interface.ai

The product is conversation and action, with agents holding voice and chat exchanges that resolve requests end to end rather than routing through scripts or menu trees, built on language models paired with graph grounded retrieval from approved sources. Speech understanding, intent resolution, personalised response and task completion are all inference, and the company positions explicitly against the brittle scripts and decision trees that preceded it. Apply the removal test and an interactive voice response system remains, which is precisely the incumbent technology being displaced.

The handoff is the designed control and it is described properly: the agent acts as first point of contact, resolves routine inquiries, then transfers complex or sensitive matters to a human agent with full context carried across, which avoids the failure where a caller repeats themselves to a person who knows nothing. Compliance guardrails restrict risky content in the moment rather than in a later review, and analytics dashboards give supervisors visibility into agent performance. What is not published is the boundary itself: no stated confidence threshold that triggers escalation, and no sampling of the conversations the agent resolved alone.

Grounded and auditable knowledge from approved sources is the central transparency property, because a compliance officer can inspect what the agent is permitted to say rather than interrogating a model, and analytics dashboards report on agent performance in production. Customer reported automation rates are specific enough to be tested during a pilot.

What is missing is the vendor's own measurement: no accuracy, containment or misrecognition rates are published, no evaluation methodology or model documentation was located, and no statement addresses support for an institution's own model validation.

The strongest customer evidence in this lane. Multiple institutions are named individually, including several credit unions and a community bank, each with an executive quoted by title and a described deployment, and the outcomes are specific and attributed rather than aggregated: call handling up from 50 to 90 percent within two days at one institution, 40 percent call automation on day one at another, more than two thirds of member inquiries automated at a third, and up to 75 percent of queries resolved generally. Multiple credit unions are stated as live in production on the newer agentic platform, and the company describes a decade of operation across billions of banking interactions.

Two design decisions are stated and both address the failure mode that matters in this setting. Knowledge is grounded and auditable from approved sources only, so the agent answers from the institution's own material rather than from open ended generation, and compliance detection is trained on the credit union regulator's rules and the interagency examination standards to catch violations and restrict risky content during the interaction.

Training the bot from existing content in minutes is described as proprietary. What is absent is evaluation evidence, model provider disclosure, and any statement on whether conversations from one institution inform models serving another.

The platform handles authenticated banking conversations, meaning account details, balances and transaction history spoken aloud, and it performs risk based authentication combining device biometrics and caller identification forensics, which brings voice and device data into scope. Grounding responses only in approved sources limits what the agent can disclose, which is a real control. No published privacy framework, retention schedule for call recordings, subprocessor list or position on state biometric statutes was located.

No trust centre, enumerated certification list, attestation scope or audit period was located in this pass. Credit unions run vendor management programmes under examiner scrutiny and would have required attestations before member calls and core banking access were granted, so the control environment is certainly stronger than the published record. The grade reflects what a prospective buyer can verify without entering procurement.

interface.ai supplies technology and holds no licence, the expected posture, and its regulatory anchoring is specific to its buyer in a way generic conversational vendors cannot match. Compliance detection is trained on the credit union regulator's rules and the interagency examination council's standards, which are the two frameworks a credit union examiner actually applies, and accessibility compliance under the disabilities act is addressed directly as a customer outcome. Collections outreach is offered with a stated compliance posture, which matters because that activity carries its own conduct rules. No formal admission programme is evidenced.

Accessibility is treated as a first class outcome, with the platform positioned as delivering disabilities act compliance quickly and offering quality banking experiences for all customers, which is more attention to inclusion than most conversational vendors show. Measurement is absent.

Voice agents carry documented recognition variation by accent, dialect, age and speech impairment, and a member who cannot be understood is a member who cannot reach their own account, which falls hardest on the older and non native callers community institutions often serve. No per accent or demographic resolution rates were located, and the collections use case adds a conduct exposure the material does not address.

Two mechanisms give an institution practical protection. Grounding responses in approved sources means the agent should not invent an answer about a product or a rate, and compliance detection restricts risky content during the interaction rather than flagging it afterwards, which is prevention rather than reconstruction. Neither is a commitment by the vendor.

No accuracy guarantee, remediation term or published error rate was located, and no route is described for a member who could not be understood, was misinformed by the agent, or was subjected to automated collections outreach in error.

The architecture is described more openly than the chain, with the platform stated to combine modern large language models with graph grounded retrieval over approved knowledge, and proprietary technology named for rapid training from a customer's existing content. That tells a buyer the shape without telling them the parties.

No model providers are named, no speech recognition or synthesis vendors are identified despite voice being the flagship channel, and no subprocessor list discloses whose infrastructure processes authenticated member calls.

Integration reaches the systems that make a banking conversation useful rather than merely pleasant, spanning core banking, loan origination, customer relationship and knowledge management platforms, plus contact centre and online banking systems, and the platform is described as prebuilt to work with banking systems from day one.

That matters disproportionately in this segment, because a small institution cannot fund a long integration project, and customers cite integration ease with online banking and call centre systems as a selection reason. The employee copilot also works inside existing tools rather than adding another application.

Delivery is cloud hosted serving North American institutions, so cross border complexity does not arise as it does for global vendors here. Residency still matters given the content, since recorded member calls and authentication data carry retention obligations and examiner interest. No hosting regions, tenancy model, residency options or subprocessor chain were located in this pass.

No rates, tiers, billing unit or minimum were located. The buyer here is often a small credit union with a constrained technology budget weighing the platform against an outsourced after hours answering service whose cost it knows precisely, so a published price would let that comparison be made directly, and the displacement of third party overflow providers is one of the company's own stated benefits.

The focus is deliberately narrow and it is the strategy: credit unions and community banks, an underserved segment most conversational vendors treat as too small, with material written in the sector's own vocabulary of members rather than customers and a stated understanding of the credit union space cited by customers as a reason for selection. Channel coverage spans voice, chat, employee copilot and outbound collections. The boundary is equally clear, with nothing for large banks, insurers, wealth, lending institutions or capital markets, and no evidence of operations outside North America.