Omilia

The removal test leaves a keypad menu. Every layer is a model the company built itself, spanning speech recognition, voice biometrics, dialogue management and speech synthesis, with a self learning agentic layer above them that learns across the whole customer journey including live agent interactions. More than 300 models are trained specifically on banking and finance intents. Detecting a synthetic caller in real time or holding an unscripted conversation that completes a payment is achievable no other way.

Automation extends to money movement, with payments processed entirely within automated interactions and no routing to a person, and the platform's headline metric is containment, meaning success is defined as the caller never reaching a human. Three things temper that.

Authentication runs continuously in the background through voice biometrics so identity is verified before anything consequential happens, full audit trails are retained for every transaction, and an investor describes the platform's distinguishing property as glass box auditability, the deliberate opposite of an opaque system. The company states it preserves the human touch where it matters most without describing where that boundary sits or what triggers escalation.

This company publishes what almost nobody in this index does: direct model performance figures rather than business outcomes. Intent understanding accuracy is stated at 97 percent and word error rate at 2 percent, both measures of whether the system understood correctly rather than whether it saved money, and one deployment's semantic accuracy above 90 percent was assessed by a named independent evaluator rather than self reported.

Customer specific figures follow the same pattern, including 98 percent voice accuracy at a named logistics client. Its investor identifies glass box auditability as the structural differentiator, and analytics tooling exists specifically to identify where the system failed to resolve a query and why.

Four financial institutions are named including two of the largest United States card issuers and a major Canadian bank, alongside enterprises in logistics, utilities, automotive, government and food service. Both major analyst houses recognised the platform within three months of each other in 2026, one as a leader in its conversational AI evaluation and the other as a visionary in its quadrant.

A 67 million dollar growth round followed, adding to earlier institutional backing, and two channel partnerships extend distribution across the Americas, Europe and German speaking markets. The company has operated since 2002, and published outcomes are specific and repeated across deployments.

No data boundary statement was located and the platform's own description makes the question acute. It is stated to have been trained on billions of real customer interactions, and its self learning layer improves by learning across the entire customer journey including live agent conversations, which is precisely the material a bank would consider confidential.

Nothing states whether learning is tenant isolated, whether one institution's calls improve models serving a competitor, what happens to voice biometric enrolments if a customer leaves, or what a buyer can decline.

The payment path is specified in unusual detail: capture certified to the highest card industry tier through both speech and keypad, with real time redaction ensuring card data is never stored or transmitted in clear text, and full audit trails retained for compliance. That directly addresses the hardest privacy problem in voice, which is that a caller reading out a card number puts it into a recording.

Held at B because voice biometric templates are biometric data governed by separate and stricter regimes in several jurisdictions, and no retention, consent or deletion policy for them was located, nor any data processing agreement or subprocessor list.

Certification to the highest tier of the card industry data security standard is held and stated, which is an audited assessment against a defined control set rather than a claim of alignment, and it is the relevant one for a platform capturing card payments by voice. Anti fraud capability adds to the picture, with deepfake, synthetic voice, spoofed number and replay attack detection operating in real time. No general information security attestation or trust centre was located, which is what an enterprise buyer would request alongside the payment certification.

One standard is named precisely and it is the right one for the function: payment capture certified at the highest tier of the card industry data security standard, which is an audited certification rather than an alignment claim, supported by real time redaction and retained audit trails. Partner material notes that new European regulatory requirements are accelerating demand for production grade agentic systems without identifying them. No financial supervisor, conduct rule or biometric data regime is named, which is the notable gap given that voice biometrics are separately regulated in several of the markets served.

The exposure here is acoustic rather than financial and it is unaddressed. Speech recognition and voice biometric systems are documented across the field to perform unevenly across accents, dialects, age, and speakers with atypical speech or non native fluency, so a published two percent word error rate is an average that conceals who sits at its tail.

Those same customers are the ones a containment focused system will fail to contain, and if biometric authentication does not match them they face additional friction proving who they are. No subgroup performance data, accessibility analysis or fallback policy for repeatedly failed recognition was located.

No guarantee, indemnity or correction process was located. The institution is well served by full audit trails and published accuracy figures. The caller has nothing described, and their position matters more here than at most voice vendors because the system authenticates them biometrically and moves their money: nothing states what happens when voice authentication wrongly rejects a legitimate customer, how a disputed automated payment is investigated, whether callers consent to biometric enrolment or can opt out, or how someone reaches a person when containment is the metric being optimised.

The stack is fully proprietary and disclosed component by component, with the company naming its own speech recognition, voice biometrics, dialogue management and speech synthesis engines individually rather than describing a general capability, and a systems integrator partner independently confirms it as a fully proprietary stack proven in demanding regulated environments.

For a bank that is the material fact: there is no third party model provider in the path whose pricing, availability, versioning or data handling could change beneath the deployment, which is the dependency almost every other conversational vendor in this index carries and does not disclose.

Integration into contact centre infrastructure is demonstrated rather than claimed, with a named bank deployment describing integration with its existing contact centre platform vendor, and support for both speech and keypad input meaning the platform coexists with legacy telephony rather than requiring its replacement. Two systems integrator partnerships extend delivery capability across regions. What is absent is any named core banking, card or payment system, and no developer documentation was located, so the depth of connection to systems of record is undescribed.

The platform is offered on premise as well as in the cloud with modular options deployable within days, and an on premise path is the strongest sovereignty answer available to a voice vendor because it means recordings, biometric templates and payment interactions never leave the institution's own environment. That is what allows adoption by banks in strictly supervised markets. Held at B because no hosting provider, region selection or residency commitment is published for the cloud path, which is what most customers will actually take.

No pricing, packaging or basis of charge was located. An investor cites cost predictability as a structural advantage over competitors, which is a claim about the pricing model rather than a disclosure of it, and matters in this category because usage based voice pricing makes budgets unpredictable at exactly the point automation succeeds. Deployment speed is quantified at days rather than months, which addresses implementation cost.

Financial coverage is genuine rather than a vertical landing page, with more than 300 models trained on banking and finance intents, telephone banking journeys spanning balances, transfers, payments and account management, and named deployments at card issuers and retail banks. Geographic reach spans North America, Europe, Latin America and German speaking markets through partners.

Held at B because the company is not a financial services specialist: its customer list includes logistics, utilities, automotive, government and restaurant brands, so financial services is its strongest vertical rather than its only one.