Financial Services Use Cases
Five candidate use cases for a controlled intelligence layer in regulated financial operations
The use cases below are the recommended starting points for a first institutional deployment. They are deliberately internal, low-risk and reviewable. Each use case is in development: the technology basis exists in the GENIE platform; the OmniOmni integration is being built now. Every case states where the assistant stops and a person decides.
Pilot metrics are listed as measurement structures only. Target values are defined per pilot with the deploying institution and are not published here.
1. Operations intelligence: case and conversation memory
In development — technology basis: fact memory, revision history, structured call artefacts (Available)
| User and problem | Relationship managers and operations staff lose context between calls, tickets and handovers. Agreed steps live in personal notes. |
| Workflow | Conversations and notes are captured; facts, decisions and open items are extracted, linked to their source and kept with a revision history. A colleague taking over sees the current state and its history. |
| Data and integrations | Conversation content and case notes. Reading case context from OmniAsset or OmniEagle is an integration in development. |
| Agent action | Extracts facts, proposes a summary, lists open items. |
| Human decision | The responsible person reviews extracted facts before they are relied on for a case, and confirms handovers. |
| Controls and evidence | Source-linked facts, revision chain, sensitivity-scoped recall, per-account scoping, tool invocation log. |
| Pilot metrics (structure) | Share of handovers with a complete context sheet; corrections per extracted fact; time to first useful context on a re-opened case. |
| Exclusions | No customer-facing decisions; no automatic writes to any system of record. |
2. Knowledge and policy assistance with sources
In development — technology basis: research tool with source references (Implemented, not enabled)
| User and problem | Staff need answers from internal manuals, product documentation and policies, with a reference they can check. |
| Workflow | The assistant searches authorised internal sources (for example OmniWiki content), answers with references, marks uncertainty and hands over to the responsible department when sources are insufficient. |
| Data and integrations | Internal documentation. Ingestion of OmniWiki content is an integration in development; no content interface exists today. |
| Agent action | Retrieves, summarises, cites. |
| Human decision | Interpretation and application of a policy remain with the responsible function. |
| Controls and evidence | References per answer, logging of retrieval calls, access restricted to authorised sources. |
| Pilot metrics (structure) | Answer rate with at least one valid reference; escalation rate; reviewer-assessed correctness on a sampled set. |
| Exclusions | No legal or regulatory interpretation presented as authoritative; a behaviour that declines to answer below a relevance threshold is a design in development, not an existing mechanism. |
3. KYC and compliance case preparation
In development — technology basis: structured artefacts, tool sensitivity tiers, logging (Available)
| User and problem | Compliance teams assemble documents, findings and history from several places before a case can be assessed. |
| Workflow | The assistant collects the documented status of a case, reads existing OmniEagle findings and process history, identifies missing items and prepares a draft case file with references. |
| Data and integrations | Read access to identity and transaction assessment results and their history in OmniEagle (documented read interfaces exist; the OmniGENIE integration is in development). |
| Agent action | Assembles, structures, flags gaps, drafts. |
| Human decision | Risk assessment, overrides and the release of the case remain with the compliance officer. OmniEagle remains the detection and scoring component. |
| Controls and evidence | Every source in the draft is referenced; the draft is labelled as such; no scores are produced or altered by the assistant. |
| Pilot metrics (structure) | Completeness of draft case files as judged by reviewers; preparation time per case; number of missing-document flags confirmed as correct. |
| Exclusions | No autonomous risk scoring, no suspicious-activity reporting, no decision on customer acceptance. Whether a given assistance use case falls under a preparatory-task exception of the AI Act is assessed case by case by the institution's legal and compliance function. |
4. Issuer and client interaction agent
In development — technology basis: real-time voice, telephony, structured call artefacts (Available)
| User and problem | Issuers, investors and partners raise routine questions and requests by phone or text outside office hours or across teams. |
| Workflow | The assistant discloses that it is an AI system, takes the request, classifies it, records a structured note and routes it to the responsible team or a human operator according to rules. |
| Data and integrations | Contact and request data; routing rules. Reading issuer or investor context from OmniAsset is an integration in development. |
| Agent action | Takes, classifies, documents, routes. |
| Human decision | Any substantive answer about a specific security, transaction or account, and any commitment, is given by a person. |
| Controls and evidence | AI disclosure at the start of the interaction; call notes and tasks; handover rules; call-recording obligations remain with the institution and are met with its own recording infrastructure. |
| Pilot metrics (structure) | Share of requests correctly classified; handover time; caller-reported satisfaction on a sample. |
| Exclusions | No investment advice, no execution of orders, no identity verification by the assistant. Conversation and safety behaviour are currently optimised for German (see Limitations). |
5. Controlled workflow orchestration
In development — technology basis: graded agent routing, delegation model, approval token pattern (Implemented, not enabled)
| User and problem | Operations staff perform recurring multi-step workflows (for example preparing a corporate action) that require data from several components and formal approval before execution. |
| Workflow | After technical hardening, the assistant prepares the inputs for an OmniOmni action, validates them against documented parameters and hands the prepared action to the existing approval path. Execution happens only after a human approval through the component's authorised mechanism. |
| Data and integrations | Read interfaces of OmniAsset and OmniSafe; the Proof-of-Action approval flow through OmniPersona. All of these integrations are in development. |
| Agent action | Prepares and validates a draft action. |
| Human decision | Approval and execution stay with the authorised user and the component. Fully autonomous execution is disabled for financial-market actions in the target design. |
| Controls and evidence | Action tiers (in development), approval token bound to the exact payload, logging of the prepared action, the approval and the outcome. |
| Pilot metrics (structure) | Share of prepared actions approved without change; rejected preparations and their reasons; zero unauthorised write attempts as a hard exit criterion. |
| Exclusions | This use case is last in sequence. It requires the controls in development on Security, Governance & Human Oversight to be built and tested first. |
Sequencing
Start internal
Operations intelligence (1) and knowledge assistance (2) involve no customer-facing decisions and no writes to systems of record.
Add case preparation
KYC and compliance case preparation (3) adds read integrations and a clear human release step.
Open the interaction channel
The issuer and client interaction agent (4) adds disclosure, routing and recording requirements.
Orchestrate only after hardening
Controlled workflow orchestration (5) is gated on tenant isolation, enterprise identity, action tiers and the OmniPersona approval integration.
Hard exit criteria for any pilot: an unauthorised write access to a system of record; a successful prompt-injection breakthrough in red-teaming; latency above the threshold agreed with the institution. Values are set jointly before a pilot starts.