Every AI step,
under your rules.
Risk OS checks every action an AI takes against your bank's own credit rules before it runs. Risk signs off on the rules once. Every decision after carries a person's name.
How a check works.
1
The AI proposes
Any AI in Risk OS, or one you connect, wants to do something: change a grade, send a memo, propose a line change, search for borrower news.
2
Your rules check it
Before anything runs, the action is checked against your credit policy, limits, approval authority, watch list and data rules.
3
One answer comes back
Allow, hold for a signer, block, or stop all. The AI can't skip the check or argue with it.
4
The decision is recorded
Who proposed it, what it wanted, the evidence, the rule that applied and who signed, kept for audit and the examiner.
Four answers. No grey area.
- Allow
- The action is inside your rules, so it runs. Most routine work, like drafting a memo from a confirmed spread or flagging a covenant test, lands here.
- Hold
- The action needs a person. It goes to the right signer with the reason and the evidence. Nothing changes in your systems until they decide.
- Block
- The action breaks a rule that has no exception path, such as sending a borrower's confidential rating outside the bank. It never runs.
- Stop all
- One switch pauses every AI action across the bank until someone with authority turns it back on. For incidents, audits or plain caution.
What your rules can cover.
- Credit policy and limits
- House limits, hold limits, policy exceptions and product rules, applied to every action that touches a loan or an account.
- Approval authority
- Your delegated authority matrix: who can approve what, by exposure and grade, from relationship manager to board committee.
- Grade movement
- How far a grade can move and on whose say. An AI proposal is never the assigned grade until a reviewer signs it.
- Watch list and risk signals
- Borrowers and cohorts that need a second pair of eyes get one, automatically.
- Evidence requirements
- Actions must carry their evidence: linked financials for a grade change, a current spread for a memo, complete coverage for a monitoring plan.
- Data handling
- What borrower and customer data an AI may read, and where it may never go.
What it looks like in practice.
| Book | The AI wants to | Answer |
|---|---|---|
| Commercial | Memo AI drafts a renewal memo from a confirmed spread | Allow |
| Commercial | Memo AI proposes moving a grade from BB+ to BB−, two steps | Hold · senior credit officer |
| Commercial | A $7.5M increase above the $5M house limit | Hold · loan committee |
| Commercial | Research AI puts a borrower's internal rating into a web search | Block |
| Retail | Line increase for an eligible card cohort within policy caps | Hold · credit strategy approval |
| Retail | Monitoring plan submitted without complete data coverage | Block until coverage is complete |
| Any | Unexpected behavior from a vendor's AI model | Stop all |
Illustrative examples. Your rules decide the answers.
A record of every decision.
The record is written at the moment of each decision, by the same layer that made the check. Nobody has to piece it together for the exam.
- Which AI proposed the action, and at what level of authority
- What it wanted to do, on which loan, account or cohort
- The evidence it used
- The rule that applied, and its version
- The answer, and who approved or declined it
- When each step happened
Works with the AI you already have.
Keep your loan platform, your core and the AI tools you've bought. Risk OS sits between the AI and your systems, so the same rules and the same record cover Aarvion's credit AI, your vendor's assistant and your own models.
See the controls at work in Commercial Credit and Retail Credit.
Questions risk teams ask.
Which AI does Aarvion control?
Any AI working through Risk OS: Aarvion's own commercial and retail credit AI, and AI you connect, such as your loan platform's assistant, a vendor model or an in-house model.
Who writes the rules?
Your credit and risk teams. In a pilot we turn your existing credit policy, limits and approval authority into rules together, and risk signs off on them before anything goes live.
Does the AI ever make the final credit decision?
No. The AI prepares work and proposes actions. Grades are assigned by reviewers, approvals follow your authority matrix, and anything outside your rules waits for a person.
What does the stop-all switch do?
It pauses every AI action in the bank at once. Work already approved by people stays approved. AI actions resume only when someone with authority turns the switch back off.
What can an examiner or auditor see?
For every AI action and every approval: who proposed it, what it wanted to do, the evidence, the rule and version that applied, who signed, and when.
