What happens when the model is wrong
Deterministic rules run before any model does, and roughly 69% of narration-audit records settle on rules alone. Below the confidence threshold the workflow escalates to a human rather than guessing.
The same handful of pain points, in every conversation with teams in this industry.
Two analysts can look at the same application and reach different calls, and there's rarely a clean way to show why one call was right and the other wasn't.
KYC and AML screening often runs on a sample, because reviewing every new account the way policy requires would take longer than the business can wait.
Fraud and audit teams spend most of a case's cycle assembling the file, not deciding what's in it, and the backlog grows while they do.
Portfolio review prep eats the hours that should go to the client sitting across the table, not the system behind them.
Not a bolt-on feature for each pain point. One governed platform that removes the root cause of all of them.
Zara builds the same six-role team for every application. It doesn't have a good day or a bad day, and it doesn't disagree with itself between reviewers.
Deterministic rules run first and extend coverage to every account, not just the ones that happened to get pulled for review.
The defence file builds as the case moves through Zen Pilot, not after someone finally asks a regulator to reconstruct it.
A SPAR adversarial pass checks the first conclusion against how a similar case was handled, before a named officer signs off on either.
Illustrative, drawn from the shape of workflows modelled for this industry. The real number depends on the case, the team, and how it's configured.
across credit, fraud, and onboarding workflows modelled for banking teams.
without adding headcount to the compliance function.
when the first pass already carries a documented, checked rationale.
so relationship manager time goes to the client, not the file.
Specific workflows, worked through end to end.

If yours isn't listed yet, that's a conversation worth having with Zara.
RBI, CBUAE, FCA, and the data residency requirements that come with each. Deployment runs as SaaS, private cloud, on-premise, or fully air-gapped, with model choice, OpenAI, Anthropic, Google, open-weight, or your own, set per agent at runtime. GDPR compliant, SOC 2 and ISO 42001 certified, regardless of which option you choose.
Most banks evaluating this work compare an internal build team against a horizontal, no-code platform. Neither ships with banking-native agents, jurisdiction-aware regulatory packs, or an adversarial check built into the cycle. Zentis starts with all three already in place.
The full detail lives on the Trust page.
Deterministic rules run before any model does, and roughly 69% of narration-audit records settle on rules alone. Below the confidence threshold the workflow escalates to a human rather than guessing.
Your choice per deployment: SaaS, private cloud in your own subscription, on-premise inside your data centre including the models, or fully air-gapped with no egress at all.
Every decision carries a defence file: source, rule applied, confidence, adversarial record, and the named approver. Regulatory overlays ship per geography, versioned alongside the workflow.
Banking & Credit work goes through SPAR like everything else. The domain knowledge changes; the governance does not.
Model choice is configuration, per agent, at runtime — including your own on-premise model.
A working session, on your data, with the challenge step switched on. If the controls it extracts are wrong, you'll see that immediately — which is the point.