Bank AI (10)

Purpose and scope (4)
Internal AI assistant for bank employees
Helps prepare credit audits from customer files
Runs entirely inside the bank network
Status: proposed design for review
Architecture (4)
Components (7)
Browser interface for employees (1)
4 panes: upload, generated files, chat, answers
Application server (API? and workflows)
Local language model? runner (no external AI)
Search index? of bank policies and procedures
Database: users, audit log?, job queue
Document service: Word template to PDF
Monitoring and backup
One server folder is the only data area (8)
models/ read-only
users/{id}/ own conversation history
customers/{id}/input/ uploaded files
customers/{id}/generated/ created files
templates/ credit audit Word template
knowledge/ bank policies, read-only
audit-log/ append-only? record
batch?/ night jobs
Credit audit flow (7)
Employee uploads customer files
Virus scan before any processing
Model extracts fields into a fixed structure
Values validated, unknown stays empty
Word template filled, PDF created
Employee reviews and approves
Result saved in customer folder and logged
Connections (3)
Employees: browser over HTTPS, bank login
Other: log server and backup only
No connection to the internet
Data protection (8)
Customer data never leaves the bank network
No cloud services, no external AI
Telemetry? and update checks switched off
Customer data not used to train models
Encrypted in transit (TLS?) and at rest?
Each user sees only own history
Customer files only for assigned employees
Retention? time set by the bank
Network security (6)
Server in an internal network segment
Firewall?: default deny?, in and out (2)
Two layers: network and host firewall?
IPv4? and IPv6? both covered
Inbound: only HTTPS from employee network (2)
admin via jump host? with MFA?
everything else dropped, logged, alerted
Outbound: allow list? only (2)
DNS?, time, log server, backup, package mirror?
everything else dropped, logged, alerted
Models installed offline with checksum?
Test: a request to the internet must fail
Isolation (7)
The AI sees only one root folder
Separate non-root? OS users per service
Containers? mount only the root folder (2)
read-only where possible
read-only root filesystem, no extra rights
AppArmor? or SELinux? profile
Model has no shell and no file tools
App rejects paths outside the root
Test: reading /etc/passwd via the app fails
Access and audit (4)
Login through the bank directory (SSO?)
Roles and need-to-know? per customer
Audit log? of every action (2)
who, when, which files, which result
append-only?, sent to the bank SIEM?
Periodic access review
AI safeguards (7)
Employee approves every generated audit
Source quote next to each extracted value
Unknown values stay empty, never invented
Fixed output structure, validated by code
Documents cannot give the model commands (1)
model has no tools beyond fixed steps
Fixed model version, change needs approval
Anonymized? test cases for regular checks
Availability and operations (8)
Backups with a tested restore
Monitoring of load, disk, GPU? and errors
Warning when many users load the server
Night batch? for heavy jobs (17:00 - 08:00)
Patch policy? for OS and tools
Change control: test, approve, rollback?
Incident runbook? with named contacts
Exit plan?: data in open formats
DORA support (5)
ICT? risk management (2)
system in asset inventory, owner named
software list with versions
Incident reporting (1)
logs and alerts, runbook?, supplier support
Resilience testing (1)
yearly restore, load and security tests
Third-party risk (3)
contract per Art. 30, subcontractor list
data for the register of information?
exit plan?
Information sharing (1)
named contact for threat information
Bank decisions needed (6)
Number of concurrent users
GPU? hardware available
Languages of documents
Who approves a generated audit
Retention? time for histories
Is the system a critical function