Backbase

Channels, journeys and onboarding run without the models, as they did for banks that bought the platform before any agent existed, so the agents sit on top of a working engagement platform. Backbase rebuilt that platform as an AI native Banking OS in April 2026 and added Kasisto's agentic platform in June 2026.

Sentinel requires a decision token before any customer, employee or software agent acts, checks each action against bank policy before execution, logs it and lets authority be revoked in any domain. A confidence threshold and sampling audits of automated decisions are not published.

Banks set which information, tools and actions agents may use and where human approval is required, any consequential request needs the customer's explicit confirmation, and exceptions route to an authorized employee.

Backbase AI Research presented a peer reviewed study at the 2026 meeting of the Association for Computational Linguistics on a 12 billion parameter banking model trained to say it does not know when evidence is missing. Aggregate accuracy and override rates for production agents are not published, and the work is not mapped to a supervisory framework.

Over seven months at a large United States institution the model raised query resolution 7.1 points across 3,297 sampled queries, scored 6.21 against 5.72 for GPT-4.1 on independent evaluation and improved citation grounding to source documents, and every agent action carries a decision token while every human override is recorded.

I&M Bank moved onboarding from one day to five minutes and cut customer acquisition cost 64 percent and cost to serve 72 percent on Backbase, and Techcombank acquired 1.9 million customers digitally in 2024. These figures measure the digital banking platform, not its AI agents.

The agent results are not tied to a named bank: one of South Africa's four largest banks handled 22 million interactions with Conversational Banking and raised self service containment from 20 to 70 percent, and a seven month trial at a large United States institution raised query resolution 7.1 points across 3,297 sampled queries.

Whether feedback from resolved cases stays inside one bank or improves a model shared across customers is not stated, and neither are retention terms for conversation data. Every resolved case, human override and completed journey feeds back into the models, a learning loop presented as a headline feature.

One shared semantic record of every customer, account and case, plus conversational history from the acquired assistant, concentrates identified financial behavior in a single layer. Retention, minimization and the position of bank customers who populate that layer are unaddressed.

A current report date, the auditor, coverage beyond managed hosting and a trust center are not published. The managed hosting service, used by more than 40 customers, is SOC 2 Type 2 attested, a named report type for that delivery option.

Customers hold the banking authorizations, and Backbase claims no financial license or supervisory test of its model layer. It presents the authority layer as the way a bank can show a regulator what an agent was allowed to do, without a regulator on record endorsing that design.

Bias testing or an AI management system certification is not published for agents that handle disputes, onboarding and service. Governance here means authority: decision tokens, policy checks and revocable permissions for every agent. The peer reviewed study of its banking model measures accuracy and refusal rather than outcomes across groups of customers.

Who carries the loss when an agent executes a wrong transaction or resolves a dispute badly is not published, and a customer's route to contest an outcome runs through the bank. What exists is a record: each action carries a decision token, Conversational Banking keeps the context, decision and outcome of every request for review, consequential requests need the customer's confirmation, and a case passes to human service with its context intact.

Which models and hosts sit behind the production agents is not stated, so a bank cannot yet trace what reasons over its customer records. Kasisto's agentic platform came into the operating system with the June 2026 acquisition, and Backbase AI Research trains its own banking model, a 12 billion parameter model tuned from an unnamed base.

The platform sits above core banking, payments and card systems without replacing them, and it can land on any existing installation, Backbase or not. Every actor works from one shared record, and Conversational Banking runs alongside a bank's existing online and mobile banking.

Hosting regions and residency terms for the managed service, and where the model layer processes data, are not stated. The deployment choice itself is wide: on premises, public, private or hybrid cloud, or managed hosting on Microsoft Azure with multi zone replication and disaster recovery, which more than 40 customers use.

Enterprise licensing is priced on request, with no rate, module price or minimum. The license is one part of a multi year architectural commitment that also carries implementation and change costs, so total cost of ownership cannot be estimated from anything public.

More than 120 banks run the platform, among them Navy Federal Credit Union, VeraBank, Techcombank, TPBank, I&M Bank across five East African countries and Jet Bank, Albania's first digital only bank. Retail and commercial banking, credit unions and private banking are covered, and Backbase works from 16 offices on five continents, with headquarters in Amsterdam and Atlanta.