Seventy to eighty-five percent of AI projects in financial services fail to reach production or deliver measurable business value, according to research from Gartner, FinTellect, and the RAND Corporation. The figure is consistent across markets — AP institutions fail at the same rate as their counterparts in North America and Europe. What makes APAC different is neither the failure rate nor the root causes, which are universal, but two things specific to the region: the competitive cost of failure is higher because digital-first challengers are gaining ground faster, and the sequencing of which decisions to address first is genuinely less obvious in a market that spans regulatory frameworks, data environments, and competitive landscapes as different as Singapore, Indonesia, India, and Japan.
Both problems are solvable. But solving them requires being honest about why AI programmes fail and applying a structured approach to prioritisation that most institutions in the region have not used.
Volume and value are not the same thing — and the hidden prize in APAC is different
The instinct that drives most AI prioritisation is the same everywhere: identify decisions made at high volume, calculate the automation cost saving, and build a business case on that basis. The decisions that attract AI investment first are account administration, STP routing, low-value card authorisation, and other high-frequency operationally repetitive processes. These are not wrong choices. They are simply not where the most material return sits.
The decisions with the largest value at stake combine high volume with severe financial consequence per error. Payment fraud, AML screening, and large credit decisions all sit in this category across every market. In APAC, however, there is a dimension that the North American and European frameworks do not capture with the same clarity: the bottom-right quadrant — decisions with lower volume but catastrophic consequence per error — is disproportionately large. Large loan origination decisions in corporate and SME banking, trade finance approvals, and correspondent banking screening all carry consequences per error that can exceed the entire annual value of the simpler automated decisions in a single event. In APAC’s trade-intensive markets, with their correspondent banking complexity and sanctions exposure, this quadrant funds AI programmes on its own.
The regulatory pressure in the region is now functioning as a sequencing accelerant. MAS, RBI, HKMA, OJK, BNM, and BOT have all issued AI governance frameworks requiring explainable, auditable decisioning. That requirement lands specifically on the high-consequence decisions — credit scoring, fraud detection, AML screening — not on the high-volume low-consequence processes. Institutions that read the regulatory signal correctly understand that the compliance requirement and the commercial opportunity are pointing at the same decisions.
The three reasons AI programmes fail — and none of them are technology problems
For institutions that correctly identify the high-value decisions and commit to targeting them, the failure patterns are consistent across every AP market assessed. They break into three categories in roughly the same proportions every time, drawn from analysis of AI initiatives across AP banks between 2019 and 2024.
Governance failure accounts for around 42% of failed initiatives. The pattern is recognisable: the initiative was framed as “implementing AI” rather than “improving a specific decision.” Without a named decision, a quantified economic impact, and a sponsor with budget authority, proofs of concept produce data science results rather than business change. Technology gets selected before the problem is defined. Vendor demonstrations are approved before a sponsor is identified. The proof of concept proves the model works. Nobody was ever assigned to act on it.
Sequencing failure accounts for around 31%. The organisation had insufficient labelled outcome data, fragmented systems, or feature sets that could not be constructed in time. Rather than surface this early, teams proceeded with inadequate data and produced models too weak to deploy. In the AP context, this failure has a specific texture: data fragmentation is often more severe than in North American or European institutions because of the combination of legacy core banking systems, regulatory data localisation requirements across multiple jurisdictions, and the relative immaturity of data governance infrastructure at many institutions. The real data posture is discovered after commitment has been made, not before.
Measurement failure accounts for the remaining 27%. The model achieved its technical targets — strong AUC, meaningful false positive improvement — but nobody had defined what a good business outcome looked like in dollar terms or assigned accountability for achieving it. Model performance metrics become the proxy for business value. No economic baseline is established at the outset. When the programme is reviewed, there is no credible answer to what it delivered, and the production funding case cannot be made.
| Failure mode | Share of failed initiatives | Pattern |
|---|---|---|
| Governance failure | ~42% | No named decision. No sponsor with budget authority. PoC success defined as model accuracy not business outcome. Technology selected before problem defined. |
| Sequencing failure | ~31% | Data readiness assumed not assessed. Fragmentation — common in AP multi-jurisdiction institutions — discovered after commitment made. Models too weak to deploy. |
| Measurement failure | ~27% | No economic baseline established. Model performance metrics used as proxy for business value. Production funding case cannot be made. |
Source: Asia Pacific Banking Practice analysis of AI initiatives 2019–2024, consistent with findings across North America, EMEA, and ANZ markets.
None of these failure modes are technology problems. A more accurate model does not fix a programme with no named sponsor. A better architecture does not compensate for data that was not ready. And a higher AUC score is not a business outcome. The implication is direct: investment in model sophistication generates diminishing returns if the governance, sequencing, and measurement conditions are not in place. In APAC, where the competitive pressure from digital-first challengers makes the cost of delayed or failed implementation higher than in other markets, the case for getting the programme design right before selecting technology is stronger, not weaker.
The sequencing question in APAC requires a structured answer
The decision portfolio analysis tells institutions where the value sits. It does not tell them which decision to start with. In a single-jurisdiction market with relatively homogeneous data infrastructure, the highest-value decision and the easiest-to-implement decision often align. In APAC, they frequently do not. A decision with the highest financial consequence per error — large corporate loan origination — may have data readiness that makes a successful AI deployment eighteen months away. A decision with lower consequence but strong data readiness and acute regulatory urgency — real-time payment fraud — may be deliverable in four months. Starting with the highest-value decision and failing because data was not ready is worse than starting with the most deliverable decision, building production capability, and using the result to fund the harder programme.
The four criteria that determine where to start are financial consequence per error, decision volume, data readiness, and regulatory urgency. Scoring each candidate decision against all four and applying a priority ranking — rather than selecting on any single dimension — consistently produces a better starting point than intuition. Applied to the common AP decision set, the analysis typically produces the same prioritisation across markets: real-time payment fraud and AML transaction screening score highest for immediate action, combining high volume, high consequence, acceptable data readiness in most institutions, and acute regulatory pressure. Retail loan origination sits just behind, dependent primarily on data readiness. Trade finance screening and customer onboarding KYC are plan-next decisions where regulatory urgency is building but data readiness is typically insufficient for near-term deployment.
The value of this exercise is not the output — experienced practitioners can often intuit the ranking. It is the process: a structured cross-functional conversation that surfaces data readiness gaps before commitment is made, identifies the sponsor who will own production outcomes, and establishes the economic baseline against which the programme will be measured. These are precisely the conditions whose absence drives 85% of AP AI programmes to fail. The prioritisation framework is not a planning tool. It is a failure-prevention tool.
Part 2 of 3.
Sources
Gartner. AI Project Failure Rates. Multiple editions, 2023–2025. FinTellect AI. Why 80% of AI Projects in Finance Fail. 2024. RAND Corporation. AI Project Success and Failure Rates. Referenced in industry literature. Asia Pacific Banking Practice. Decision Portfolio and Failure Mode Analysis. Internal analysis of AI initiatives 2019–2024, consistent across North America, EMEA, and ANZ markets.
