AI begins to reshape SME credit assessment at Chinese lenders

Written by Nanda Lakhwani

Mar 18, 2026
8 min

Discussions at the Shanghai International AI Finance Summit 2026 highlighted how Chinese banks are using AI to address structural constraints in SME lending — Jiangsu Su Merchants Bank by industrialising its internal credit workflow, and SPD Bank by restructuring its institutional model to reduce borrower uncertainty before lending begins. Early portfolio indicators look stable, although the risk-adjusted case is still being written.

Structural constraints in small and medium-sized enterprise (SME) lending remain entrenched across banking systems. Thin credit histories, limited collateral and the weak unit economics of small-ticket exposures have long constrained banks’ ability to scale these portfolios profitably. Chinese lenders are beginning to explore how AI deployment is changing the way these frictions are managed, as discussed at the Shanghai International AI Finance Summit 2026.

In China, the strategic context is sharpening those pressures. Industrial policy increasingly emphasises the integration of technology-sector SMEs into national supply chains, with state initiatives identifying more than 17,000 designated “little giants” and a broader tier of over 100,000 specialised enterprises as priorities for industrial upgrading. For lenders, this reinforces pressure to extend credit to innovation-driven firms whose growth prospects may be clearer than their financial track records.

The lenders’ operational data suggested that AI deployment is reshaping the mechanics of SME lending rather than its fundamental risk profile. Technology-enabled processes are accelerating decision cycles and widening client reach. Evidence on risk-adjusted returns remains limited, though the banks pointed to stable NPL ratios and improving re-performance rates in support of their assessments.

Industrialising credit assessment

Jiangsu Su Merchants Bank is a Nanjing-based mid-tier institution founded in 2017 with a core mandate to expand lending under China’s inclusive SME finance framework. Ouyang Tao, general manager of the bank’s fintech department, said first-time technology borrowers historically required up to three times longer to process than repeat clients, with disbursement cycles extending to around one month. He attributed this to non-standard application materials, difficulty verifying borrower disclosures, uneven underwriting judgement and the risk of hidden liabilities or falsified documentation.

Ouyang said the bank has rebuilt its credit process around multiple specialised AI applications, supported by technology investment equivalent to roughly 6% of annual operating revenue and a workforce in which technology staff represent more than 51% of headcount. Designed from inception as an online-to-offline institution, AI infrastructure forms a core operating capability rather than a retrofit. Ouyang outlined an architecture built on four pillars: data asset integration, domestically sourced high-performance computing, in-house model development combining open-source and proprietary approaches, and deep embedding of AI capability into business workflows.

The performance figures Ouyang cited were detailed. A voice-based outreach system generated RMB 110 million ($15.2 million) in new lending and RMB 160 million ($22.1 million) in deposits within a year, while reducing workload equivalent to 7.5 full-time employees. A multimodal document-review agent reduced application rejection rates from around 50% to 10% and improved per-document processing efficiency by 275%; the system also identified more than ten suspected cases of document tampering in its first year.

A transaction-flow analysis system now processes more than 80 bank-statement formats and generates more than 200 risk labels per borrower profile, reducing average processing time from roughly one hour to under two minutes. Ouyang said a credit-review agent trained on historical case data and senior underwriter judgement patterns has shortened approval cycles from ten days to fewer than three, with more than 90% of system-generated risk prompts accepted by human reviewers. Ouyang said AI outputs are intended to inform rather than replace human judgement.

AI tools have also been applied post-disbursement. Ouyang said an automated investigation platform for non-performing loan cases has generated close to 2,000 review reports, each reducing manual analysis time by around ten working days. A contracts-analysis system has processed more than 60,000 loan agreements, saving the equivalent of more than 1,500 working days annually.

On portfolio performance, Ouyang said the bank serves more than 4,000 technology enterprises within an inclusive-finance book exceeding RMB 20 billion ($2.8 billion), and that the proportion of re-performing loans in this portfolio has improved by four percentage points since AI deployment. In 2024 the bank established a joint financial large model engineering research centre with the Nanjing municipal government, which has since become a regional platform for AI innovation in banking.

On AI costs, Ouyang’s answer was precise. Jiangsu Su Merchants Bank operates a deliberate low-cost strategy as a governance principle: prioritising domestic chips over imported hardware, maintaining elastic compute supply to avoid resource waste during low-utilisation periods and sourcing AI talent through university partnerships rather than the open market. The position, he said, is that infrastructure self-reliance is a commercial prerequisite — not just a preference — because dependency on imported hardware or external talent creates institutional risk for a regulated lender. For a mid-tier institution, he concluded, all-in AI commitment is not a strategic option but a commercial necessity.

Expanding the risk perimeter

Shanghai Pudong Development Bank (SPD Bank), one of China’s 12 national joint-stock commercial banks and a major lender to technology enterprises, has approached the same problem from a different direction — deploying AI not primarily to accelerate credit assessment, but to change what is being assessed and who is being reached.

Zhong Quan, deputy general manager of the bank’s technology-finance department, argued that the financing constraints facing many technology SMEs are not principally a function of creditworthiness. The most capable early-stage firms, he said, often attract multiple lenders; those that cannot access financing typically lack commercial track records rather than genuine technical merit. Underwriting models centred solely on financial statements may therefore provide an incomplete basis for lending decisions, he argued.

According to Zhong, the bank’s AI-driven prospecting platform, known internally as Moxiangying (Sperm Whale), aggregates public-domain and commercial data sources to identify and engage potential technology-sector clients. He said the system onboarded more than 9,000 new technology enterprises in its first year — a volume he said would previously have required three years to build through conventional relationship-management channels.

Zhong said the bank has introduced a five-capability credit-assessment framework that incorporates non-financial indicators including track record, team competitiveness, team composition and innovation capability — enabling the bank to assess a technology enterprise’s sector position and market trajectory, not just its current financials. Products are structured to follow firms from seed stage to maturity, with digital onboarding and proactive engagement calibrated to growth phase. An internal science and technology finance knowledge base transmits the bank’s analytical approach consistently across its sales network, he said, ensuring uniform assessment standards across branches.

Zhong argued that SPD Bank’s most significant governance decision is structural. The bank has established Waitan FTC — the Bund Financial Technology Community — as a legally incorporated, non-financial platform company: the first of its kind among Chinese banks. Registered with RMB 50 million ($6.9 million) in capital, FTC co-locates more than 10,000 investors, representatives of more than 100 top-tier universities and a broad range of professional services firms within a single ecosystem. In less than six months it hosted more than 1,100 events and served tens of thousands of technology enterprises.

Zhong explained that the significance of FTC lies in its structure. By establishing a non-financial legal entity outside the bank’s licensed perimeter — open to all banks, not proprietary to SPD Bank — the bank has separated the ecosystem function from the lending function. His framing is that SPD Bank is both a user and a provider of financial technology infrastructure, and that the next phase of AI finance will require shared platforms of this kind rather than proprietary development within individual institutions. His stated objective is to reduce the commercial uncertainty surrounding early-stage borrowers by assembling the resources that make their outcomes more predictable before a formal credit decision is reached.

On portfolio performance, Zhong said more than 50% of the bank’s new credit authorisations now relate to technology enterprises, approximately 90% of which are SMEs. He said overall non-performing loan (NPL) ratios have not diverged materially from pre-AI levels despite this expansion, and that technology-finance activities contribute more than 30% of total bank revenue.

Zhong acknowledged that the stability in NPL ratios partly reflects the rapid expansion of the denominator — the portfolio has grown fast enough that the ratio holds even as the absolute book scales. Whether the underlying credit quality of newly originated cohorts will perform in line with the existing book, he said, remains to be seen as the loan cycle matures.

Implications for SME credit models

The presentations addressed AI adoption from different entry points: Jiangsu Su Merchants Bank focused on automating internal credit workflows, while SPD Bank concentrated on widening the information base and institutional relationships available before a credit decision is reached. The session pointed to three governance observations.

First, cost architecture is a governance choice: Jiangsu Su Merchants Bank treats infrastructure self-reliance — domestic hardware, elastic compute, university talent — as a commercial prerequisite for sustained AI deployment. Second, institutional structure shapes what AI can do: SPD Bank's incorporation of FTC as a non-financial, open-access entity separates ecosystem development from licensed banking activity. Third, both banks preserve human decision authority over AI-generated risk assessments.

Both banks operate within an enabling environment specific to China — dense data infrastructure, domestically sourced compute and active policy direction toward technology enterprise development. Whether AI-originated SME lending produces structurally better credit outcomes over a full cycle remains to be demonstrated.

Keywords: