Regulatory compliance & delivery methodology
Compliance as architecture
What sets our agent products apart is that regulatory compliance is built into the product architecture. Every requirement maps to an engineering mechanism, and the products ship with six go-live red lines — cross any one, and it does not go live.
Three regulations, six gates
Our products provide an engineering answer to each of the six gates these regulations establish: human oversight, explainability, review of customer rights, training-data red lines, high-risk approval, and filing and accountability. In FMCG, pharmaceuticals and other industries, the same mechanisms also carry general requirements such as the Personal Information Protection Law, as well as industry rules.
- Interim Measures for the Administration of Generative AI ServicesIn force since 2023
- Measures for Labeling AI-Generated Synthetic ContentIn force since 2025
- NFRA Jin Fa [2026] No. 8: Guidelines on the Safe Development and Application of AI in Banking and Insurance32 articles
Every regulatory requirement lands on an engineering mechanism
One principle, four planes, ten components
The principle is “trade control for autonomy”. Four planes form a bank-wide shared foundation; each scenario is simply a tenant on that foundation.
- Plane 01Business process plane
- Plane 02Agent runtime plane
- Plane 03Capability & integration plane
- Plane 04Control & evidence plane
How scenarios are prioritized
We follow a priority matrix of “scalable augmentation first, reversible, well-evidenced” — an irreversible scenario never enters the priority zone, however valuable it is.
L levels govern applications. A levels govern actions.
L0–L5 and A0–A4 do different jobs and work together. The L scale asks how much autonomy an agent application may have, for tiered admission of applications and scenarios. The A scale asks how far a single action may be automated, as a bank-wide measure for every action. New scenarios define the domain, then the level, then start work.
| Level | Meaning | Typical actions |
|---|---|---|
| A0 Retrieve / read-only | Read-only retrieval, citing sources and summaries; no source, no answer | Policy Q&A, clause lookup |
| A1 Advise | Generates advice, scores and drafts; writes to no business system | First-draft credit opinions, red-flag alerts |
| A2 Pre-fill | Writes drafts and to-dos that stay “unsubmitted” and can be withdrawn in full | Application pre-fill, draft reports |
| A3 Execute after approval | After an authorized signature, calls allow-listed deterministic commands | Submitting for approval, sending notices |
| A4 No autonomy | May never be triggered or decided by an agent; supporting material only | Funds transfers, regulatory filings, account freezes |
Each phase ends in an accepted deliverable — the ticket to the next
- Days 1–15
Decide and map
Review how the core principles fit your bank; map the first scenario’s real process to roles and systems, and mark every step that needs a human signature.
DeliverableScenario selection report + human checkpoint list - Days 16–45
Build the hub, fix the contract
Stand up the minimum viable Harness: admission, planning, state machine, tool gateway, policy gate, validator; lock in the task contract and mandatory stop conditions.
DeliverableTask contract + control-module list + test set - Days 46–90
Run end to end, pass the gates
Run the full chain from read-only assistance to draft write-back, pass dual go-live threshold tests and emergency-stop drills, and produce a complete audit trail.
DeliverableValidation report + admission decision + go-live plan
Let AI take on roughly 80% of cognitive and operational work — while fact-finding, exception judgment, major risk decisions, customer-rights decisions and accountability stay firmly inside a controlled governance system.
- The modelthinks
- The agentdoes
- Rulesdecide whether it may be done
- Harnessmakes sure it is done safely
- Peoplehold final accountability
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.