We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    [Case Study] SGD 1 Billion in AI Value: How DBS Industrialized 2,000 Models

    What a bank-wide portfolio of more than 430 use cases teaches leaders about reusable infrastructure, value accounting, and model governance.

    Conscious Engines

    Industry: Banking and financial services
    Organization: DBS
    Use case: Bank-wide AI portfolio across customers, risk, operations, and employees
    Evidence basis: DBS 2025 annual reporting
    Disclosure: This is an independent analysis by Conscious Engines based on DBS public disclosures. The economic-value figure uses the bank's internal accounting methodology, which is not published model by model.

    1. Outcome at a Glance

    DBS reported operating more than 2,000 AI models across more than 430 use cases and generating approximately SGD 1 billion in economic value during 2025.

    Key Outcomes

    More than 2,000

    AI models

    Bank-wide portfolio reported in the cited case.

    More than 430

    Use cases

    Applied business scope reported in the cited case.

    Approximately SGD 1 billion

    Economic value in 2025

    Internal value accounting.

    MeasureReported resultEvidence note
    AI modelsMore than 2,000Bank-wide portfolio
    Use casesMore than 430Applied business scope
    Economic value in 2025Approximately SGD 1 billionInternal value accounting

    The ratio of models to use cases, roughly 4.7 models per use case, is strategically revealing. A production use case often needs multiple components: propensity, risk, classification, document extraction, language generation, fraud signals, and monitoring. It is rarely one model connected to one screen.

    The SGD 1 billion figure is significant because it is reported in an annual report, but it should still be treated as management's estimate. The public disclosure does not expose every baseline, counterfactual, cost allocation, or realization rule.

    2. The Operational Problem

    Banks have thousands of repeatable decisions and information tasks. They must authenticate customers, detect fraud, assess risk, monitor transactions, serve customers, process documents, support employees, and meet regulatory requirements.

    The challenge is not finding possible AI use cases. It is industrializing them without multiplying inconsistent data, duplicated integrations, unmanaged models, and hidden risk.

    A pilot can be built around one dataset and a small team. A 2,000-model estate requires model inventory, ownership, lineage, validation, monitoring, access controls, incident management, and retirement. It also requires value measurement that can compare very different workflows.

    Generative AI adds new risks, including unsupported text, prompt injection, data leakage, and variable inference cost. Banks need to combine probabilistic models with deterministic controls and human authority.

    3. What Was Built

    DBS's public description points to an enterprise AI factory rather than a single application.

    System at a Glance

    Governed data

    Reusable customer, transaction, product, and risk features.

    Model platform

    Train, deploy, version, and monitor models consistently.

    Decision services

    Expose predictions and recommendations to workflows.

    Language layer

    Search, summarize, extract, and assist with controlled knowledge.

    Governance

    Inventory, validation, approvals, explainability, and audit.

    Value office

    Attribute operating, revenue, risk, and experience benefits.

    LayerBank-wide role
    Governed dataReusable customer, transaction, product, and risk features
    Model platformTrain, deploy, version, and monitor models consistently
    Decision servicesExpose predictions and recommendations to workflows
    Language layerSearch, summarize, extract, and assist with controlled knowledge
    GovernanceInventory, validation, approvals, explainability, and audit
    Value officeAttribute operating, revenue, risk, and experience benefits

    The model portfolio likely contains classical machine learning alongside language models. That is the correct architecture. Fraud scoring, propensity, forecasting, and anomaly detection should not be replaced merely because generative models are fashionable.

    Task-specific language models can add value in document-heavy and service workflows. Enterprise RAG can ground policy answers. Speech models can support regulated contact centers. Model routing can reserve expensive models for ambiguous tasks while smaller models handle extraction and classification.

    4. How It Reached Production

    Industrialization requires repeatable gates.

    Create a use-case P&L. Every model system needs a baseline, named owner, benefit mechanism, full operating cost, risk class, and review cadence.

    Reuse platform controls. Identity, data access, evaluation, logging, deployment, drift detection, and incident response should not be rebuilt by each team.

    Validate by risk. A marketing recommendation and a credit decision require different evidence, explainability, and authority.

    Measure realized value. Forecast benefit should be separated from observed operational change and finance-approved realized value. Double counting across use cases must be prevented.

    Retire models. A large estate needs decommission criteria when performance, cost, policy, or business relevance changes.

    For generative systems, banks should add groundedness, citation precision, harmful-action testing, prompt-injection resistance, sensitive-data leakage, and cost per accepted outcome.

    5. What Financial-Services Leaders Should Take Away

    DBS demonstrates that enterprise AI advantage comes from a portfolio and an operating system. More than 430 use cases can create material value only when they share data, controls, and delivery methods.

    A production-ready banking system combines private RAG, document intelligence, regulated speech and voice agents, domain SLMs, routing, and evaluation. The bank should own the policy logic, evaluation data, and decision history.

    The central metric is net realized value per use case after model, infrastructure, control, and change costs. The SGD 1 billion disclosure establishes the potential scale, but the discipline behind the number is the more reusable lesson.

    Sources

    • DBS Annual Report 2025, CEO reflections
    • The portfolio and value figures are first-party disclosures. Readers should not assume the same accounting treatment or returns without DBS's model-level methodology.