How Is AI Being Used in Finance? Generative AI in Finance, With Singapore Examples
Generative AI in finance drafts, summarises and searches for the people who work in banks, insurers and investment firms, and it also answers customers directly, so a governed deployment sets human approval, escalation and monitoring according to the materiality of each use case. For drafts that reach a client file or a regulated record, a governed design works like maker-checker: the model makes the draft, and a person checks it at the points the design sets before it leaves the desk. Staff assistants came early, such as OCBC GPT for 30,000 employees from November 2023, and in a DBS release of 28 July 2026 DBS reported that its customer-facing assistants, DBS Joy and DBS digibot, reach more than 10 million users across Singapore, Hong Kong and Taiwan. This guide sets out, as of September 2026, what generative AI in finance covers, how it works, where it fails, which Singapore institutions use it and how, and the order in which an institution adopts it.
What Generative AI in Finance Means
As of September 2026, generative AI in finance means large language models and related models that produce new text, code, summaries or answers for financial work, as distinct from the predictive models that score, rank or flag. The category covers drafting research notes and client letters, summarising calls and documents, answering staff questions from internal policy, translating, and writing or reviewing code. It excludes the credit scorecards, fraud models and rules engines that institutions have run for years, even when those sit on the same data platform.
The boundary matters in practice because the two families fail differently. A predictive model estimates a score, probability, ranking or label, and validators test it against measured outcomes for measures such as discrimination, calibration and stability. A generative model returns fluent prose that reads as correct whether or not it is, so teams evaluate it for factual accuracy, grounding in source material, relevance, safety and task performance, and a sound deployment adds a review step to those tests.
How does generative AI differ from the AI banks already run?
The main difference between generative AI and the AI banks already run is the output: generative AI writes new text, while fraud and credit models predict a score or a label from past data. A fraud model tells an analyst which transaction to look at. A generative assistant writes the case summary the analyst then reads, corrects and signs. Teams validate the first against labelled outcomes, and evaluate the second for factual accuracy, grounding, relevance and safety, with the rate at which reviewers accept drafts unchanged as one measure among them, which puts the reviewer inside the system design rather than after it.
How Generative AI Works in Finance, From Staff Assistants to Customer Channels
As of September 2026, a common governed architecture for generative AI in finance, especially for higher-risk uses, retrieves the institution’s own approved material, drafts from it, and routes the draft to review points, set by the use case’s risk, where a person accepts, edits or rejects it before anyone files or sends it. Banks already call that control maker-checker. Retrieval and review are design choices rather than properties of the technology: some tools answer from the prompt or from the model’s general training, and DBS’s customer assistants answer authenticated customers directly and execute everyday banking tasks on their instructions, with live chat to a human agent when needed, according to the DBS release of 28 July 2026.
A retrieval-grounded deployment runs three mechanical steps:
- Retrieves the relevant passages from approved sources such as a knowledge base, product terms or research archive, so the model answers from the institution’s material rather than from its general training.
- Drafts the output in the format the job needs: a call summary, a prefilled service request, a research paragraph or a translated letter.
- Routes the draft to a named reviewer, whose acceptance, edit or rejection becomes the record and the feedback that tunes the next release.
Retrieval over the institution’s own documents
Retrieval is the step that keeps a generative answer tied to what the institution has approved, and a customer-service assistant shows it plainly. DBS described its CSO Assistant in a DBS media release of 18 July 2024 as transcribing customer queries in real time, searching the bank’s knowledge base live, and producing call summaries and prefilled service-request fields. The same release targeted 500 customer service officers in Singapore by the end of 2024 and an expected reduction in call handling time of up to 20%, using a model that DBS tailored to local languages and parlance. The officer stays on the call and owns the answer; the assistant removes the searching and the typing.
Drafting and summarising for knowledge workers
Enterprise assistants that draft and summarise for employees are among the widest staff-facing generative deployments DBS and OCBC have published. OCBC’s media release of 24 October 2023 announced OCBC GPT for its 30,000 employees globally from November 2023, after a trial of about 1,000 staff between April and September 2023 whose participants reported completing tasks about 50% faster. OCBC listed writing, research, ideation, investment research report writing, translation and drafting customer responses as the tasks, and stated that the bank keeps the information staff enter within the bank and does not make it available to Microsoft or other external parties. The DBS release cited above reports a comparable internal tool, DBS-GPT, available to over 25,000 employees for content generation and writing in a secure environment. The 50% figure is a self-reported trial result, not a measured productivity change across the bank.
Assistants inside the investment process
Investment firms apply the same pattern to diligence and committee work, with the assistant informing a decision that people still take. The Temasek Review 2026 strategy section describes a multi-model generative AI environment for employees, a customised tool that investment teams use from diligence planning through investment committee preparation to portfolio monitoring, and a customised assistant that supports the investment committee by surfacing deeper insights, independent views and potential blind spots. Temasek also reports deploying agentic AI for teams with specialised, high-value workflows. The committee remains the decision-maker; the assistant widens what it sees before it decides.
How can generative AI assist in finance?
Generative AI assists in finance by producing first drafts of work that staff already do: summarising customer calls, answering policy questions from approved documents, drafting research and credit notes, translating and writing code. It saves the drafting time, and the reviewing person keeps the decision and the sign-off.
Where Generative AI Fits Among Finance AI and the Rules and Frameworks That Apply
As of September 2026, generative AI is one branch of AI in finance, and in Singapore it falls under the same proposed supervisory expectations as every other AI use rather than under a separate regime. Predictive machine learning, agentic AI, robotic process automation and rules engines sit beside it, each with a different output and a different reviewer.
| Technology | What it produces | Typical finance job | Who reviews the output | Reference and its status |
|---|---|---|---|---|
| Generative AI (large language models) | New text, summaries, code, answers | Call summaries, research drafts, policy answers, customer chat | Set by materiality: a named employee before filing or sending a draft; escalation and monitoring for customer-facing assistants | MAS Guidelines on AI Risk Management (proposed supervisory guideline); PDPA (legislation) where personal data is used |
| Predictive machine learning | Scores, probabilities, labels | Fraud alerts, credit risk, transaction monitoring | Model owner and second-line validation | MAS Guidelines on AI Risk Management (proposed supervisory guideline); MAS FEAT Principles (principles document, not a Notice) |
| Agentic AI | Actions taken through connected tools | Multi-step workflows across systems | Human approval at defined action points | MAS Guidelines on AI Risk Management (proposed supervisory guideline; covers AI agents) |
| Robotic process automation | Scripted keystrokes and transfers | Data entry, report assembly | Process owner through exception queues | Existing operational and technology risk controls (institution’s own framework) |
| Rules engines | Fixed decisions from coded logic | Eligibility checks, limit checks | Business owner through rule change control | Existing change management controls (institution’s own framework) |
The practical consequence is that an institution can bring generative AI into the enterprise or model-risk framework it already runs, once it has assessed that framework for gaps and added the generative-specific controls, such as grounding and output review. The FSB’s November 2024 report found that existing regulatory and supervisory frameworks address many of the vulnerabilities of AI adoption, and flagged possible further work to ensure they are sufficient. The guide that reinforces Responsible AI in Financial Services through FEAT, Veritas and model-risk evidence covers MAS’s FEAT Principles and the Veritas toolkit, the fairness and transparency layer that sits over the AI rows of this table.
Is generative AI regulated separately in Singapore’s financial sector?
MAS does not regulate generative AI separately: its proposed Guidelines on AI Risk Management, which MAS issued for consultation on 13 November 2025, cover generative AI, AI agents and every other AI use by financial institutions. On 5 August 2026 the MAS Chairman told Parliament, in a written reply, that MAS will finalise them soon. The reply also states that the Guidelines apply to all AI use cases by financial institutions, including agentic AI. The same November 2025 release describes Guidelines that apply to all financial institutions in proportion to each institution’s size, activities, AI use and risk profile, sets expectations on board and senior management oversight, AI inventories, risk materiality assessments and life-cycle controls, and invited comments until 31 January 2026. As at 29 September 2026 the Guidelines remain proposals, so any decision that depends on them rechecks their status first.
Where personal data enters a prompt
Data in a prompt, a retrieved context window or a model output is personal data where it is about an individual who can be identified from it, or from it and other information the organisation has or is likely to have access to, and the PDPA obligations that follow depend on the organisation’s role. The PDPC’s Advisory Guidelines on Key Concepts in the PDPA (the linked PDF is marked “Revised 29 April 2026”) set out that definition, and explain that a data intermediary processing personal data on another organisation’s behalf under a written contract carries only the protection, retention limitation and breach-notification obligations for that processing. Singapore’s Personal Data Protection Act 2012 governs the collection, use and disclosure of personal data by organisations. A generative deployment therefore records what personal data enters the retrieval layer, the purpose of its collection, and where and for how long the system stores prompts and outputs.
The Limits and Failure Modes of Generative AI in Finance
As of September 2026, the characteristics that limit generative AI in finance are confident errors, exposure of confidential data, dependence on a small number of model providers, and its use against institutions by fraudsters. Each one follows from how the technology works rather than from a particular vendor, so each needs a control that the institution designs in before release. The wider set of definitions and frameworks behind these controls sits in the collection that categorises the AI Resources & Guides Library by question. The controls column is Synthesised.ai’s illustrative risk treatment, and the last column sources the risk rather than the control.
| Failure mode | What it looks like in finance | Where it bites | Control that reduces it (illustrative) | Source for the risk |
|---|---|---|---|---|
| Hallucination | A fluent answer citing a product term, rate or clause that does not exist | Customer responses, research notes, policy answers | Retrieval from approved documents, citation of the source passage, reviewer sign-off | ABS Handbook on Generative AI Guardrails in Banking, 24 March 2026 |
| Confidential data exposure | Client data pasted into a public tool, or surfaced to the wrong user by an internal one | Any prompt or retrieval index holding client or deal data | Enterprise tenancy, access controls on the retrieval index, logging | MAS information paper on generative AI cyber risks, 30 July 2024 |
| Prompt manipulation | Instructions hidden in a document that redirect the model | Assistants that read external emails or files | Input filtering, restricted tool permissions, testing against known attacks | MAS information paper on generative AI cyber risks, 30 July 2024 |
| Third-party concentration | Many institutions depending on the same few model providers | Outages, pricing changes, correlated errors | Exit planning, model-agnostic design, outsourcing controls | FSB report, 14 November 2024 |
| Fraud and disinformation | Deepfaked voices or documents aimed at staff and customers | Payments, onboarding, market communications | Call-back verification, document-authenticity checks, biometric liveness detection where onboarding uses it, staff training | FSB report, 14 November 2024 |
Confident answers that are wrong
A generative model produces a plausible answer even when the right one is missing from its sources, and in finance a plausible wrong figure costs more than a visible gap. The Association of Banks in Singapore’s Handbook on Generative AI Guardrails in Banking (24 March 2026, PDF), listed on the ABS data management publications page separately from the ABS SCDM Gen AI Guardrails and Controls Library of 19 May 2025, which the ABS Standing Committee on Data Management developed with member banks, groups banks’ enterprise generative AI use into seven use-case categories, names hallucination among the top risks and sets out nine guardrail approaches. The controls that reduce the error in practice are grounding every answer in retrieved, approved passages, showing the reviewer which passage the answer came from, and blocking outputs that cite nothing. Reduction is the realistic goal, and the reviewer remains the last check.
Confidential data in prompts and outputs
Every prompt is a data flow, so a generative deployment is also a data-protection and cyber deployment. MAS circular MAS/TCRS/2024/05 of 30 July 2024 published an information paper on the cyber risks of generative AI, giving financial institutions an overview of the key threats, their risk implications and mitigation measures. The design questions follow directly: which tenancy the model runs in, whether the provider retains prompts, who has access to which documents through the retrieval index, and whether the system logs outputs for review.
Concentration on a few model providers
Institutions building on the same few foundation models share their outages, their price changes and their blind spots. The Financial Stability Board’s report of 14 November 2024 lists four AI-related vulnerabilities with systemic potential: third-party dependencies and service provider concentration, market correlations, cyber risks, and model risk, data quality and governance. For a single institution, Synthesised.ai’s illustrative response is an exit plan for each model provider and an architecture in which the model is replaceable without rebuilding the retrieval and review layers around it.
Fraud and disinformation aimed at the institution
The same technology that drafts a client letter also forges one, and the FSB report cited above states that generative AI increases the potential for financial fraud and disinformation in financial markets. Voice-cloned instructions and synthetic documents target payment approvals and onboarding. In Synthesised.ai’s illustrative control design, call-back verification to known numbers, document-authenticity checks and, where onboarding uses biometrics, face matching with liveness detection counter the attack at the point where it lands.
What are the risks of generative AI in finance?
The main risks of generative AI in finance are confident wrong answers, confidential data leaking through prompts or outputs, dependence on a few model providers, and fraud or disinformation aimed at the institution. The Financial Stability Board’s November 2024 report lists third-party concentration, market correlations, cyber risk and model risk as the AI vulnerabilities with systemic potential.
Generative AI Applications in Finance, With Singapore Examples
As of September 2026, generative AI applications in finance fall into six jobs, and Singapore’s three local banks have each published deployments in at least one of them. The six jobs are service-officer assistance, employee knowledge assistants, research and report drafting, investment process support, customer-facing assistance, and code and data work. The table maps each job to the reviewer who checks its output. The review intensity column is Synthesised.ai’s illustrative assessment, not a regulatory classification.
| Application | What the model does | Typical user | Singapore example (published) | Review point (typical design) | Review intensity (illustrative) |
|---|---|---|---|---|---|
| Service-officer assistance | Transcribes, searches the knowledge base, summarises, prefills | Contact-centre officer | DBS CSO Assistant | Officer confirms before the request is lodged | Medium |
| Employee knowledge assistant | Answers questions from internal policy and documents | Any employee | OCBC GPT; DBS-GPT; UOB build-your-own bots | Employee judges the answer before use | Low to medium |
| Research and report drafting | Drafts research and client-report text | Analysts, advisers | OCBC GPT (investment research report writing) | Analyst and compliance review before publication | High |
| Investment process support | Organises diligence, prepares committee material, surfaces blind spots | Investment teams, committees | Temasek customised tools and committee assistant | Committee decides | High |
| Customer-facing assistance | Answers customer questions and, after login, executes everyday banking tasks | Corporate or retail customers | DBS Joy (corporate) and DBS digibot (retail) | Agentic actions only for logged-in customers on their own instructions; live chat to a human agent when needed | High |
| Code and data work | Writes and reviews code, queries data dictionaries | Engineers, data teams | UOB Daita IQ (data dictionaries) | Code review and testing before deployment | Medium |
How is AI being used in finance?
As of September 2026, AI in finance spans several technologies, including agentic AI, robotic process automation and rules engines, and this guide focuses on two major model families: predictive machine learning that scores fraud, credit and transaction risk, and generative AI that drafts, summarises, searches and answers. Generative models run call summaries, research drafts, policy answers, code and customer chat. Singapore’s disclosed generative deployments now serve both staff and customers: OCBC GPT supports 30,000 employees, and DBS reported on 28 July 2026 that DBS Joy and DBS digibot reach more than 10 million users across Singapore, Hong Kong and Taiwan.
MAS’s Managing Director told the Global FinTech Fest on 11 September 2026 that common AI use cases in Singapore’s financial sector include fraud detection, credit underwriting, risk management, regulatory compliance, marketing, customer service and document processing, and that these use cases have moved beyond pilots to deployment at scale.
DBS: models at scale and customer assistants at regional scale
DBS reports running AI at a scale that makes generative tools one layer of a much larger programme. The DBS Annual Report 2025 reports approximately SGD 1 billion of economic value from data analytics and AI/ML, over 2,000 models across more than 430 use cases, and DBS Joy, a generative AI corporate banking bot that the 2025 report puts at over 20,000 customers. DBS’s 28 July 2026 release updates that scale: DBS Joy and DBS digibot bring generative and agentic capabilities to more than 10 million users across Singapore, Hong Kong and Taiwan, DBS Joy lets 350,000 corporate users in Singapore execute everyday banking tasks, and DBS digibot is available to over five million customers in Singapore and four million across Hong Kong and Taiwan. The SGD 1 billion is DBS’s own measure and covers analytics and machine learning as a whole, so it does not isolate the contribution of generative AI.
OCBC: an assistant on every desk
OCBC describes generative AI as one part of an AI estate that already makes millions of decisions a day. The OCBC Annual Report 2024 AI section reports about six million decisions that AI makes in the bank daily, over 300 use cases, OCBC GPT supporting 30,000 employees worldwide, and OCBC Buddy, a knowledge assistant OCBC credits with an employee productivity boost of up to 50%. The productivity figures are OCBC’s own reported results rather than independent measurements.
UOB: build-your-own bots and Copilot
UOB publishes a self-service approach in which teams build their own domain chatbots on a shared generative platform. UOB’s Gen AI Enablement page for the Singapore FinTech Festival 2025 describes Build-Your-Own-Bot for onboarding, policy clarification and operational guidance, Microsoft Copilot embedded in Microsoft 365 for drafting, summarising and translation, and the Daita IQ chatbot for technical information from data dictionaries. The page lists expected benefits and publishes no usage or time-saving figures.
Which financial institutions in Singapore use generative AI?
DBS, OCBC and UOB each publish generative AI deployments: DBS-GPT, a service-officer assistant and the customer assistants DBS Joy and DBS digibot at DBS, OCBC GPT for 30,000 OCBC employees, and build-your-own bots with Microsoft Copilot at UOB. Temasek describes a multi-model environment and an investment committee assistant, and GIC sits in MAS’s Project MindForge consortium.
How Financial Institutions Adopt Generative AI
As of September 2026, institutions adopting generative AI in Singapore work from a regulator-published reference for the controls, MAS’s information paper of 5 December 2024, and can follow Synthesised.ai’s recommended five-step sequence, informed by those MAS materials, that runs from an AI inventory to a reviewed staff-facing use case before anything faces customers. MAS does not prescribe this order. The paper sets out good practices for AI and generative AI model risk management that MAS observed in its mid-2024 thematic review of banks, covering governance and oversight, key risk management systems and processes, and development and deployment. Project MindForge, the MAS-led industry consortium whose phase two members include DBS, OCBC, UOB and GIC, has since published an AI risk management toolkit of executive and operationalisation handbooks and implementation examples.
What does a first generative AI deployment in finance look like?
A low-risk first generative AI deployment in finance serves staff rather than customers: an assistant that answers questions from approved internal documents or summarises calls, with review points set by the use case’s risk. It starts from an inventory of existing AI use, passes a fixed test set before release and stays under sampled review after it.
Step 1: Inventories every AI use, including tools staff already use
The first step lists every AI use in the institution, including generative tools staff adopted on their own, and rates each by materiality. The proposed MAS Guidelines expect accurate, current AI inventories and risk materiality assessments that factor impact, complexity and reliance, so the inventory doubles as the evidence base.
Step 2: Picks a staff-facing job with a checkable answer
The first use case is a staff-facing job where a trained person already checks the work and a wrong draft is caught before it leaves the building. Service-officer summaries and internal policy questions fit that test; unsupervised customer advice does not.
Step 3: Grounds the model in approved documents
Grounding points the model at a controlled set of approved documents and makes it cite the passage it used. The team decides the document set, its owner and its refresh cycle before it chooses the model, because answer quality depends on what the model retrieves as much as on the model itself.
Step 4: Tests before release and keeps testing
A release passes a fixed set of test questions with known answers before launch, and reviewers sample live outputs after it. MAS’s 20 March 2026 release on the MindForge toolkit describes a toolkit that a consortium of 24 banks, insurers, capital market firms and other industry partners developed. Its Operationalisation Handbook has four sections, scope and oversight, AI risk management, AI life-cycle management and enablers, and the release states that the four sections align with the proposed Guidelines.
Step 5: Decides buy, configure or build
Once the use case and controls are clear, the institution chooses between a packaged tool, a configured platform and a custom build on data residency, integration and audit needs. The comparison that validates AI Tools for Finance: Buy vs Build decisions against six criteria sets out those criteria for finance teams and IT buyers.
Who owns generative AI risk inside the institution?
Under the proposed MAS Guidelines, the board and senior management “play a key role in the governance and oversight of AI risk management”, including for generative AI, according to MAS’s consultation release (13 November 2025). Day-to-day ownership sits with the business owner of each use case and the model owner, and review intensity and approval points follow each use case’s risk materiality, in line with the proposed Guidelines’ expectation that controls apply based on relevance and in proportion to assessed risk.
Scope a Generative AI Build for Your Institution
As of September 2026, Synthesised.ai recommends starting a generative AI programme in finance with a staff-facing use case, a controlled document set and a named reviewer, and the next decision depends on whether the work sits in a financial institution’s lines of business or in a finance function. Carrying maker-checker discipline into the design turns a pilot into a production system, and a scoped engagement catalyses a governed build through Generative AI Consulting that matches use cases to guardrails before the institution buys any licence. Synthesised.ai serves organisations in Singapore, Australia, New Zealand and South East Asia with one practical approach, Discover, Prototype and Productionise, adapted to the business and use case, with phase lengths agreed for each engagement. Synthesised.ai does not publish a rate card and scopes fees through an enquiry.
For banks, insurers, wealth managers and fintechs
Institutions whose products are financial services start from the use cases their model-risk framework already governs. The industry guide that prioritises AI for Financial Services use cases by model-risk readiness sequences fraud, credit, service and reporting work.
For finance teams in any industry
CFO offices start from the close, where every automated posting still needs an audit trail. The function guide that automates reconciliation and the close through AI for Finance Teams covers close, reconciliation, payables and forecasting.
Start a scoped conversation
Every commercial step begins with a scoping conversation, and a written proposal follows it. No list prices appear on this site. Tell us the problem, the stack you already run and the timeline.
Book a discovery call →