Do-It Agent
Describe the outcome · the agent plans and executes
The agent spans all three prototypes. It plans the steps, runs the models, stages the action, and leaves a human in the loop for anything that writes to the real world. Never approves a decision alone, never breaches a configured guardrail.
Branch · Ruwi
Who should Ruwi branch call tomorrow? Rank by expected value and cap at 15 contacts.
Who should Ruwi branch call tomorrow? Rank by expected value and cap at 15 contacts.
Margin
Give me 50 basis points of margin without breaching the 5% attrition appetite.
Give me 50 basis points of margin without breaching the 5% attrition appetite.
Committee paper
Prepare the ALCO paper for Thursday on the rollover-repricing play.
Prepare the ALCO paper for Thursday on the rollover-repricing play.
Risk
Give me the early-warning watchlist for next quarter — rank by stage migration probability.
Give me the early-warning watchlist for next quarter — rank by stage migration probability.
Trigger · design
Build a trigger for customers whose kids are starting school.
Build a trigger for customers whose kids are starting school.
End-to-end
Walk Salim Al Hinai from trigger → credit decision → funding cost.
Walk Salim Al Hinai from trigger → credit decision → funding cost.
Operating envelope
Hard-coded guardrails
- No real-world writes. Agent stages actions; a human maker-checker must approve.
- Will not breach configured attrition appetite, DBR ceiling, LTV floor, LCR appetite.
- Will not assert regulatory compliance or issue a Sharia opinion.
- Will not fabricate a figure the dataset cannot support — says so instead.
- Every tool call carries a trace and lands in the Audit Trail.
- Role of the active user governs tool access & field unmask.
Tools wired
| DP | Deposit Pricing · Elasticity · pricing · what-if |
| TL | Trigger Lending · trigger engine · opportunity query |
| CD | Credit Decisioning · affordability · policy · ECL |
| All | Shared forecast engine |
| All | Audit log writer |
| Draft | Committee pack generator |
| Human | Maker-checker queue |
Trust & override
Every step is reviewable & interruptible. If a plan proposes a step that lies outside policy, the step is blocked with the policy named, and the agent asks whether to request an exception.
Current session role: Credit officer. The agent can read all queues and stage actions; it cannot deploy policy changes or unmask bulk PII without an elevated role.