Support · A typical retail bank
A failed transfer, across the database and the logs
-
The situation
A customer says a transfer failed at 14:32 yesterday. The support engineer needs the transaction record from core banking on Oracle and the matching API access log from three app servers.
-
How it works today
Two tickets: one to the DBA for an extract, one to operations for log files. A day or more. The extract often arrives as a full table dump with every column.
-
The request
One request from the support template: the transactions table, five columns, filtered to that account and day, plus the API access logs for 14:00 to 15:00. The team lead approves by rule.
-
What leaves the perimeter
Nothing. The engineer queries inside the sandbox; the answer is on their screen.
-
What the auditor sees
Who asked, what they asked for, who approved, and when the sandbox ran.
Template
Failed transfer investigation
- Sources
- core_banking (Oracle) + api access logs (SSH)
- Columns
- txn_id, account_id, amount, status, created_at
- Account
- 4471
- Window
- yesterday, 14:00–15:00
Why the logs matter
Logs are high volume, and support needs a sliver of them: one hour, a few fields. The request names the window, so the sandbox loads one hour from three servers, not ninety days from every host. Nobody copies a log archive to answer one ticket.
Analytics · A typical regional bank
Monthly product profitability on core banking
-
The situation
Finance wants profitability by product and segment. The data is in the core banking database, which may not be copied to the cloud.
-
How it works today
A quarterly extract job, owned by a DBA, that breaks when the schema changes. Or a federated query that hammers production at month end.
-
The request
A template joining accounts, products, and balances: selected columns only, last full month. The virtual table joins them for the analyst; row filters apply when the sandbox loads. The source is read once. The analyst's forty follow-up queries hit the sandbox, not production.
-
What leaves the perimeter
No raw rows. The analyst views query results in the explorer, inside your network.
-
What the auditor sees
The approved scope, the template version, and the requester.
Template
Monthly product profitability
- Source
- core_banking (Oracle)
- Tables
- accounts ⋈ products ⋈ balances
- Columns
- product_id, segment, balance, fee_income
- Window
- last full month
AI outside banking
Plain-English questions over data that stays put.
People ask in plain English in the explorer; AI agents connect over MCP. Either way, each question runs against an approved sandbox, so the answer can only draw on what was approved.
Network operations assistant
An engineer asks, in plain English, why a cell site dropped calls last night. The question becomes SQL over syslog files on SSH servers and the alarm table, for that site and window. Syslog stays on the operator's servers, as retention rules require.
Claims triage
An adjuster asks for the policy and prior-claims history behind one claim, from an on-prem policy system. The answer draws on the requested columns for that claim, not the policy database.
Quality investigation
A quality engineer asks which batches ran on line 4 during a temperature excursion. The data comes from plant databases and machine logs that never leave the plant network.