Skip to content

Responsible AI, Limitations & Availability

Known limitations, transparency, non-roles and the phased plan without dates

Principles

  • Human decision. OmniGENIE prepares; people decide. Every write or transactional action requires an explicit human approval before execution.
  • Transparency towards people. Where OmniGENIE interacts directly with natural persons by voice or text, it is designed to disclose that they are interacting with an AI system.
  • Traceability without exposing reasoning. Operational artefacts are recorded; the model's internal reasoning is neither published nor stored as an audit artefact.
  • Purpose-bound data. Memory serves the defined operational purpose; deletion is supported subject to statutory retention.
  • Honest status. Every capability carries a status label. Nothing is described as available that is not active in the productive platform today.

What OmniGENIE does not do

OmniGENIE does not act as, and must not be described as:

  • a crypto securities registrar or register operator — that role stays with the licensed partner operating OmniAsset;
  • a custodian — that role stays with the licensed custodian operating OmniSafe;
  • a compliance officer or money-laundering officer, or a replacement for either;
  • an investment adviser — it gives no investment recommendations;
  • a credit decision-maker — it does not set credit lines or risk classes;
  • an approved, certified or regulator-endorsed system in its own right.

Current limitations

These limitations describe the GENIE platform today. They are the reason several capabilities on the previous pages are labelled in development.

LimitationConsequencePlanned response
Conversation and safety-critical language understanding are optimised for German. Several safety and control behaviours rely on German-language rules and are not validated for other languages.Non-German interactions are not supported for safety-relevant behaviour.Multilingual control flow is a documented gap; no date is published.
Sign-in uses consumer Google OAuth only.No federation with an institution's identity provider.Enterprise identity federation (In development).
Data isolation is per account, not per organisation.No institution-level tenant model.Institution-level isolation (In development).
No general policy engine or action-tier model for financial actions; approval and routing mechanisms exist for other purposes.Existing mechanisms cannot be cited as bank-grade approval gates.Action tiers and Proof-of-Action approval (In development).
Personal-data redaction exists for one intake feature only, not pipeline-wide.Conversation content reaches model providers for inference.Pipeline-wide data-loss prevention (In development).
No legal hold; no automatic retention schedule.Deletion cannot be blocked technically for retention obligations.Legal hold and retention classes (In development).
One language-model family via a gateway; automatic fallback exists for text-to-speech only.Provider dependency for the language model.Model provider failover (In development).
Kill switches are manual (feature switches, administrative allowlist).No automatic circuit breaker.Automatic circuit breaker (In development).
Operational logs record argument digests, not full payloads.Exact argument values are not reconstructable from the log alone.Complete per-action artefact set (In development).
No packaged on-premises or customer-VPC deployment.Deployment is a self-hosted stack run by one operator.In development (In development).
Automated continuous integration on every commit is not configured.Tests run through a documented local script.Release process with evaluation reports (In development).

Availability by capability group

Available
Active in the productive GENIE platform today (technology basis).
Implemented, not enabled
Exists in code and tests, but is switched off or narrowly scoped in production.
In development
Being built now as part of the OmniGENIE integration into OmniOmni. Not yet generally available.
Roadmap
Planned for a later phase. Not available; no dates are published.
Capability groupStatus
Operational memory (extraction, revision history, sensitivity, correction, erasure)Available
Real-time voice, telephony, structured call artefactsAvailable
Tool sensitivity tiers and invocation loggingAvailable
Graded agent routing, delegation model, approval tokenImplemented, not enabled
Voice recognition of enrolled persons, group and guest modes, web researchImplemented, not enabled
OmniOmni integrations (all components)In development
Action tiers, enterprise identity, tenant isolation, legal hold, circuit breakerIn development
Meeting companionRoadmap
OmniGENIE as an OmniOmni componentIn development

Phased plan

Development is organised in phases. No availability dates are published. Phases advance only when the controls they depend on are built and tested.

Phase 1 — Foundation and governance

Institution-level isolation, enterprise identity federation, action tiers for observe, draft and recommend, log export, the first two use cases (operations intelligence, knowledge assistance).

Phase 2 — Financial operations preparation

Read integrations with OmniAsset and OmniEagle, KYC and compliance case preparation, prepared actions handed to the Proof-of-Action flow in OmniPersona.

Phase 3 — Interaction and coordination

Issuer and client interaction agent with disclosure and routing, controlled workflow orchestration, multilingual control flow, meeting companion review.

How to get involved

OmniGENIE is being built with institutions, not for them in isolation. Design partners help define the first use cases, agree a pilot scope with hard exit criteria (no unauthorised write access, no successful prompt-injection breakthrough in red-teaming, latency within the agreed threshold) and review the control catalogue with their own risk, compliance and information-security functions. Contact ecrop.

Naming

"OmniGENIE — Agentic Intelligence for Financial Operations" is a working title. Alternatives are being evaluated for clarity, suitability for banks and fit with the existing component names. No trademark availability is claimed.

This page describes technical and organisational capabilities and their status. It is not legal advice. Institutions should consult qualified legal counsel to confirm the regulatory classification of their specific use case.