In October 2025, a major consulting firm agreed to refund part of its AUD 440,000 fee to the Australian government. Errors, including references that did not exist, had been found in a report it had produced with the help of generative AI.1 The firm did not say that AI caused the errors, and the department said the recommendations were unchanged. Yet the episode captured a question that many boards are now asking: what is our AI investment actually returning?
The wider evidence is sobering. More than three-quarters of organisations now use AI in at least one business function, yet more than 80% report no tangible impact on enterprise-level EBIT from their use of generative AI.2 By some estimates, more than 80% of AI projects fail.3
Exhibit 1
The adoption to value gap
>3/4
of organisations use AI in at least one business function
Sources: McKinsey Global Survey on AI, March 2025 (1,491 participants, self-reported); RAND, 2024.
Value at scale is possible. DBS reported approximately SGD 1 billion of economic value from data analytics and AI/ML in 2025, spanning more than 430 use cases.4 What separates these outcomes is rarely the technology alone. It is whether data, AI and transformation are run as one agenda.
01The problem
Functional success is not enterprise value
In recent conversations with senior leaders across banking and the public sector, and at major industry events, I have noticed that data, AI and transformation are often discussed in silos. These discussions explore each discipline in depth but rarely address the value of bringing them together. In my view, integrating these three disciplines can create greater value than pursuing each independently.
An AI tool may make a task faster. Whether that creates a better customer outcome depends on the data it receives, the process around it and the decisions people make with its output. For leadership teams, the challenge is to turn that capability into sustained enterprise value. This requires treating data, AI and transformation as a shared business agenda, with clear accountability for outcomes, investment, regulatory compliance, and risk.
Too often, these disciplines pursue different definitions of success. The data team improves availability, the AI team improves model performance, and the transformation team delivers milestones. Each may meet its targets while customers still face delays and inconsistent decisions, and the institution falls short on compliance or expected returns. Leadership must resolve this gap between functional success and enterprise performance.
02Why it happens
Three Disciplines, One Purpose
Data is the information an institution uses to understand and act. It includes customer information, transactions, operational records, external information, and the outputs passed between teams and organisations. Its usefulness depends on accuracy, timeliness, context, and the right to use it. More data does not automatically mean better decisions.
More data does not automatically mean better decisions.
Arjun Chib
AI uses computational models to recognise patterns, make predictions, generate content, or support actions. It has evolved over time, with rapid advances in recent years; the expansion of generative AI has made it accessible to many more people. Its outputs can be useful without being reliable enough for every decision. Leaders need to distinguish between assistance and delegated authority and remain accountable for how both are used.
Transformation changes how work gets done: the process, responsibilities, incentives, technology, and skills needed to deliver a better outcome. Its objective should be to improve service and economics while keeping risk within an explicitly agreed appetite. New capabilities may introduce new risks, even when they reduce existing ones. Leadership must ensure that the organisation has the ability and capacity to recognise, challenge and manage those risks.
An improvement journey begins by challenging steps that no longer serve a purpose or deliver their intended value. Then simplify what remains and redesign the journey end-to-end. Required controls should be retained or redesigned with the appropriate owners, rather than removed in the pursuit of speed. AI can then support the decisions and activities where it demonstrates value. The sequence matters: eliminate unnecessary work, simplify the process and then decide where AI belongs.
The sequence matters: eliminate unnecessary work, simplify the process and then decide where AI belongs.
Arjun Chib
In my corporate experience, several AI experiments did not deliver the expected outcomes. They exposed limitations in the technology and assumptions I needed to revisit. That experience sharpened my judgement about where AI could add value and where it was unsuitable. For leaders, experimentation is valuable when it changes an investment or operating decision, including the decision to stop. The discipline is to define what must be learned before committing more capital.
Exhibit 2
Where AI investments stall between pilot and value
Pilot
1Each function hits its own target
The data, AI and transformation teams meet their goals while customers still face delays.
2AI is added to work that should not exist
Eliminate unnecessary work and simplify before deciding where AI belongs.
3Data is not fit for the use case
Ownership, definitions, provenance and quality thresholds are not set before deployment.
4Outputs are not verified
Fluent drafting does not establish factual reliability.
5Time saved is not redeployed
Released capacity never becomes service, demand served or lower expenditure.
6Perceived speed is mistaken for value
Total delivery time, review and rework go unmeasured.
Enterprise value
03What good looks like: an individual firm's perspective
Lessons from recent public and private sector examples
Results from 2024 to 2026 show how data, AI and changes to the way people work can reinforce one another. Their value is clearest when the operational objective and the evidence are specific.
Case 01
DBSBanking at enterprise scale
Reported result
DBS reported approximately SGD 1 billion of economic value from data analytics and AI/ML in 2025, spanning more than 430 use cases. 4
Evidence caveat
This is the bank’s own reported value measure, not an independently established AI-only profit or net ROI figure.
Leadership lesson
Manage AI as a portfolio of business investments: prioritise value, assign benefit owners and assess realised outcomes rather than count deployments.
Case 02
UK GovernmentAdoption and productive capacity
Trial and result
The UK government provided Microsoft 365 Copilot licences to 20,000 employees in a September–December 2024 trial. Findings published in June 2025 showed that survey respondents estimated average savings of 26 minutes a day. 5
Evidence caveat
These were self-reported savings, not independently measured cash benefits. The report identified training and change management as important to adoption, alongside limitations on nuanced and complex work.
Leadership lesson
Fund adoption alongside technology: provide training, establish appropriate data access and make business owners accountable for converting released time into better service or productive capacity. Distributing licences is only the start of transformation, not the end.
Case 03
A major consulting firmVerification and accountability
What happened
In October 2025, a major consulting firm agreed to refund part of its AUD 440,000 fee for a report commissioned by the Australian government after incorrect references were identified. 1
Evidence caveat
The revised report disclosed generative AI use; the firm did not confirm that AI caused the errors. The department said the recommendations were unchanged and published further corrections in February 2026. 6
Leadership lesson
Accountability for quality remains with the institution, including when work is outsourced or supported by AI. Verify material citations against original sources, retain an evidence trail and give a qualified reviewer the time and authority to withhold sign-off. Fluent drafting does not establish factual reliability.
Case 04
METRMeasured developer productivity
Measured result
METR’s early 2025 randomised trial involved 16 experienced developers and 246 tasks in familiar open-source projects. Access to AI increased completion time by 19%. 7
Evidence caveat
This was a specific setting and generation of tools, not a universal verdict on AI-assisted coding.
AI increased completion time in this early 2025 trial
Figure 1. Normalised task completion time from METR’s February–June 2025 trial. Without AI = 100; with AI available = 119. Reported confidence interval for the increase: 2%–39%. The 2026 follow-up below is essential context. Sources 78.
2026 update
In February 2026, METR suggested that newer tools were likely to offer greater benefits, but selection effects made its follow-up data unreliable for estimating the current productivity gain. 8
Leadership lesson
Use evidence from your own operating environment, including total delivery time, review and rework. Reassess the investment as tools and workflows change; perceived speed is not a sufficient basis for scaling.
Fluent drafting does not establish factual reliability.
Arjun Chib04The framework
How to turn the combination into durable value
Scaling creates more dependencies, handoffs and competing priorities. Small organisations may align more quickly but can lack controls and depth of expertise. Large institutions can integrate successfully when decision rights and incentives support the same outcome. The leadership task is to create that alignment through shared priorities, clear authority and a consistent measure of value.
I would make six disciplines part of the leadership agenda, brought together in ASCEND: Agreed outcomes, Single accountable owner, Credible data, Explicit risk controls, Net realised value and Durable adoption. Together, they connect ambition with operating accountability.
Exhibit 3
ASCEND: turning data, AI and transformation into sustained value
One shared outcome. Clear ownership. Continuous accountability.
A
A Agreed outcomes
Choose a customer or operational problem before choosing the technology. Establish a baseline for cost, turnaround time, quality and failure rates, then agree the improvement that would justify investment.
Compare AI with simpler alternatives, including process changes and rule-based automation. Set the criteria for scaling, redesigning or stopping before the pilot begins.
S
S Single accountable owner
Build a team spanning operations, data, technology, AI, risk and frontline users. Give the business owner accountability for the end-to-end outcome and realised benefits, along with the authority to resolve delivery trade-offs. Establish clear escalation routes. Align functional objectives with that outcome. Preserve independent risk challenge; shared delivery does not require everyone to report to the same leader.
C
C Credible data
Prioritise the data needed for the selected use case. Establish ownership, definitions, provenance, permitted use and quality thresholds. Make data readiness a condition of deployment, with explicit acceptance of any residual limitations. Validate AI-generated outputs before reusing them as inputs; otherwise, an initial mistake can circulate as apparently authoritative evidence.
E
E Explicit risk controls
Translate risk appetite into operating boundaries: where AI may assist, where it may act and where it must abstain or escalate. For critical decisions, ensure human reviewers have the expertise, time and authority to challenge or override the output. Test performance across unusual cases, different customer groups and changing conditions, with particular attention to customers who may face greater risk of harm. For generative AI, verify answers against approved sources and test for unsupported claims, data leakage and malicious instructions. Restrict action permissions, require approval for consequential steps and assign clear authority to pause the system and activate a fallback.
N
N Net realised value
Agree with finance how benefits will be measured and validated. Assess incremental contribution, realised operating savings and any separately evidenced loss reduction against the full costs of implementation, data preparation, integration, model usage, oversight, training, ongoing maintenance, without double counting benefits. Time saved becomes financial value only when capacity is productively redeployed, demand is served or expenditure is actually reduced. Review customer outcomes and risk alongside the financial return.
D
D Durable adoption
Treat workforce readiness as a leadership responsibility. Train employees to use AI and challenge its outputs, involve them in redesigning the process and provide a way to raise concerns. Scale in stages after meeting agreed service, quality and risk thresholds. Assign ongoing ownership for monitoring model drift, overrides, complaints, incidents and cost per successful outcome. Reassess when the model, data or workflow changes; deployment is the beginning of operating accountability.
Time saved becomes financial value only when capacity is productively redeployed, demand is served or expenditure is actually reduced.
Arjun Chib05Immediate task
What leaders should do immediately
1
AChoose one problem. Pick a customer or operational problem and set its baseline for cost, turnaround time, quality and failure rates.
2
SName one owner. Give a single business owner accountability for the end-to-end outcome and realised benefits.
3
EDraw the boundaries. Decide where AI may assist, where it may act and where it must abstain or escalate.
4
NAgree the measure. Agree with finance how benefits will be measured and validated against the full costs.
5
ASet the exit. Set the criteria for scaling, redesigning or stopping before the pilot begins.
The outcome that matters
For institutions, success means sustainable value and clear accountability. For customers, it means reliable, accessible service, fair treatment, protection of their information and access to human help when needed. Regulators have different mandates, but institutions should be able to demonstrate how their systems operate, how risks are controlled and how people can obtain redress.
The power of combining data, AI and transformation depends on leaders connecting ambition with operating discipline. Define the outcome, simplify the journey, invest where evidence supports value and retain meaningful human oversight. Keep learning and be prepared to change course. The leadership test is whether the institution delivers better outcomes, earns trust and can sustain both as it scales.
Trust sets the rhythm for Data & AI
Turning data, AI and transformation into sustained value.
Reference numbers correspond to the article. Dates refer to the underlying activity unless a publication or update date is specified.
1
The Guardian, 6 October 2025. Report on a consulting firm's partial refund to the Australian Department of Employment and Workplace Relations for a AUD 440,000 report that contained errors, after it disclosed using generative AI.The Guardian report
2
McKinsey & Company. The state of AI: How organizations are rewiring to capture value. Published 12 March 2025; survey fieldwork 16 to 31 July 2024, 1,491 participants. More than 80% of respondents say their organisations are not seeing a tangible impact on enterprise-level EBIT from gen AI; 17% attribute 5% or more of EBIT to gen AI.McKinsey State of AI 2025
3
RAND Corporation. The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. 2024. States that "by some estimates, more than 80 percent of AI projects fail"; RAND cites this estimate rather than measuring it.RAND report
4
DBS. Annual Report 2025, Letter from Chairman and CEO. More than 430 use cases delivering economic value of approximately SGD 1 billion in 2025.DBS Annual Report 2025
5
UK Government Digital Service. Microsoft 365 Copilot Experiment Cross Government Findings Report. Published 2 June 2025; trial September to December 2024. The 26-minute daily saving was estimated from self-reported survey responses.UK Government Copilot trial findings
6
Australian Department of Employment and Workplace Relations. Targeted Compliance Framework Assurance Review, final report page, recording the corrected versions including further corrections on 3 February 2026.DEWR report and corrections
7
METR. Measuring the Impact of Early 2025 AI on Experienced Open Source Developer Productivity. 10 July 2025; tasks undertaken February to June 2025. Sixteen developers, 246 tasks, 19% longer completion time with AI available.METR early 2025 study
8
METR. We are Changing our Developer Productivity Experiment Design. 24 February 2026. Reports selection effects in the follow-up; newer tools likely offer more benefit, but a reliable current effect cannot be inferred.METR February 2026 update
9
Monetary Authority of Singapore. Guidelines on Artificial Intelligence Risk Management. Issued 7 October 2026; effective 7 October 2027.MAS media release | Guidelines (PDF)