AviaryAI

AviaryAI runs a proprietary LLM trained on financial services data rather than adapting a general purpose one. It also operates a second distinct model class, independent safety models whose function is supervising the agents for compliance and call transparency. Conducting a natural outbound conversation that completes a card activation or negotiates a repayment is not achievable without models. Without them, untrained staff are left making calls nobody wants to make, which the company names directly as the status quo it replaces.

Four distinct mechanisms operate at different points. Independent safety models form a separate model class supervising the agents for compliance and call transparency, adaptable as regulation changes. That is machine supervision of machine behavior rather than a promise of care. During calls, members are escalated where escalation is warranted.

Staff are notified through their own collaboration tool for approvals, warm transfers and exceptions, so a person is reachable rather than theoretically available. Afterwards, every call is audited without sampling. The stated philosophy matches the architecture: the technology should let staff concentrate on relationships rather than replace the relationship.

The safety model design is the substantive answer. It places a separate model class between the agent and the member specifically to enforce transparency and compliance, the verifier pattern some document vendors use, applied to live conversation. Auditing every call rather than sampling gives complete coverage for retrospective review, and several engagement metrics are published, including the proportion of answered calls that progress beyond a greeting.

What is missing is direct measurement of the models themselves. No accuracy, error, misstatement or hallucination rate is published, and for agents that complete card activations and discuss debts, that is the figure a risk function would want.

The predecessor product is the strongest evidence and it belongs to the same team: an outbound debt negotiation assistant that helped more than 80,000 consumers resolve and reduce collections debt, which is real conversational volume with real consumers before this company existed.

Performance figures are numerous and specific, including a 42 to 47 percent contact rate, an 84 percent reduction in cost per call, a 70 percent conversion rate on fee income campaigns, a claimed 174 percent improvement in collections success against a stated national average, and from its accelerator launch, 26 percent of calls producing successful interactions with 72 percent of answered calls engaging beyond a greeting. Revenue is reported at roughly $1 million annually with ten employees and $3 million raised. No institution is named as a customer.

The boundary is architectural rather than contractual: models are private and proprietary, trained on financial services data rather than shared with a general purpose provider, which is the strongest form of separation available to a voice vendor and is stated plainly as the reason institutions can use generative technology at all. What is unaddressed is provenance in the other direction.

The founding team's earlier product conducted debt negotiations with more than 80,000 consumers, and nothing states whether those conversations trained the current models, whether one institution's member calls improve agents serving another, or what a customer can decline.

Two structural properties do real work here. The company runs private models rather than sending member conversations to a third party interface, which removes the exposure most voice vendors leave open. It also audits every call afterwards, so a complete record exists of what was said to whom. A security attestation is claimed alongside these. No data processing agreement, retention schedule, recording policy or subprocessor list is published, and outbound calls to members about debts and products generate sensitive material at volume.

A SOC attestation is claimed, and the company's own materials conflict on its status. One page describes the certification as held. Another states the team is currently working toward obtaining the Type II version, which most plausibly means the point in time report exists while the period of operation report does not. The weaker claim is the verifiable one. Even so, an attestation program in progress at a ten person company is real disclosure, and it is published alongside named compliance guardrails and complete post call auditing.

Outbound calling is designed around the Telephone Consumer Protection Act (TCPA), which is named repeatedly and treated as a design constraint rather than a footnote. It is the correct and specific regime for automated outbound calling and for collections contact. Right party contact capability exists to satisfy it. Full regulatory auditability is claimed, and the safety models are described as adapting to regulatory change. No supervisor is named, no examination or state regime appears, and the product is sold to insurers without any reference to insurance conduct rules.

Right party contact and statutory compliance protect the contact side, addressing the harassment risk that automated dialing creates for people in arrears. That is the real everyday harm in outbound collections. The unaddressed side is what the calls sell. A published 70 percent conversion rate on non interest income campaigns describes a highly effective automated machine for placing fee generating products with members.

The same platform calls people about collections, so persuasion capability is pointed at both the indebted and the profitable. Nothing describes suitability checks, campaign exclusions for members in difficulty, or analysis of who converts and who is called most.

There is no published guarantee, indemnity or correction process. The institution is well equipped, since auditing every call gives it a complete record to defend or to correct from. The member benefits because warm transfer to a person is designed in rather than absent, and that is where it stops.

Nothing states whether members are told they are speaking to an automated agent, what happens if an agent misstates a balance or a product term, or how someone disputes what occurred on a call about their debt.

The company owns its model dependencies, which keeps the picture clear. It describes a proprietary LLM trained on financial services data and a separate class of independent safety models, and states explicitly that this is not a general purpose model repackaged. That tells a buyer the chain is internal rather than resting on an external interface that could change beneath them. What is not disclosed is any base model the proprietary version was built from, the hosting arrangement, the telephony carrier, or a subprocessor list.

The company identifies the hard part accurately, noting that automating post call workflows such as opening accounts or activating cards is difficult because several systems must work together, and it builds toward that rather than stopping at the conversation. Staff notification runs through a named enterprise collaboration platform, which matters because it puts approvals and warm transfers where staff already work.

Implementation in under two weeks with minimal technical resource is itself an integration claim suited to institutions without engineering teams. No core banking, member management or telephony system is named.

Running private models implies infrastructure control, and where that infrastructure sits is not described. There is no published hosting provider, region selection, residency commitment or private deployment option. Depository institutions have supervisors who expect documented third party arrangements covering member call recordings, so that is a routine question published material does not answer.

Actual pricing figures exist. Pricing is reported as tiered subscriptions running up to $19,000 alongside usage fees, giving a buyer both the model and a ceiling, and annual revenue of around $1 million indicates the scale of the customer base behind it. Implementation cost is addressed separately and concretely, at an average of under two weeks with minimal technical resource. That is the second question after price for an institution with no engineering capacity. The pricing detail comes from third party reporting rather than a published schedule.

The buyer definition is deliberately precise, covering credit unions, community banks, regional banks and insurance companies, which is a coherent segment sharing constrained staffing, member relationship models and identical regulatory exposure.

Functional coverage across that segment is broad, spanning collections, onboarding, loan document chasing, card activation, cross sell, retention, deposit renewals and pre migration outreach, with a separate internal knowledge product for staff and real time multilingual support. Coverage is domestic and excludes large national banks by design.