Kasisto

The removal test leaves a management dashboard and a set of connectors. The core asset is a proprietary language model fine tuned exclusively on banking conversations, policies, regulatory filings and financial knowledge sources rather than adapted from a general purpose model, and everything the platform does depends on it: conversational agents handling account enquiries and money management, generative answers drawn from an institution's knowledge base, and the newer agentic layer executing multi step work inside banking systems. An independent analyst characterises the company's distinguishing choice as staying focused on banking while the large platform providers pursued general artificial intelligence.

The safeguard here is the rarest in this lane and it is built into the model rather than wrapped around it. The proprietary language model follows what the company calls an accuracy first design and is trained to answer that it does not know when information is unavailable, with the stated reasoning that incorrect guidance in a financial setting creates legal and compliance exposure.

Calibrated abstention addresses the failure mode that actually matters for a banking assistant, which is not silence but confident error. Around it sit auditable agents embedded into banking systems, a central dashboard for managing and monitoring agent performance, and an orchestration framework through which one institution runs an array of assistants under common control. The chief executive framed the launch of the agentic platform against exactly this problem, that most artificial intelligence cannot be trusted in banking.

Three design decisions reduce risk before any monitoring is applied. The model is domain restricted, trained only on banking material, which narrows the space in which it can be wrong. It is designed accuracy first with explicit abstention when information is unavailable, which is a direct control on fabrication rather than a caution around it. And the agents are described as auditable, with a dashboard exposing performance across an institution's estate.

That combination is stronger than most in this lane. What is missing is measurement: no accuracy, containment quality, escalation precision or hallucination rate is published, and abstention rates are not disclosed either, so a buyer cannot tell whether the safeguard fires usefully or too often.

The named customer list spans the full size range and four continents, including one of the largest United States banks, a major British headquartered international bank, Australia's oldest company and first bank serving more than twelve million customers, and a large Canadian banking group, alongside three of the biggest United States credit unions and named community institutions in the Midwest.

One of those relationships is stated to date from 2020 and involves an array of assistants managed through a single orchestration framework, which is duration and depth rather than a logo. More than 90 million dollars was raised before the company was acquired by a banking software group in 2026. An independent conversational AI analyst describes it as demonstrably one of the longest standing specialists in this field. What is absent is quantified outcome: no containment, resolution or handle time figure is published.

The training description raises the boundary question directly and then leaves it unanswered. The proprietary model is stated to be trained only on banking industry conversations and data, and banking conversations are by definition exchanges between institutions and their own customers, so whether that corpus draws on deployments at the named customers is the material question and nothing addresses it. Those customers include direct competitors in the same markets. A partial answer exists for one component, since the generative answer product is described as drawing on the institution's own knowledge base, but that concerns retrieval rather than training.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located. The payload is conversations between financial institutions and their customers covering balances, transactions and money management, together with what the company describes as valuable information about each customer's financial situation surfaced through those interactions. Nothing published states how long conversation records persist, how they are separated between institutions, or what a customer is told about the processing.

No attestation, certification, trust centre or enumerated framework was located, though bank grade performance and secure agents are both asserted. Deployments at several of the world's largest banks mean security assessment has been passed repeatedly at the most demanding standards available, and none of that assurance is published, so a prospective institution begins its own review with nothing to read.

No supervisor, statute or instrument is named. Regulatory filings appear as a category of training material and compliance is cited as a design goal for the agentic platform, but neither identifies an obligation. The gap is material for this function, because an assistant answering customer questions about accounts and payments operates inside electronic transfer error resolution rules, prohibitions on unfair or deceptive practices in customer communication, and accessibility requirements, and none is engaged in published material.

Conversational systems vary in how well they understand different speakers and writers, with dialect, phrasing, literacy and non native usage all affecting comprehension, and this is the twelfth recorded instance of that finding in this index. Two features work against the harm. Operating in English and Spanish from day one is a genuine access property in a market where a large population banks in Spanish and is often served worse.

And the abstention design means a customer the system does not understand receives an admission rather than a confidently wrong answer about their money. No accuracy across populations, languages or channels is published, and no analysis of who the assistants serve least well was located.

No guarantee or indemnity was located, and the design decision that matters most for the customer is the abstention behaviour. A system that answers it does not know rather than guessing removes the specific harm this product could most easily cause, which is a person acting on wrong guidance about their own money, and the company states that reasoning explicitly rather than leaving it implied.

Auditable agents mean an institution can reconstruct what was said and why when a customer disputes an interaction. What is missing is anything after the fact: no correction process is described for a customer given wrong information, and nothing states whether a customer is told they are speaking to a machine.

The model layer is the company's own and the training corpus is described by category rather than left blank, covering banking conversations, institutional policies, regulatory filings and financial knowledge sources, which is more disclosure than most conversational vendors offer and means no external frontier provider sits in the path. The distribution chain is named, with three digital banking platforms pre integrated and further core banking and engagement partners identified. What is not disclosed is the origin of the conversational training material, no speech or language service is named for any component, and no subprocessor list or hosting arrangement was located.

Pre integration covers the digital banking platforms this market actually runs on, with three named providers between them serving a very large share of United States community institutions, and further partnerships named across core banking and payments software vendors. A customer engagement platform used by more than 600 banks and credit unions carries the technology into those institutions as well.

An orchestration framework lets a large bank manage multiple assistants across channels under one control layer rather than as separate deployments. Following acquisition, the platform is being folded into a banking operating system, which extends the integration surface further into the institution's core.

No hosting provider, region selection, residency commitment or private deployment option was located. The question is live because deployments span North America, the United Kingdom and Australia, each with its own expectations about where customer conversation data may be processed, and the acquisition by a European headquartered software group adds a further dimension that published material does not address.

No pricing, packaging or basis of charge is published. One comparative claim is made for smaller institutions, that community banks and credit unions can have the same assistants as the largest institutions at a fraction of the cost with no coding and negligible upkeep, which describes relative positioning rather than a rate. Nothing indicates whether charge falls per institution, per conversation, per agent or per end customer.

Coverage runs from the largest global banks to single branch credit unions on the same platform, which almost nothing else in this lane demonstrates with named customers at both ends. Geographic reach is genuine rather than aspirational, with deployments across North America, the United Kingdom and Australia.

Functionally it serves customers and members on one side and bankers and contact centre staff on the other, and the product is built for a specific institutional stress point, the surge in support demand during mergers, system upgrades and conversions. Bilingual English and Spanish operation from day one matters materially for United States retail banking coverage.