In most financial institutions, back-office staff spend a large share of their time on work that has nothing to do with their expertise: finding out where a case is stuck, filling in a form whose information is already recorded somewhere else, or building a report that is needed only once and that no one will spend development time on. These are exactly the places where a language-model assistant can free up time.
At the same time, a financial institution is no place to let a language model produce figures or make decisions on its own. Dara's AI assistant was designed with both of these considerations in mind.
An assistant on the same platform
The Dara assistant is a thin layer over the same modules, operations and permissions the user works with in the portal every day, and it has no database of its own. When a staff member asks the assistant for something, the assistant calls the same operation the staff member could have called with a few clicks, and with the same permissions.
Where it helps
Tracking a case
One of the most common questions in operations teams is where a particular case has stalled. The assistant checks the Process Engine for the step the case is at, which department's work queue the task is sitting in and since when, and when that step's deadline runs out. The answer comes with its source, for example “Credit department work queue, committee approval step.”
Filling in forms and starting processes
A staff member types “File a Qard al-Hasan loan application under the marriage scheme for the customer with this national ID.” The assistant fills in the form with the customer's existing information, highlights the empty fields and shows the draft for review. The staff member makes the final submission with their own click.
Detecting drift and anomalous behavior
The assistant can compare the current configuration with the expected one, flagging, for example, a deposit category whose daily withdrawal limit is out of line with similar categories, or a user whose activity pattern has suddenly changed. Each finding is reported with its evidence, and the decision about it rests with a person.
Reviewing permissions before an audit
Ahead of a periodic audit, the assistant compares users' permissions with their roles and lists anything that falls outside the pattern, such as a user who can post vouchers but does not work in the finance department.
New reports without code
A manager who needs a report that is not among the ready-made ones describes it in plain Persian. The assistant builds the query only over data the manager is allowed to see, and shows the result together with the filter conditions it applied.
Where it must not decide
The assistant's boundaries are built into the architecture itself and do not depend on how the model is configured:
- No new level of access is created. The assistant sees what the user would see after signing in to the portal, and not a single row more.
- No change is recorded without human approval. Posting a voucher, deactivating an account, publishing a process and sending correspondence all wait for a person's click.
- Figures are calculated by the system. Totals, trial balances, account balances and days overdue are all computed by the system; no figure comes from text the model has generated.
- Credit and identity decisions are not automated. Final identity verification and any credit or pricing decision are never handed over to the assistant, under any circumstances.
- Every answer has a source and every action is logged. An answer that refers to data states which voucher it came from and what condition was applied.
- The chain of steps is bounded. If a request runs past the set number of steps, execution stops so that a manager can decide.
Why figures must not come from the model
Language models are good at writing fluent text, and that very fluency can be misleading. A figure stated confidently in a well-formed sentence sounds convincing even when it is wrong. In Dara, the assistant takes figures from system operations, and its only job is to place them in a sentence. If a user asks for a branch's total overdue installments, the figure comes from a system report, and the assistant says which report it read and with which filter.
Where the model runs and where the data goes
The assistant works on the modules the organization has deployed. Where the language model runs, and whether any data leaves the organization's network, are decided separately for each project, in line with that organization's security policy.
- #AI
- #AI_assistant
- #Audit
- #Permissions
