Most banks and fintechs already use AI. Building it well is the hard part.
The Bank of England and the FCA found that 75% of UK financial firms were already using AI in 2024, with another 10% planning to start within three years. In the same survey, 46% of firms said they only partly understood the AI they use, and a third of use cases came from third-party providers. McKinsey estimates that generative AI could add $200 billion to $340 billion a year to banking, equal to 9% to 15% of operating profits. Yet in a Gartner survey of 644 organizations across industries, only 48% of AI projects reached production, and the move from prototype to production took about eight months on average.
The gap is not interest. It is execution. This guide covers what involves, which use cases fit which techniques, how each build stage maps to the controls regulators expect, and what Singapore's new AI risk guidelines mean for your roadmap.
As an AI & Blockchain Innovation Partner, SotaTek helps financial institutions plan, build and run AI systems with security, audit trails and human review designed in from the start.
What Is AI Development for Financial Services?
AI development for financial services is the work of designing, building, testing, deploying and monitoring AI systems for regulated financial workflows. Those workflows include lending, payments, fraud control, compliance and customer service.
It differs from general AI development in three ways:
- Decisions affect money and rights. A wrong credit decline or a missed fraud alert has direct customer and legal impact.
- Regulators expect evidence. You need documentation, validation and audit trails, not just a model that works.
- Legacy systems set the pace. Core banking, payment gateways and KYC tools were not built for AI, so integration often takes longer than modeling.
Most financial AI systems combine four layers: data, models, a workflow layer that decides what happens with each output, and a governance layer that records and controls it all. Our guide to enterprise AI architecture covers these layers in more depth.
Where AI Delivers Value: Use Cases by Function
McKinsey's banking analysis puts the largest AI value pool in risk and legal, at $385 billion across generative AI, advanced AI and analytics. Corporate banking (321billion)andretailbanking(306 billion) follow.
| Function | Typical use case | Common technique | Risk to design for |
| Fraud and AML | Transaction scoring, alert triage, account network analysis | Machine learning, graph analytics | False positives, drift as fraud tactics change |
| Credit and underwriting | Default prediction, affordability checks | Predictive ML | Explainability, fairness testing |
| KYC and onboarding | ID extraction, liveness checks, sanctions screening | Computer vision, document AI | Spoofed documents, edge cases |
| Customer service | Assistants for status, card and product questions | LLM with RAG | Wrong answers, safe handoff to staff |
| Compliance and reporting | Case summaries, policy search, draft reports | LLM with RAG | Source traceability, human sign-off |
Start with one workflow where you already know the baseline, such as alert review time or onboarding turnaround. For more on automation use cases, read our guides to AI automation in fintech and AI in fintech.
Choose the Right AI Technique Before You Build
McKinsey also notes that generative AI is not always the right tool, and that traditional analytical AI is sometimes the better fit. That matters in finance, where a simpler model is often easier to explain and validate.
| If the task is... | Use | Why | Watch for |
| Scoring or classifying structured data (default risk, fraud) | Classical ML | Cheaper, faster, easier to explain | Needs clean, labeled history |
| Answering questions from policies and product documents | LLM with RAG | Answers come from current, approved documents | Retrieval quality, stale files |
| Matching a fixed tone, format or vocabulary | Fine-tuning | Changes how the model behaves | Training data cost, re-tuning |
| Reading IDs, forms and statements | Document AI, computer vision | Turns files into structured data | Poor scans, forged documents |
| Multi-step work across systems | AI agents | Coordinates tools and steps | Permissions, action limits, human override |
A simple rule helps. If a rule or scorecard solves the problem and your regulator already understands it, keep it. Add AI when it beats that baseline by a margin you can measure.
To compare retrieval and tuning in detail, see RAG vs fine-tuning. For multi-step agents, see our guide to AI agent platforms.
The AI Development for Financial Services Lifecycle, Mapped to Regulatory Controls
Most guides list build stages. Few show what a supervisor will ask at each one. The table pairs common build stages with the control themes MAS names for the AI life cycle: data governance, testing, human oversight, cybersecurity, monitoring and change management. The pairing is our own reading, not an official MAS table.
| Stage | What the team does | Evidence to keep |
| 1. Scope and risk-rate | Define the decision, baseline and owner. Rate impact, complexity and reliance on the AI | AI register entry, risk rating, named owner |
| 2. Prepare data | Check relevance, accuracy, completeness and recency. Classify and secure data | Data lineage, quality checks, access rules |
| 3. Build and test | Train, compare with the baseline, test bias, robustness and edge cases | Test reports, fairness results, limits of use |
| 4. Validate | Get independent review before go-live for high-risk use cases | Validation report, sign-off |
| 5. Integrate and deploy | Connect to core systems. Set approval rules, human override and a fallback process | Architecture record, override log, tested fallback |
| 6. Monitor | Track drift, errors, overrides, complaints and security events | Dashboards, thresholds, incident log |
| 7. Change and retire | Version models, re-validate after material changes, decommission cleanly | Version history, change approvals, retirement plan |
Two details are easy to miss. Monitoring should cover data drift (the inputs change) and concept drift (the link between inputs and outcomes changes). And for high-risk use cases, MAS expects fallback options such as manual processes, and expects firms to test them. For the build side of each stage, see our AI development process guide.
What the New MAS AI Guidelines Mean for Your Build
On 7 October 2026, the Monetary Authority of Singapore issued its final Guidelines on AI Risk Management. They apply to all financial institutions MAS supervises and cover all forms of AI, including generative AI. They set supervisory expectations that firms can apply in proportion to their risk profile and scale of AI use.
The timeline is phased:
- 7 October 2027: Sections 3 and 4 apply. They cover board and senior management oversight, plus AI identification, inventory and risk-materiality assessment.
- 7 October 2028: Sections 5 and 6 apply. They cover life cycle controls, including third-party AI, and capability and capacity.
MAS asks firms to rate each AI use case on three dimensions: impact, complexity and reliance. Here is a practical way to read them. It is our interpretation, not MAS wording.
| Dimension | Question to ask | Example of higher risk |
| Impact | What happens to customers or the firm if the output is wrong? | A credit decline, a missed sanctions hit |
| Complexity | How hard is it to explain, test and reproduce the output? | A large language model or a stack of linked models |
| Reliance | How much does the business depend on it, and how much autonomy does it have? | AI acts without review and no manual fallback exists |
MAS also says firms should apply life cycle controls to high-risk use cases as soon as possible, not wait for the end of the window. Three things are worth building in now:
- An AI register entry for every system, from day one. Record the owner, purpose, data sources, model or vendor, autonomy level and risk rating. Include AI embedded in vendor software.
- Clear third-party accountability. Using a vendor model does not move accountability away from your institution. Ask for independent assessments rather than self-attestations, and plan an exit.
- Limits for agentic AI. MAS plans to consult the industry on agentic AI guidance in 2027. Set permission limits, action caps and human override now.
If you work outside Singapore, check your own regulator's position. This section is general information, not legal advice.
Build, Buy or Partner?
Third-party implementations made up a third of AI use cases in the Bank of England survey, up from 17% in 2022. Most institutions will mix routes.
| Route | Best when | Trade-off |
| Buy a packaged tool | The task is standard and low risk, such as meeting notes | Less control over data, updates and testing, while you keep the risk |
| Build in-house | The workflow sets you apart and uses proprietary data | Needs MLOps, validation and security skills on your team |
| Partner to build | You need speed and control, and your AI team is small | Needs clear terms on IP, security and knowledge transfer |
For deeper criteria, see our custom AI vs off-the-shelf AI comparison and a scorecard for evaluating AI development companies.
Budget for the whole lifecycle, not just the build. According to IBM, as cited in SotaTek's AI development cost guide, annual maintenance and monitoring runs about 15% to 20% of the original build price. SotaTek's AI outsourcing price guide gives indicative figures of about $40,000 for a straightforward fintech automation and $150,000 or more for an end-to-end departmental workflow.
Five Mistakes That Stall Financial AI Projects
- Starting with the model instead of the workflow. Define the decision, the owner and the handoff to humans first.
- Skipping the baseline. Without current alert, onboarding or approval times, you cannot prove value later.
- Judging success by accuracy alone. Track false positives, override rates, time per case, complaints and drift too.
- Underestimating data work. Poor, duplicated or outdated records hurt performance. An AI readiness checklist helps you find gaps early.
- Having no fallback. Decide what happens when the model is down, drifting or wrong, and test that process.
Conclusion
Strong AI development for financial services comes down to five habits. Pick a workflow with a measurable baseline. Use the simplest technique that beats it. Build evidence and controls into every stage. Treat vendor AI as your own risk. Start your AI register now, because the MAS timeline begins in October 2027.
Planning an AI build in banking, insurance or fintech? Talk to SotaTek about a risk-rated use-case review and a controlled pilot plan.
https://www.sotatek.com/ 
(+84)2466585248 
