Internal Operations in Financial Services
SanadAn Arabic-First Enterprise Assistant
An AI assistant concept that answers a financial institution’s staff from its own policies and procedures, citing its sources and respecting permissions.
01
The challenge
In financial institutions, operational knowledge is scattered across policy PDFs, intranet pages, regulatory circulars, ticketing systems and email. A customer-service or credit officer can lose valuable time hunting for the right procedure, and often finds two conflicting versions of the same document.
Generic chatbots do not solve this. Their Arabic is uneven when faced with rich morphology, spelling variants and mixed Arabic–English terminology, and they can fabricate an answer with complete confidence. Worse, they have no idea who is asking, so they may surface content the employee is not entitled to see.
On top of that sit non-negotiables: in-Kingdom data residency, the Personal Data Protection Law (PDPL), and the ability to trace every answer back to its source for audit and compliance.
02
Proposed solution
Sanad is built on retrieval-augmented generation with permissions enforced at query time. Access-control lists are synced from the identity provider and source systems, and no passage is retrieved unless the person asking could already read it. The model never sees what it is not allowed to say.
Retrieval blends BM25 over carefully normalized Arabic (unifying alef, yaa and taa marbuta forms, stripping tatweel and diacritics) with multilingual dense embeddings, then reranks with a cross-encoder. Documents are chunked by structure (headings, clauses, tables) rather than fixed token windows, so citations stay accurate to the paragraph.
Every request flows through a provider-agnostic model gateway that routes to self-hosted models in the Kingdom or to external services depending on data classification, behind guardrails for PII redaction and prompt-injection detection. The assistant can call a narrow set of tools, such as opening a ticket or summarizing a customer file, but anything that changes data requires the employee’s explicit confirmation.
03
Design rationale
Trust is the first design criterion. Every answer carries clickable citations that open the source passage highlighted, and the assistant says plainly when it lacks enough evidence instead of guessing. The source document’s last-updated date is shown so no decision rests on a superseded policy.
The chat surface is designed for Arabic from the ground up: RTL layout with bidi isolation for English fragments, numbers and codes; tolerance for colloquial phrasing and common spelling slips; and a clear Arabic typeface sized for long-form reading. Users can switch the answer language without rephrasing the question.
An “escalate to a specialist” action is always within reach, and one-tap answer ratings feed an Arabic golden test set used to measure quality before every model or index update. The interface is fully keyboard- and screen-reader-accessible, in dark and light themes.
04
Example architecture
An illustrative layering of the system, from the screens people use down to the infrastructure. A starting point we would validate during discovery, not a final design.
Layer 01
Experience
- Next.js chat workspace with citation viewer
- Embeddable widget for internal portals
- Admin console for sources & evaluations
Layer 02
Orchestration
- LLM gateway with policy-based routing
- LangGraph agent with scoped tools
- Prompt & version registry
- Human-in-the-loop confirmations
Layer 03
Retrieval
- Structure-aware ingestion (Apache Tika)
- Arabic normalization & tokenization pipeline
- Hybrid index: OpenSearch BM25 + pgvector
- Cross-encoder reranking
Layer 04
Governance & safety
- ACL sync from the identity provider
- PII redaction & prompt-injection filters
- Immutable answer audit log
- Arabic golden-set evaluation harness
Layer 05
Infrastructure
- vLLM inference on GPU nodes
- Kubernetes in an in-Kingdom region
- Keycloak SSO
- Per-request OpenTelemetry tracing
05
Relevant technologies
Tools we would consider for this concept. They indicate engineering ecosystems, not partnerships or certifications.
- Python
- FastAPI
- LangGraph
- vLLM
- OpenSearch
- PostgreSQL
- pgvector
- Apache Tika
- Redis
- Next.js
- TypeScript
- Keycloak
- Kubernetes
- OpenTelemetry
06
Key features
Source-cited answers
Each answer links to specific passages in the original documents, with their update dates, one click away.
Permission-aware retrieval
The assistant only sees what the person asking is entitled to see, based on permissions synced from source systems.
Arabic-aware search
Orthographic normalization, morphology-aware processing and hybrid search that handles synonyms and mixed Arabic–English terms.
Tools with a human in the loop
Narrow actions such as opening a ticket or drafting a reply, with employee confirmation before anything changes data.
Continuous evaluation
Arabic and English golden sets measure accuracy and citation quality before every release.
Complete audit trail
Questions, retrieved sources, answers and routing decisions are retained for review and compliance.
07
Potential benefits
These are design goals the solution is built toward, not measured results.
Aims to cut the time staff spend searching for procedures and policies
Consistent answers grounded in the approved version of each document
Lower risk of information leakage through permission-scoped retrieval
Faster onboarding of new staff onto institutional knowledge
A governance framework that eases audits and supports data residency
Related services & industry
Industry
Secure digital banking, payments and analytics platforms engineered for trust and scale.
Services involved
LLMs, agents, RAG and computer vision grounded in your data, with clear controls on quality, security and cost.
Reliable pipelines, governed warehouses, and executive dashboards that turn scattered data into decisions.
Security built into architecture, code, and delivery pipelines from day one, not bolted on after launch.
REST and GraphQL APIs, payments, webhooks, and event-driven integrations that make separate systems work as one.
Research-led product design, from user journeys to production-ready design systems for complex enterprise interfaces.
This is a concept project created by AMA GROUP to demonstrate our approach. It is not a client engagement: the interface uses sample data, and no client, logo or result is implied.
Have a similar challenge?
Share the problem you are facing. We will come back with a clear approach, scope and next steps.