ICBC
Rebuilding the small-business credit portfolio
+RMB 610Bannual lending uplift
- 60+ small-business products consolidated into 7 — 80% less complexity
- Approval rate 15% → 32%; drawdown rate 50% → 65%
- Average M1 delinquency down 168 bps
Industry solutions · Banking
AI for banking — from acquisition and underwriting to drawdown and collections
Banking is where we started. We were early to bring large models and AI agents into end-to-end banking workflows, built around SAM scorecard modeling, VDC variable derivation, DH S1 real-time decisions, the Enterprise Agent Platform and the Harness governance engine — across retail, small-business and corporate credit.
Retail · small-business · corporate creditAcquisition · underwriting · drawdown · post-loanEvery product deploys on-premises
2 trillion+RMB lent by the institutions we serve
50 billion+RMB net profit earned by those institutions
50+leading financial institutions served worldwide
100+solutions delivered into production
Trusted by state-owned, regional, private and rural commercial banks, financial holding groups and banks in Europe and the USSelected clients
ICBCChina
China Construction BankChina
Bank of NingboChina
Bank of SanxiangChina
Pizhou Rural Commercial BankChina
China UnionPayChina
TCL Financial HoldingsChina
INGNetherlands
Wells FargoUnited States
JPMorganUnited States
CitibankUnited States
BarclaysUnited Kingdom
SAM in banking
Every credit model and strategy, built on one pipeline
We have designed 30+ credit products in China, and that experience lives on as SAM’s end-to-end product templates: every model runs the same fourteen-step pipeline, with statistics computed by deterministic code and sign-off by a named approver.
Acquisition
Marketing whitelists and response models
Find high-intent, low-risk segments in properly consented data and direct marketing spend.
Underwriting
Application scorecards and anti-fraud
Master-scale models, admission rules and fraud-ring detection, combined with limit and pricing decisions.
Drawdown
Behavior scorecards and limit strategy
Usage models and strategies, drawdown-rate, renewal and revolving models that adjust limits over time.
Post-loan
Early warning and collection scorecards
Post-loan warning, collection strategy and asset preservation, with resources allocated by risk tier.
Enterprise Agent Platform · Harness governance engine
The model proposes. Harness decides.
A large model is a brain. An agent is brain + hands + workflow. Enterprise agents on the platform are built from three layers and twelve components — and an “agent” without its governance layer never touches a production system.
01
Cognition layer · thinks
LLM core · task goals & contract · planner · memory & state
02
Knowledge & action layer · acts
RAG retrieval · Skill packs · tool calls via MCP / API · workflow engine
03
Governance layer · controls
Harness runtime control · rules engine · permissions & authorization · audit trail
The five inequalities
Capability≠authority
Reasoning≠fact
Advice≠decision
Decision≠execution
Execution≠unlimited execution
Scenario agents · Corporate credit
AI does 80% of the preparation. People keep 100% of the decision.
A far-from-clean RMB 30 million credit application is classified as high-risk and locked at autonomy level L2 — the agent’s toolset contains no approval or disbursement tools at all, so overreach is structurally impossible. The case advances on conditions, not intent, through nine stages before an authorized approver decides.
- 01RMB 30MCredit facility requestedIntake immediately surfaces four initial anomalies, including high customer concentration
- 02RMB 21MMinimum of four constraintsProduct cap, remaining group limit, collateral haircut cap, working-capital estimate
- 03RMB 17.5MCandidate limitAdjusted by a 0.85 risk factor, rounded down to the nearest RMB 0.5M
- 0492%Evidence coverage below 100%Independent validation refuses submission and requests documents; an authorized approver signs off after
Case figures are illustrative and do not represent any real customer.
Results in banking
Measurable business outcomes
Institutions we serve in China have lent more than RMB 2 trillion and earned more than RMB 50 billion in net profit, with no deviation in the default-probability predictions of the products we support. Three representative before-and-after results:
Bank of Ningbo
A self-operated consumer loan, from zero to one
RMB 2B → 100Bannual lending
- Average rate down 800 bps; average ticket up RMB 43k
- Approval rate 10% → 21%; drawdown rate 20% → 35%
- Average M1 delinquency down 570 bps
Bank of Sanxiang
Deepening construction-machinery lending
RMB 0.4B+ → 3.2B+annual lending
- Approval time down 90%; due-diligence time down 60%
- Approval rate 45% → 56%; drawdown rate 35% → 45%
- Average M1 delinquency down 600 bps
Governance & compliance
Compliance as architecture
For China’s Interim Measures for Generative AI Services, the Measures for Labeling AI-Generated Synthetic Content and the NFRA’s Jin Fa [2026] No. 8 Guidelines on the Safe Development and Application of AI in Banking and Insurance, every regulatory requirement maps to an engineering mechanism.
01
Human oversightTiered approval gates and a suspendable state machine
02
Explainability & traceabilityEnd-to-end tracing and a unified case file where every conclusion points back to its source
03
High-risk admissionDual go-live thresholds, shadow runs and emergency-stop drills
04
Data red linesOn-premises deployment, field-level permissions and training isolation
05
SecurityGuardrails, sandboxes and red-team testing
06
Lifecycle accountabilityApplication registry, version-bound approvals and tamper-proof audit logs
Built for banks
End-to-end solution, scenario agents and customer results
P&E-HOPE
P&E-Hope Retail Credit Solution
Products, marketing, risk, operations, agents, models, algorithms and a financial-LLM foundation in one stack, governed end to end by Harness.
Learn more
SCENARIO
Scenario Agents
The corporate credit agent plus ten applications: corporate due diligence, AML, trade documents, consumer protection, reporting, collections and contact centers.
Learn more
RESULTS
Customers & Results
Measured before-and-after results and talent development at ICBC, Bank of Ningbo, Bank of Sanxiang and more.
Learn more
Get started
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.