• Welcome to Digital Expressway Consulting
  • +91 98180 30135
  • info@digitalExpressway.net

AI in Finance: How DBS Reached S$1 Billion in Measured Value, and 4 Habits Other Banks and Finance Teams Can Copy

← Industry Foresight & Innovation Hotbeds · Banking and finance

AI in Finance: How DBS Reached S$1 Billion in Measured Value, and 4 Habits Other Banks and Finance Teams Can Copy

According to Forrester, DBS Bank in Singapore recorded S$1 billion of economic value from AI in 2025, up from S$370 million at the end of 2023. The bank did not take that number on faith. Forrester reports that DBS compares customer outcomes from AI-powered services against control groups, the way a drug trial compares patients against a placebo.

Now look at a different result. Information Age reports that Commonwealth Bank of Australia announced in July 2025 that 45 customer service roles would be replaced by an AI “voice bot”, then reversed the decision weeks later and apologised to staff. Same technology, opposite outcomes. The difference was not the AI model. It was how carefully each bank tested the claim before acting on it.

That gap matters to anyone working on AI in finance, whether in a bank, an insurer, a corporate finance team or a global capability centre (an offshore hub that runs work for a multinational parent). The tools are now cheap and widely available. The discipline to measure, govern and train is not. This article shows what the leaders actually do, where efforts stall, and how to start.

What you will take away from this article

  • Four habits of the leaders: how DBS, Axis Bank, JPMorgan Chase, Citi and BNY run AI, with sourced results.
  • Five reasons efforts stall: including a real case of job cuts that had to be reversed.
  • Where AI in banking already works: fraud, financial crime, customer service and the finance function.
  • What regulators expect: the Reserve Bank of India’s FREE-AI framework and the EU AI Act, in plain words.
  • A 30-day starter checklist: and the habits that separate trained teams from the rest.

About a 12-minute read.

First, a quick definition. “AI in finance” here means software that learns patterns from data or generates text, and is used to support financial work. It covers older machine learning (for example, models that score a card payment for fraud risk) and newer generative AI (tools such as large language models, or LLMs, which read and write text). It also covers “agents”: AI systems that can carry out a sequence of steps, such as checking a record, drafting a correction and sending it to a person for approval.

4 Moves the Leaders Make

  • Prove value against a control group. According to Forrester, DBS benchmarks customer outcomes from AI against control groups and traces value to three buckets: extra revenue, cost savings and risk avoided. Forrester says this means the S$1 billion DBS reported for 2025 reflects measured benefits, and that it meets a five-year target stated in the bank’s 2022 annual report. What you can borrow: before scaling any AI use, agree with finance how you will measure the “without AI” baseline.

  • Pass every use case through a scorecard, net of AI running costs. Inc42 reports that Axis Bank, India’s third-largest private lender, tests every use case on five measures: accuracy and speed, real adoption by users, operational improvements (such as fewer customer hand-offs), cost reduction and revenue impact. According to Inc42, Axis measures cost and revenue after deducting model, infrastructure and token (usage) costs. What you can borrow: a pilot that cannot show net benefit on a written scorecard should not move to production.

  • Give everyone access, then let peers do the teaching. McKinsey’s interview with JPMorgan Chase’s chief analytics officer reports that nearly 250,000 people have access to the bank’s internal LLM Suite and about half use generative AI tools daily. In the same interview, JPMorgan says the rollout was opt-in, backed by an “AI Made Easy” training programme and “prompt of the week” emails. AI News reports that Citi invited staff to volunteer for its “AI Champions” and “AI Accelerators” programmes, building an internal AI workforce of roughly 4,000 employees. AI News also reports that more than 70% of Citi’s roughly 182,000 staff use firm-approved AI tools in some form. What you can borrow: name one trained volunteer per team before you buy more licences.

  • Treat AI agents like staff, with a named supervisor. Microsoft, which supplies some of the technology BNY uses, reports that BNY runs nearly 140 “digital employees”. According to Microsoft, each has login credentials, an employee number and a human supervisor. Microsoft says one payments agent cut payment validations from five or six minutes to under 30 seconds, with corrections sent for human review. What you can borrow: give every agent an owner, an access list and an approval step before it touches money or customers.

Why Most AI in Finance Efforts Stall

1. Cutting people before proving the tool. Announcing savings before the system works is the fastest way to lose trust. The Commonwealth Bank case below shows the cost.

Case in point: Commonwealth Bank of Australia

What happened: According to Information Age, the bank announced in July 2025 that 45 customer service roles would be replaced by an AI voice bot.

What went wrong: According to Information Age, the Finance Sector Union said call volumes were in fact rising, with overtime offered and team leaders answering some calls. The bank told Information Age its initial assessment “did not adequately consider all relevant business considerations”.

What changed: Information Age reports that the bank reversed the decision in August 2025, apologised and said it was reviewing its internal processes.

The lesson: Measure real demand and real AI performance first. Decide on roles last.

2. Pilots that never reach the profit and loss account. Gartner’s 2025 survey of 183 finance leaders (May to June 2025) found 91% reported only low or moderate impact at first. Gartner also found 25% were unsure how to move from planning to piloting. Inc42 quotes Axis Bank saying AI is not yet a material driver of its profit and loss, and that “the hard work is getting to that first 1%.”

3. Weak data. The Bank of England and Financial Conduct Authority survey of 118 UK firms (published November 2024) found four of the top five perceived AI risks were about data: privacy, quality and security. Gartner’s finance survey also names poor data quality and low data skills as the largest obstacles.

4. Using models you do not fully understand. The same UK survey found about a third of AI uses were built by outside suppliers, up from 17% in 2022. Only 34% of firms said they had a complete understanding of the AI they use, and 46% said partial. Gaps were mostly linked to third-party models.

5. Too little training. BCG’s AI at Work 2026 survey of 11,749 people found only 36% felt they had received sufficient training. Tools without skills produce shallow use.

What the Successful Few Do Differently

What they do What it means in practice The evidence
Measure against a baseline Compare AI-assisted outcomes with a group that did not get AI Forrester reports DBS uses control groups and reached S$1 billion of value in 2025
Count costs honestly Deduct model and infrastructure fees before claiming savings Inc42 reports Axis Bank measures cost and revenue net of model and token costs
Make training social Volunteer champions, prompt libraries, peer badges AI News reports Citi’s volunteer champions programmes and roughly 4,000-person internal AI workforce; McKinsey reports JPMorgan’s peer “prompt of the week” emails
Govern agents like people Named supervisor, credentials, human review of outputs Microsoft reports BNY’s digital employees each have a human supervisor
Set a public value target Commit to a number so progress can be tracked Evident reports RBC is one of only seven of 50 tracked banks with an AI return target, up to CAD 1 billion by 2027

These examples and surveys show what leaders do alongside their results. They show correlation, not proof that each habit caused the gains.

Where AI in Banking Already Works

Financial crime and anti-money laundering (company-reported, on the supplier’s blog). Anti-money laundering (AML) means spotting and reporting money from crime moving through accounts. Writing on Google Cloud’s blog, an HSBC financial crime executive reported that Google’s AML AI found two to four times as much suspicious activity as HSBC’s old rules-based system, with 60% fewer alerts. HSBC compared 12 months after launch with the 12 months before, adjusted for season.

Card fraud (company-reported modelling). Mastercard said in February 2024 that its generative AI fraud tool, in initial modelling, improved fraud detection by 20% on average. Mastercard also reported more than 85% fewer false positives, meaning genuine payments wrongly flagged as fraud.

Customer and employee service (company-reported). Bank of America says its virtual assistant Erica passed 3 billion client interactions in August 2025, with nearly 50 million users since launch. The bank also says its internal version, used by over 90% of employees, cut IT help desk calls by 50%.

The finance function (independent survey). Gartner’s 2025 survey found 59% of finance leaders use AI in their finance function, barely changed from 58% in 2024 (and up from 37% in 2023). The top uses were knowledge management (finding policies and past answers, 49%), accounts payable automation (37%), and error and anomaly detection (34%). For corporate teams weighing AI for finance work, these are the building blocks of faster month-end close and reconciliation.

India view. The Reserve Bank of India (RBI) FREE-AI committee report, summarised by law firm Khaitan & Co, found 20.8% of surveyed entities (regulated firms and fintechs) were deploying AI, and 67% wanted to explore it. MediaNama reports, citing an RBI right-to-information reply, that 23 banks had implemented MuleHunter.AI by December 2025. This is an RBI Innovation Hub tool that spots “mule” accounts used to move fraud proceeds. Inc42 reports that Axis Bank’s internal knowledge search is used by about 40% of its nearly 100,000 staff, with wrong or fallback answers cut below 10%.

How It Works

Most bank and finance-team AI has four parts.

  1. Data and documents. Transaction records, customer files, policies, contracts and invoices. Quality here sets the ceiling for everything else.
  2. Models. Scoring models flag risk; language models read, summarise and draft. Inc42 reports that Axis Bank tests several models rather than relying on one supplier.
  3. Workflow and agents. Software that moves a case through steps, calling the model where it helps and stopping at defined checkpoints.
  4. Controls the regulator expects. Logs, human approval, monitoring and a record of every model in use.

Illustrative workflow (a composite, not a real bank). A mid-sized lender’s operations team handles failed cross-border payments. An agent reads the failure message, pulls the customer record and checks the address and country code against reference data. It drafts a correction and an explanation. A payments officer reviews and approves the fix. The agent then resubmits the payment, logs every step and updates the case.

[DIAGRAM: Failed payment alert → Agent pulls customer record → Agent checks address and country code → Agent drafts correction → Payments officer approves (human approval point) → Agent resubmits and logs → Case closed]

What regulators expect, in plain words. India’s RBI published the report of its FREE-AI committee (Framework for Responsible and Ethical Enablement of Artificial Intelligence) in August 2025; the RBI says it sets out seven guiding principles and 26 recommendations. According to Khaitan & Co’s summary, its principles say firms remain accountable for AI decisions however autonomous the system, should disclose AI use, and should let people override AI decisions. Khaitan & Co also lists recommendations for a board-approved AI policy, an inventory of AI models and use cases, AI incident reporting and audits, and an AI sandbox for safe testing. In Europe, McCann FitzGerald explains that the EU AI Act classes AI used to assess a person’s creditworthiness or set their credit score as “high-risk”. Fraud detection systems are excepted, but anti-money laundering systems are not. According to McCann FitzGerald and Leaseurope, the start date for these high-risk rules has moved from 2 August 2026 to 2 December 2027. Leaseurope says EU lawmakers adopted this change in June 2026, following the Commission’s “Digital Omnibus” proposal of November 2025. McCann FitzGerald says the obligations themselves have not otherwise changed.

Value It Unlocks

The measured gains are real, but concentrated in a few organisations.

  • Bank-wide value: Forrester reports DBS went from S$370 million (2023) to S$1 billion (2025). In McKinsey’s interview, JPMorgan says the gross benefit it credits to AI has grown 30–40% a year, tracked per initiative.
  • Process speed: Microsoft reports BNY’s payments agent cut open payment investigations by nearly 80%, and that client inquiries once taking days now resolve the same day.
  • Industry signals: Evident’s 2025 index of 50 banks reports twice as many banks now disclose total AI returns as a year earlier, and AI headcount at those banks grew more than 25%. NVIDIA, which sells AI hardware, reported in February 2025 that nearly 70% of financial services respondents saw AI lift revenue by 5% or more; the page did not state the sample size.

Gains are uneven. Evident reports the top 10 banks raised their scores by more than double the average bank’s increase. Gartner found 91% of finance leaders saw only low or moderate early impact. Inc42 quotes Axis Bank expecting about 1% profit and loss impact in its 2027 financial year.

Where Claude Fits, and Where It Doesn’t

Claude, the AI assistant made by Anthropic, is a general-purpose language model. In finance it is suited to reading-heavy and writing-heavy work:

  • Summarising long policies, credit memos, contracts and regulatory circulars.
  • Drafting variance commentary, board-pack narratives and customer replies for review.
  • Explaining reconciliation breaks and suggesting likely causes for an analyst to check.
  • Helping teams write and test code for reporting and data checks.
  • Checking a draft against a policy or a checklist and flagging gaps.

Where a person must stay in charge:

  • Credit decisions and anything affecting a customer’s access to money.
  • Filing suspicious activity reports and other regulatory submissions.
  • Posting journal entries, releasing payments and signing off accounts.
  • Any figure going to a regulator, auditor or the board.

Inc42 quotes Axis Bank’s view neatly: “we will not need makers, but the checkers won’t go away.” The principle: let AI do the first draft, and let accountable people make the decision.

How to Start: A 30-Day Checklist

  • [ ] Week 1: Pick one process with clear volumes and pain, such as payment repairs, invoice queries or policy questions.
  • [ ] Week 1: Record the baseline — time per case, error rate and backlog — agreed with finance.
  • [ ] Week 1: Check data rules — what can and cannot leave the bank’s controlled systems.
  • [ ] Week 2: Write a one-page scorecard covering accuracy, adoption, operational gains, cost net of AI fees, and revenue.
  • [ ] Week 2: Name a human approver for every output that touches money or customers.
  • [ ] Week 3: Run a small trial with a closed group, keeping a comparison group working the old way.
  • [ ] Week 3: Log errors and overrides, and add the use case to an AI inventory.
  • [ ] Week 4: Compare results with the baseline and the comparison group.
  • [ ] Week 4: Decide to scale, fix or stop, and write down why for risk and audit teams.

Why Trained Teams Succeed

“Anti-patterns” are habits that look sensible in the moment but reliably cause failure later. In finance they are costly, because errors reach customers, regulators and the accounts.

Short-sighted habit What a trained team does instead
Announces savings or job cuts before the trial Measures demand and AI performance first, then decides on roles
Counts gross time saved Counts net benefit after model, infrastructure and review costs
Pastes customer data into any tool Uses approved tools and keeps personal data on controlled systems
Lets the agent post or pay directly Builds in a named human approver and full logs
Treats the pilot as an IT project Involves risk, compliance and finance from week one

Training changes behaviour. BCG’s AI at Work 2025 survey of 10,635 employees found 79% of those with more than five hours of training used AI regularly, against 67% of those with less. Yet BCG’s 2026 edition found only 36% felt properly trained. The leading banks above all invest heavily here: Microsoft reports BNY has logged 171,000 employee learning hours on AI.

Build the skills to run AI in finance safely and measurably. The CCAO-F prep course from Digital Expressway Academy prepares professionals who run AI work without writing code (consultants, managers, analysts, operations and delivery leads) for Anthropic’s Claude Certified Associate exam, building the habits above: clear scoping, human review and honest measurement. See the course

Bringing AI to your whole team? Talk to us about our Industry Advisory, Innovation and Pilot Build, and Mentoring and Skill Development services.

Sources

All sources were opened and checked in October 2026. Figures described as company-reported come from the organisation itself.

Digital Expressway Insights. Published November 2026; last reviewed October 2026. Not affiliated with or endorsed by Anthropic. Claude is a product of Anthropic. Examples marked “illustrative” are composites created to explain an idea and do not describe a real organisation.