantified results and not one named customer, which under this index's bar is the lower grade however specific the figures are. Published case study outcomes include 43 percent higher collections, 75 percent lower call costs and a doubling of promise to pay over a year across a four phase deployment, with results stated to appear within 30 to 60 days.

Every one of those sits against an unidentified client, the supporting testimonial is unattributed, and no lender, bank or agency is named anywhere in the material reviewed. The omission is conspicuous rather than ordinary at this scale: a vendor claiming more than a billion consumer interactions and publishing percentage improvements has a reference base and does not draw on it publicly. The offsetting fact, credited on the model risk axis rather than here, is that every engagement is said to begin with a live pilot in which a prospect measures these claims on its own accounts.

The safety half is genuinely well specified and the stewardship half is untouched. Red teaming for harmful or non compliant output is named, guardrails and escalation rules are encoded into the agent, real time screening sits between the model and the consumer, and safety requirements are stated to be set per use case and per jurisdiction rather than once globally. Against that, no boundary is described anywhere.

Nothing states whether conversations from one creditor's accounts inform models serving another, whether recordings are used for model improvement, or whether a consumer's interaction history follows them across creditors on the platform. At a claimed billion interactions across many lenders, an aggregated behavioural picture of individual consumers in distress is an asset the material never acknowledges existing.

No privacy policy detail, data processing agreement, subprocessor list or retention schedule was located, and the data in question is about as sensitive as consumer financial data gets. The platform holds debt balances, payment histories, bankruptcy status, litigation history and recorded voice conversations about a person's inability to pay, obtained largely without that person having chosen the vendor.

Interactions are stated to be recorded and securely stored with audit trails, which is a retention practice described without a retention period, and PCI DSS certification is credited on the security axis. Nothing states how long recordings are held, whether a consumer can request deletion, or how data from one creditor's accounts is separated from another's.

Three certifications, and the mix is well matched to what the product actually does. SOC 2 Type II is held and described as covering the platform end to end, ISO 27001 addresses the information security management system, and PCI DSS matters more here than it would for most vendors in this index because the agents negotiate and take payment during live consumer calls. That is a deeper stack than most of this pocket carries.

Held off the top grade because the certificates are asserted without the surrounding detail a buyer would use: no scope statement, no audit period, no named auditor, no control implementation figures and no trust centre from which a report or certificate can be requested or read.

The firm is a technology supplier to regulated collectors and holds no licence itself, which is the correct posture and carries no penalty. It sits closer to the regulatory perimeter than most vendors in this index, because the rules it encodes are the operating law of its customers' industry and it states that it updates those rules as regulations change, but under the standing index ruling that is a compliance capability credited on the autonomy and security axes rather than supervisory standing of its own.

No regulator engagement, sandbox participation or supervised test of the agents was located, and no position is stated on the pending federal rulemaking covering artificial intelligence generated calls, which is the single regulatory development most likely to alter how this product may operate.

Bias and fairness evaluation and red teaming for harmful or non compliant output are named as steps in the published responsible AI process, which puts this vendor ahead of most of the index on an axis where silence is the norm. Naming the practice is where it stops. No methodology is described, no results are published, no differential outcome analysis across demographic or geographic groups appears, and no independent audit is referenced.

The exposure is concrete rather than theoretical: an agent that negotiates payment terms and decides how to escalate is making judgements about people already in financial distress, and voice systems are known to perform unevenly across accents and dialects, so a per cohort containment or escalation figure would say something a statement of intent cannot.

No liability position, error rate, remediation commitment or correction path is published, and the regulatory setting makes that absence unusually pointed. Every control the vendor describes exists because contacting the wrong party, calling outside permitted hours or exceeding the contact cap carries statutory consequences under the federal collection and telephone consumer protection regimes.

The architecture is preventive throughout and there is no stated route afterwards: a consumer wrongly called, misidentified or given incorrect information by an agent has no described way to have the interaction reviewed or corrected, and nothing states whether the creditor or the vendor answers for a violation the automation produced.

No model provider, family, version, hosting arrangement or country of processing is disclosed. The material describes what the agents do and what constrains them without ever stating what produces the speech and the reasoning, whether any component is licensed from an external provider, or whether consumer conversations are processed by infrastructure the vendor does not control.

For a product whose recordings capture consumers discussing debt and disclosing payment details, whose infrastructure carries that audio is a question a creditor's own vendor risk process would be obliged to ask.

The architectural position is stated plainly and the named counterparties are data sources rather than platforms. The product is positioned as a voice layer integrating into an existing collections platform rather than replacing it, which is a real integration commitment and the harder of the two paths.

Pre dial screening implies and describes live connections to several external references: do not call registries, bankruptcy records, statute of limitations data and litigation histories, plus skip tracing, phone validation and email refresh services.

Held off the top grade because no collections platform, dialler or case management system is named as supported and no integration count is published, where at least one competitor in this pocket names the specific recovery platforms it connects to.

No deployment options, hosting regions, residency commitments or tenancy model are published. The product is delivered as a hosted service and nothing states where calls are processed, where recordings are stored, or whether a customer can require that consumer audio remain within a particular jurisdiction.

The gap is narrower in practical effect than it would be elsewhere, since the customer base and the statutory design are both United States centred, and it is still unaddressed for a buyer with state level data handling obligations or an eye on expansion.

No pricing, tier structure or billing basis is published, and the route to a number is a conversation. One real disclosure sits alongside that: the engagement model itself is stated, with every deployment opening as a live pilot, which tells a buyer something about commitment shape even though it says nothing about cost.

Competitors in this pocket are reported to publish per minute rate models, so the reticence is a choice within this category rather than a category norm, and a collections operation weighing automation against agent headcount cannot make that comparison from published material.

Real breadth of buyer type and real volume, bounded by geography and by an absence of names. The stated customer base spans large consumer lenders, banks and collection agencies including business process outsourcers, which is three distinct buyer shapes rather than one, and the claimed scale of more than a billion consumer interactions is substantial for a single vertical.

Coverage is concentrated on United States regulated collections, and the compliance architecture is built specifically around that statutory regime, so a lender outside it would be buying a product engineered for someone else's rules. Competitor comparisons place its sub industry coverage across buy now pay later, healthcare, auto and insurance receivables, and those are recorded as leads rather than findings because each appears on a rival's own domain.