2026 Greater China financial services AI deployment survey: how many POCs actually make it to production? Three reasons projects stall

Last quarter, we reviewed the AI projects we've worked on and interviewed across Greater China's financial institutions. We wanted to answer a question most organizations avoid: of all those AI POCs with signed contracts, launch announcements, and demonstrations, how many actually made it to production and are being actively used?

The results were revealing.

Financial services AI deployment rate: only 18% of POCs actually make it to production

In the first half of 2026, we reviewed 116 financial AI projects launched between 2024 and 2025 across Greater China, primarily Chinese-language financial institutions in Taiwan, Hong Kong, and Singapore, spanning banks, insurance, securities, and asset management. Our tracking criteria are strict. Success requires three simultaneous conditions: deployment to production, real business users (not just the project team) using it daily, and operations continuing beyond 90 days.

By this standard, Greater China's financial services AI deployment rate is 18%.

Of every five projects launched, fewer than one reaches production with active users. Over 80% of financial AI projects in our sample stall between the demonstration and pilot phases.

The funnel shows losses at each stage:

The sharpest drop occurs between stage two and stage three. Sixty-one percent pass technical validation, but only about half deploy. Technology itself is not the bottleneck. What comes after, the regulatory, organizational, and operational challenges, stops most projects.

Why projects stall: the three main reasons

We interviewed teams behind 41 projects that passed validation but never shipped and identified three recurring patterns. None involve the model itself.

First, compliance and security reviews arrived too late or not at all. This accounts for 43% of stalled projects. Financial services faces this challenge uniquely. A RAG knowledge system operates smoothly in testing, then encounters resistance when compliance and security reviews begin. Key questions surface: Can customer data leave the internal network? Who bears responsibility if the system generates problematic advice? How are audit trails maintained? How do systems align with data privacy and outsourcing regulations? One insurance claims assistance project we tracked was engineered in three months but spent seven months in compliance review. The organization ultimately removed it because it could not meet interpretability requirements. The real problem wasn't speed. These requirements weren't optional details to handle later. They were entry conditions that should have shaped development from the beginning.

Second, nobody actually owned adoption. This is 31% of our cases. This is the scenario we encounter most often. The system launches, training is conducted, adoption rate stalls in single digits. The reason: the implementation team delivered a system but not a change management process. Nobody was accountable for getting tellers, underwriters, or advisors to modify their routines. Front-line staff operate under performance targets, established procedures, and the pressure to avoid errors. A tool requiring them to open another window or learn another interface gets bypassed the moment workload increases. Without someone monitoring how people actually work and integrating the AI into existing workflows, adoption never exceeds 4%.

Third, POC success criteria do not align with production conditions. This is 26% of cases. Demonstrations use clean, curated data, ideal network conditions, and single-user scenarios. Production involves decades-old core systems, fields known for data quality issues, peak-time concurrency, and transactions that cannot fail. A corporate lending workflow achieved 92% accuracy in testing but dropped to the low 60s when connected to actual credit data sources. The business team rejected it immediately. A POC demonstrates whether something might work. It does not demonstrate whether it will work in your specific production environment. This gap between proof-of-concept and production deployment represents the entire challenge of implementation.

What the 18% did right

The surviving projects share a consistent approach. From the beginning, they prioritized shipping over validation. Compliance and security joined the first meetings as partners, not final gatekeepers. Someone was explicitly accountable for adoption rates, not just for handoff at launch. Success criteria were defined using real production data and realistic concurrency levels, not demo conditions. Demonstrations served as checkpoints, not destinations.

This shaped our approach: we place engineers on site working directly with underwriters and advisors, and we remain involved until the system reaches active daily use. In financial services, demonstration quality doesn't predict success. Deployment and actual use do.