Technology Insights

AI Adoption in Financial Services: Use Cases, Vendor Types, Costs and Compliance

The four most common AI use cases in financial services are customer service, compliance, fraud prevention and internal knowledge bases. When choosing a vendor, first decide whether you are buying a platform, buying a product, or hiring a team for custom integration. This article covers the data, risks and human review points for each use case, maps the Guidelines for the Use of Artificial Intelligence (AI) in the Financial Industry issued by Taiwan's Financial Supervisory Commission in 2024 to what you need to prepare, and lists cost ranges by project stage.

By Founder & CEO, JoinXPublished 10 min read
AI Adoption in Financial Services: Use Cases, Vendor Types, Costs and Compliance

Key takeaways

  • The most common AI use cases in financial services are customer service, compliance, fraud prevention and internal knowledge bases, and their data and risks differ too much to share one approach.
  • The AI guidelines issued by Taiwan's Financial Supervisory Commission in June 2024 are administrative guidance that asks institutions to apply six core principles on a risk basis and to oversee third-party vendors.
  • When AI is outsourced, the financial institution must still be able to explain the system's logic, keep human oversight, and agree on how data and systems migrate if the engagement ends.
  • Financial AI vendors fall into three types: cloud and model platforms, packaged AI products and custom integration teams; the difference is whether they can connect to existing core systems and permissions.
  • Based on JoinX's custom project pricing in Taiwan, a single-use-case PoC costs about NT$300,000 to 800,000, a production launch about NT$800,000 to 2.5 million, and cross-department, multi-system projects are quoted by scope.
Contents

When a financial institution adopts AI, the first decision is not which model to use, but who is responsible when the use case goes wrong and how to recover. On June 20, 2024, Taiwan's Financial Supervisory Commission (FSC) issued the Guidelines for the Use of Artificial Intelligence (AI) in the Financial Industry, turning six principles (governance, fairness, privacy, security, transparency and sustainability) into concrete practices for adopting, using and managing AI. Below we cover what financial institutions should look at when evaluating AI projects, by use case, compliance, vendor type and cost.

What are the most common AI use cases in financial services?

The most common are four: customer service, compliance, fraud prevention and internal knowledge bases. They differ in data sources, cost of errors and where human review sits, so evaluate them separately rather than with one acceptance standard.

Use caseTypical usesMain dataHuman review pointCost of errors
Customer serviceAnswering common questions on products, procedures and fees; handing off to staffProduct descriptions, FAQs, operating rulesQuestions involving transactions, complaints or personal accounts go to staffMedium: wrong answers affect customer rights and complaint volume
ComplianceComparing regulatory changes, looking up internal rules, first-pass document reviewLaws, official letters, internal rules, contractsCompliance staff confirm the resultHigh: wrong judgments can lead to penalties
Fraud preventionFlagging unusual transactions, identifying suspicious accounts and conversationsTransaction records, account behavior, reported casesStaff decide whether to block or reportHigh: false positives affect normal customers, misses cause losses
Internal knowledge baseStaff looking up operating manuals, product specs and SOPsInternal documents, training materialsAnswers cite sources and staff judge for themselvesLow to medium: mostly internal use

Customer service: get common questions right before automating transactions

Customer service is the first AI use case to go live in most financial institutions. The safer approach is to let AI answer from a knowledge base with cited sources, and always hand off account inquiries, transactions and complaints to staff. Once the accuracy rate is stable, consider letting AI query account data through APIs. For the approach and costs, see AI customer service.

Compliance: AI screens, compliance staff decide

Compliance suits AI for "comparison and summarization": finding internal rules affected by a regulatory amendment, condensing long official letters into key points, and flagging contract clauses that need attention. The final judgment belongs to compliance staff, and the system should record which documents the AI relied on so auditors can trace it.

Fraud prevention: AI flags risk, people decide on blocking

Fraud systems usually add a layer of AI analysis on top of existing transaction monitoring, giving risk alerts on unusual transactions or suspicious accounts. Define in advance how many false positives and misses are acceptable, and keep channels for staff review and customer appeals. These systems directly affect customer rights, so pay particular attention to the fairness and explainability described in the guidelines.

Internal knowledge base: the best first use case

An internal knowledge base is used only by staff, answers cite sources, and staff judge wrong answers for themselves, so the cost of errors is relatively low. It is also the foundation for the other three: the knowledge used by customer service and compliance usually comes from the same documents. Many financial institutions start here, cleaning up documents and setting permissions first.

What do the FSC's AI guidelines require for adoption?

The guidelines are administrative guidance with no penalties, but they ask financial institutions to apply the six core principles on a risk basis and to oversee third-party vendors. In practice, when AI is outsourced, these requirements become project documents, permission design and contract terms, and the vendor must be able to help produce them.

The guidelines consist of general provisions and six chapters. The general provisions define the AI system lifecycle and risk assessment factors, and require oversight measures when engaging third parties, including a data or system migration mechanism if the engagement ends. The table maps the six chapters to what you need to prepare:

ChapterKey pointsWhat to prepare
1. Governance and accountabilityDefine AI purposes, applicable business and responsible staff; set procedures for errors and unexpected eventsUse case list and owners, incident procedure, independent third-party evaluation for high-risk systems where needed
2. Fairness and human-centered valuesWatch for bias and avoid discrimination; staff must manage the risk of generative AI outputTest data covering different customer groups, human review points written into the workflow
3. Privacy and customer rightsData minimization and leak prevention; respect customers' choice of whether to use AI servicesData classification, which data never goes into models, keeping a human service channel
4. System robustness and securitySecurity protection; risk management and oversight when using third-party AI systemsSource code and vulnerability scans, permissions and activity logs, risk assessment of third-party services
5. Transparency and explainabilityExplain the logic internally and disclose how AI is used externallyTechnical documents on architecture and decision basis, disclosure text for customers
6. Sustainable developmentInclude sustainability in AI strategy; train employeesCompute resource planning, staff training

Financial institutions that outsource operations must also follow the FSC's outsourcing rules, covering outsourcing scope, data protection, audit rights and handover arrangements. AI projects are no exception, so confirm that a vendor can support audits and documentation requirements.

What types of AI vendors serve financial services?

There are roughly three types: cloud and model platforms, packaged AI products and custom integration teams. The difference is not the strength of the AI technology, but whether they can connect to the institution's existing core systems, permissions and audit processes.

Vendor typeWhat they provideBest forWatch out for
Cloud and model platformsLarge language model APIs, cloud compute, vector databasesInstitutions with their own IT team that can integrateConfirm data location and cross-border transfer first; you do the integration
Packaged AI productsReady-made chatbots, document recognition, transaction monitoringStandard workflows close to the product's featuresLicensed or usage-based pricing; custom work to connect existing systems is extra
Custom integration teamsAI features designed around your workflow, connected to existing systems and permissionsConnecting core systems, complex permissions, audit documentationCheck financial outsourcing experience and how source code and documents are delivered

A common mix in practice: the model runs on a cloud platform, and a custom team builds the interface and system integration. Ask three questions when choosing: will data leave the designated environment, how are systems and data migrated if the engagement ends, and who can explain why the AI answered the way it did when something goes wrong.

How much does AI adoption cost in financial services?

Based on JoinX's custom project pricing in Taiwan, the requirements review is free, a single-use-case PoC costs about NT$300,000 to 800,000, a production launch about NT$800,000 to 2.5 million, and cross-department, multi-system projects are quoted by scope. The difference in financial services is mostly the work for security reviews, audit documents and system integration.

StageEstimated costTypical scopeTimeline
Requirements reviewFreeUnderstand your workflow and data, judge whether AI fits, suggest the first use case1–2 meetings
Single-use-case PoCNT$300,000–800,000One use case, limited data, tested under human supervision4–8 weeks
Production launch (single use case)NT$800,000–2.5 millionPermissions, system integration, monitoring, testing and deployment, audit documents2–4 months
Cross-department, multi-systemQuoted by scopeMultiple knowledge bases or AI agents, cross-system workflows and full governance4+ months

AI model API and cloud hosting costs are billed separately based on actual usage. These figures are a starting point for budgeting, not a fixed quote; the more systems to integrate and the stricter the acceptance criteria, the closer the cost gets to the top of the range.

What has JoinX done in financial services and AI?

JoinX's financial clients include Chang Hwa Commercial Bank, Taipei Fubon Bank, KGI Securities and Financial Information Service Co. (FISC). JoinX maintains Capital Investment Trust's website, ETF site and app, and Capital Investment Trust describes JoinX as "an IT outsourcing developer that meets the Financial Supervisory Commission's required standards." Our AI projects span insurance, research institutes, publishing and platforms.

JoinX took over Capital Investment Trust's website, app and e-commerce system from another team in 2023, first stabilizing the systems and completing security adjustments and audits under FSC rules, then rebuilding the ETF website and improving ETF electronic applications. See the Capital Investment Trust case study.

On the AI side, these are our published cases:

  • Insurance company: In 2026 we designed the conversational interface and workflow of an AI agent for insurance sales, letting AI help look up and organize information in the order sales staff work, with a person confirming at the end.
  • National Atomic Research Institute (NARI): From 2022 to 2025 we built an AI analysis and automation system for eddy current signals; AI flags suspected defects first and staff review them. See the NARI case study.
  • Linking Digital: An audiobook platform with AI text segmentation and character voices, with acceptance thresholds of 90% for character recognition and over 75% for text segmentation. See the Linking Digital case study.
  • GoodParty: We added AI search to the platform JoinX built and continues to maintain, reusing the existing data and admin back end.

Before delivery, JoinX runs its own source code and vulnerability scans; black-box testing is done by a third party, which JoinX can help arrange, and we deliver after fixing the vulnerabilities found. For more on how these projects were done, see enterprise AI adoption cases; for our financial system work, see fintech system development.

How should a financial institution start with AI?

Start with a use case where errors are manageable, data is available and results can be reviewed by people, usually an internal knowledge base or common customer service questions. Do a requirements review and PoC first, connect to production systems only after passing the thresholds, and prepare the governance documents the guidelines call for along the way.

  1. List candidate use cases and classify them. Using the risk assessment factors in the guidelines, judge whether each use case directly affects customer rights or involves personal data.
  2. Confirm data and permissions. Which system holds the data, who can read it, and what must never go into generative AI.
  3. Define acceptance thresholds. For example accuracy, citation accuracy and handoff rate, tested with the same set of real questions.
  4. Design where human review happens. Decide which results people confirm and how errors are recovered and reported.
  5. Write it into the outsourcing contract. Agree on audit rights, data protection, delivery of source code and documents, and migration if the engagement ends.
  6. Go live only after the PoC passes. At launch, complete monitoring, activity logs and customer disclosures.

FAQ

Are the FSC's AI guidelines mandatory?

No. The FSC describes the guidelines as administrative guidance without penalties, and financial industry associations can refer to them when setting their own AI self-regulatory rules. However, the governance, human oversight and third-party oversight they call for are what internal audits and security reviews usually ask about anyway, so it is wise to prepare them during adoption.

Can financial institutions use generative AI such as ChatGPT?

Yes, but classify your data first. The guidelines note that even when using third-party generative AI whose training process the institution cannot control, staff must still manage the risk of its output. Customer personal data and sensitive data should not go directly into public services, and for enterprise services or private deployment, confirm data storage and usage terms first.

Which AI use case should a financial institution start with?

Most start with an internal knowledge base, because it is used only by staff, answers cite sources, and the cost of errors is relatively low. Common customer service questions usually come next. Compliance and fraud prevention directly involve penalties and customer rights, so it is better to wait until governance processes are running smoothly.

What should a financial institution watch for when outsourcing AI development?

Confirm that the vendor can support audits and documentation requirements, how source code and technical documents are delivered, and how data and systems migrate if the engagement ends. The guidelines also require financial institutions to oversee third-party vendors, so ask vendors directly about their experience supporting financial audits.

How long does AI adoption take in financial services?

From JoinX's project experience, the requirements review takes 1–2 meetings, a single-use-case PoC about 4 to 8 weeks, and a production launch about 2 to 4 months. Security reviews and internal approvals often take the longest at financial institutions, so involve security and compliance teams from the PoC stage.

Sources

  • Guidelines for the Use of Artificial Intelligence (AI) in the Financial Industry (issued June 20, 2024)Financial Supervisory Commission, Taiwan
  • Core Principles and Policies for the Use of AI in the Financial Industry (issued October 17, 2023)Financial Supervisory Commission, Taiwan
  • Regulations Governing Internal Operating Systems and Procedures for the Outsourcing of Financial Institution OperationFinancial Supervisory Commission, Taiwan
  • AI Risk Management Framework (AI RMF 1.0, 2023)National Institute of Standards and Technology (NIST)

These are the regulations and sources cited at the time of writing. Rules may be revised; always check the latest official announcement.

About the author

Jack Lee

Founder & CEO, JoinX

Jack founded JoinX in 2016 and leads a team of 80+ people across Taiwan and Japan. JoinX has delivered 520+ projects for 220+ clients, including 56 mid-to-large systems.

Related articles