Scenario Agents
AI does 80% of the preparation. People keep 100% of the decision.
Scenario agents put agent capability into specific banking roles. Every scenario defines exactly what the agent takes on and which human checkpoints are irreplaceable — and all of them reuse one platform foundation.
How a far-from-clean credit application is handled safely
A smart-equipment manufacturer applies for a RMB 30 million credit facility. Intake surfaces four initial anomalies: high customer concentration, receivables growing far faster than revenue, suspected related-party outflows and inconsistent revenue definitions. The case is classified as high-risk and locked at autonomy level L2. The agent may query authorized data, call financial calculations and risk models, run credit rules and draft options — but its toolset contains no approval or disbursement tools of any kind. Overreach is structurally impossible.
- 1Intake & admission
Open the case, check authorization, classify as high-risk L2, pin policy and model versions, generate the execution plan.
Output: case opened · permissions narrowed - 2Document parsing & quality check
Finds a missing cash-flow statement, contracts with signature pages only, an expired collateral appraisal and a missing bill ledger.
Output: NEEDS_SUPPLEMENT - 3Business authenticity checks
Three cross-checks: invoice coverage 93.57%, collection coverage 81.24%, receivables growing far faster than revenue.
Output: RMB 34.4M gap to explain - 4Risk identification & site checks
Generates four tailored question sets — concentration, receivables, related parties, use of funds — instead of a fixed questionnaire.
Output: 20 site-visit questions - 5Rules & limit calculation
No hard-rule vetoes, but four prerequisites and two re-checks; the rules engine calculates a candidate limit.
Output: candidate limit RMB 17.5M - 6Candidate proposal
Lists supporting evidence alongside six objections, so the report never just argues for approval.
Output: proposal + counter-evidence list - 7Independent validation
Recomputes numbers, re-runs rules, verifies evidence — key-claim evidence coverage is only 92%, short of the 100% submission threshold.
Output: submission refused · documents requested - 8Human investigation & approval
With documents in, an approval pack is formed and the report keeps its V1–V4 version chain; an authorized approver signs off on RMB 17.5M with eight conditions.
Output: formal credit decision - 9Controlled write & reconciliation
The credit system performs the write; the agent only reads back and reconciles item by item, checking status first on timeouts.
Output: COMPLETED
The limit is not estimated — it is squeezed by four constraints
- 01Four constraintsProduct cap · remaining group limit · collateral haircut cap · working-capital estimate
- 02RMB 21MTake the minimum of the four
- 03× 0.85Adjust by the risk factor
- 04RMB 17.5MRound down to the nearest RMB 0.5M for the candidate limit
If the model suggested approving RMB 20 million, the policy gate would block it outright and log the deviation as a model-behavior event. Once documents arrive, the authorized approver signs off on RMB 17.5 million with eight conditions, each translated into a monitoring rule that can fire in the post-loan system.
Case figures are illustrative and do not represent any real customer.
Five business lines, ten scenarios, one platform
On a single methodology and platform foundation, our scenario agents now span five business lines: credit, risk & compliance, operations, customer service and fintech.
| Scenario | What the agent takes on | Irreplaceable human checkpoints |
|---|---|---|
| B1 Corporate working-capital loan due diligence | Document checks, three-statement reconciliation, related-party tracing, draft review report (disagreements auto-flagged) | On-site investigation and interviews; final credit-committee sign-off |
| B2 AML suspicious transactions | Aggregated triage packs, routing suggestions, structured investigation records | Case-by-case human review; the decision to file or not |
| B3 Letter-of-credit document review | Cross-document checks, special-clause parsing, discrepancy pre-review with cited rules | Compliance determination and refusal decisions |
| B4 Consumer-protection complaints | Classification and grading, case deadline management, draft responses | Resolution decisions and sign-off on response wording |
| B5 Suitability & sales-recording QA | Rating match and frequency checks, full transcription QA of recorded sales | Sales conducted in person; rulings on violations |
| B6 Regulatory reporting data quality | Definition mapping, cross-table error location, data-lineage tracing | Error determination and report sign-off |
| B7 Centralized operations post-review | Voucher image recognition and pre-fill, element checks, error tickets | Review and authorization roles; error accountability |
| B8 Collections call compliance QA | Pre-checks on frequency, timing and contact targets; real-time script-violation detection | Approval of collection strategies and relief plans |
| B9 Engineering productivity & release compliance | Coding assistant, security and open-source compliance scans, change-evidence collection | Dual review of release approval; triggering release commands |
| B10 Remote banking agent assist | Intent recognition, permission-aware knowledge with sources, ticket pre-fill | Customer confirmation for transactions; commitments on funds and rights |
Build the platform once, then onboard ten scenarios — not ten separate projects
01The first value point
Always freeing staff from cross-system searching and document gathering — the fastest payoff, least contested and lowest compliance risk.
02Human checkpoints
Consistently land in three places: formal commitments to customers, decisions affecting rights and funds, and sign-offs regulators name.
03One platform foundation
Every scenario reuses the permission-aware knowledge layer, evidence binding, gate scheduling, the release gate and monitoring.
Start with one business question. Let SAM deliver a model you can sign off.
Whether you are in banking, FMCG and marketing, pharmaceuticals or another industry, try SAM online first — or talk to our modeling experts about your use case.