Imagine walking into a finance meeting without a static spreadsheet. Instead, a “live finance cockpit” presents comparative analyses for the business’ toughest questions, each linked to evidence.
That is the future Stephen Lo, Chief Financial Officer of Prenetics, envisages: digital colleagues working around the clock to monitor cash, investigate variances and test management’s assumptions. But they will not have free rein. Payments, record changes, disclosures and consequential judgements will still require authorized human approval and appropriate controls.

This balance of ambition and restraint sits at the heart of the artificial intelligence (AI)-focused CFO’s challenge. The opportunity is not simply to automate finance, but to reshape how businesses make decisions and allocate capital, without surrendering human accountability.
For now, the gap between ambition and execution remains wide. IBM’s 2026 CFO Study found that 62 percent of finance leaders have taken on enterprise technology or AI strategy responsibilities, yet only 6 percent describe their finance function as transformation-ready. The test now is not whether AI can impress in a demonstration. It is whether people can trust it in everyday work, and see a justified return on investment.
Beyond the chatbot
At global consumer health sciences company Prenetics, Lo says AI’s greatest impact has been in strategy, marketing and real-time analytics. Selling into more than 50 countries means monitoring performance across time zones, spotting signals and deciding what to change. “AI is not a productivity tool we bolt on,” he says. “It changes the operating tempo of the company.”
For Mandy Law, CFO of Microsoft Hong Kong, the true distinction is whether AI changes the work itself, rather than simply helping an individual complete an existing task faster. “The transition happens when AI stops sitting beside the work and becomes embedded in how the work itself gets done,” she says. Achieving that requires organizations to redesign workflows and encourage learning, while keeping people responsible for judgement, accountability and business outcomes, she notes.
Bonnie Chan, APAC Generative AI and Integrity Leader at IBM, sees a similar shift across Asia Pacific: a move from isolated pilots to disciplined execution, with clear priorities, governance and visible leadership commitment. “The shift is not primarily a technology shift. It is a mindset and governance shift,” she says.
And Dr William Chen, Head of Finance, APAC, at Schurter Group, an electronic component manufacturer, agrees. For him, technology itself matters less than connecting AI to a business process, reliable data and a measurable outcome.
For Raymond Tam, CFO of biotech company Adagene, the distinction between experimentation and execution is especially stark. In drug development, a model that performs well in isolation has little value unless its findings can be validated, protected and eventually commercialized.
“Finance leaders must navigate harsh commercial realities: nearly half of all new product launches fail to meet initial financial forecasts. Therefore, AI creates premium enterprise value only when it moves past technical novelty to deliver clinically validated, patentable assets with a predictable commercial runway,” he says.
To turn experimentation into execution, Lo describes what makes that transition possible as an “AI harness”: the combination of context, data, tools, permissions, feedback, evaluation and oversight. An agent that executes multiple steps is only one component.
“The question is no longer, ‘Can a chatbot answer this?’” he says. “It is, ‘Can we redesign this workflow so AI prepares the work, a qualified person applies judgement, and the team can focus on a higher-value problem?’”
For CFOs, the investment question shifts from whether a tool can perform a task to whether it works reliably within a process, subject to human oversight.
Value depends on the business
According to experts, a worthwhile result looks different in every business. Lo tracks revenue per employee, time-to-decision, customer-acquisition economics, forecasting quality, errors and management time reallocated for higher-value work. But he questions assigning a return to every AI interaction. “This is an area where I am still learning,” he says. “The return from AI is immense, but I do not think it is useful to pretend that every prompt or token has its own clean AI return-on-investment calculation.”
With development cycles spanning years, Tam takes a longer view. Adagene assesses AI initiatives through risk-adjusted net present value and option value. “We measure hard savings from shrinking upfront physical testing requirements from thousands of samples to a small number per project, preventing severe downstream sunk capital. True ROI is realized when the platform acts as a repeatable asset factory, securing global corporate partnerships that bring in immediate upfront cash and milestone payments to protect company liquidity.”
For Dr Chen, manufacturing outcomes might mean better forecasting, fewer errors, faster reporting or greater visibility across markets. The measure must match the problem.
Across Microsoft Finance, Law describes an AI-enabled experience that brings financial reporting and performance data together, using Copilot to generate an initial layer of insights and commentary on metrics, trends and variances. “An AI agent can also analyze forecast variances using the financial materials provided to it. This means we can get to the ‘so what?’ earlier: where does the business need support? Which assumptions should we challenge? What action should we consider?” she says.
The intended benefit is not simply a faster first draft explaining results, but more time to interpret them and advise the business
Start with the foundations
The CFO’s task is to prevent activity, novelty or enthusiasm from standing in for value. This is where Dr Chen places data at the centre of AI credibility. “Data is the lifeblood of AI, and the CFO is the architect who ensures that data is clean, unified, and legally secure,” he says.
Poor data quality, he argues, is the most immediate threat to trust. “If the underlying data are fragmented, inconsistent, or inaccurate, even a strong AI model will produce outputs businesses do not trust,” he says. Weak governance and unauditable black-box decision-making compound the problem, he adds.
Law describes Microsoft Finance’s own development as a progression built over more than a decade, rather than a transformation that began with generative AI. The function moved data to the cloud, built a trusted, unified data platform, strengthened business intelligence and added predictive and automation capabilities before reaching today’s Copilot and agent scenarios.
That history matters to the investment case. Law notes that the costs extend beyond technology to skills development, change management, secure governance and workflow redesign.
“Data is the lifeblood of AI, and the CFO is the architect who ensures that data is clean, unified, and legally secure.”
Tam sees a particular version of that challenge in biotech, where data is often heterogeneous, unstructured and scattered across research, clinical and commercial systems. Preparing it for machine learning can require substantial investment. In his assessment, proprietary, high-quality information is the real competitive asset.
Lo makes a similar point from an operating perspective. At Prenetics, AI is only as useful as the context it receives: market data, customer analytics, internal performance information and the permissions that determine what it may access.
Chan applies three tests before considering a use case ready: whether its underlying data is clean, available and governable; whether a measurable outcome can be defined before work starts; and whether it supports a stated strategic priority. “If the answer to any of these is no, the use case is not ready,” she says.
Governance belongs inside the workflow
For Dr Chen, AI governance should not become a parallel bureaucracy detached from existing responsibilities. “CFOs can build AI governance into existing control frameworks by treating AI as an extension of finance, risk, and internal control rather than as a separate programme,” he says. “That approach keeps AI governance close to the work CFOs already own. It also makes it easier to define where AI changes the control environment and where it simply reinforces the framework.”
Each material use case, he says, needs a business owner, a control owner and an independent reviewer. Its purpose, scope, approved data and escalation arrangements should be documented, with monitoring for errors, exceptions and changing performance. This makes governance a practical operating discipline rather than a statement of intent.
As CFO of a Nasdaq-listed company, Tam applies a strict standard to AI applications that could affect financial forecasting or market disclosures. “If we cannot explain it to our auditors, we do not deploy it,” he says.
At Prenetics, Lo says employees operate within enterprise platforms with data controls, rather than relying on unapproved personal accounts. “Governance also has to be practical. The level of scrutiny should increase with the sensitivity of the data, the impact of a possible error and the consequence of the decision,” he says. Material financial, regulatory and external communications remain subject to human review and established approval processes. “In health sciences, the threshold is higher: AI-enabled content is not medical advice, diagnosis or treatment,” Lo adds.
Law offers an example of how boundaries can be built into the work itself. Across Microsoft Finance, an AI risk management agent guides employees through compliance questions, summarizes the information they provide and draws on specified internal compliance sources. If those approved sources cannot resolve a question, the agent directs the employee to appropriate specialists.
The boundary is part of the design: assistance within an approved information base, followed by escalation when that information is insufficient. “The goal is not to create more layers of approval. It is to build controls that give people confidence in the outputs they act upon,” she says.
For finance professionals, Law adds, that still means validating numbers, applying business context and challenging assumptions. Access permissions, protection of sensitive information and maintaining oversight of AI and agent activity must remain in place.
Chan takes a similar view, and notes that accountability, traceability, cybersecurity and compliance must be designed in from the beginning. “Governance is not a constraint on AI – it is the enabler of sustainable AI,” she says.

Know when to stop
Execution also means accepting that some initiatives should not continue. Yet Lo argues that the CFO should not select every use case from the top, and that people closest to the work often see the highest-value opportunities first. “The CFO’s job is not to pretend to know every answer. It is to create the environment: the right tools, the right data, permission to experiment, clear accountability, and enough AI fluency at the leadership level to know when a team has found a viable solution,” he says. “Then you apply discipline. A good use case needs a named owner, a defined problem, an expected outcome and a point at which we decide to scale, redesign or stop.”
When something underperforms, he starts with the work itself. “Was the task well defined? Was the data reliable and permitted for that use? Did we redesign the workflow, or simply put a chatbot on top of a broken process? Can the output be checked?” says Lo. “The decision is ultimately made by a human. You cannot blame the system.”
Tam favours phase-gated capital allocation: releasing funding in stages, tied to operational validation, with predetermined “killer criteria” for stopping unsuccessful project variations.
In biotech, he warns, a product that looks convincing computationally may fail physical testing or prove unsuitable for commercial manufacturing. AI does not remove those constraints, and management should not assume it can bypass the need for human testing.
Dr Chen sees project failure as a governance signal as much as a technology problem. If a use case cannot be measured, monitored or assigned to an owner, it is unlikely to become a dependable part of the business.
Chan’s experience points to similar causes: unclear ownership, weak data, insufficient governance, skills gaps and unrealistic expectations. “In my experience, the root cause is almost never the technology,” she says.
Capabilities that matter most
Lo identifies talent as the biggest constraint on adoption, but his definition goes beyond recruiting technical specialists.
His finance team, he says, works with agentic AI and is comfortable coding in a command-line environment, despite not coming from formal computer science backgrounds. The important qualities are curiosity, practical experimentation and a willingness to build.
Effective adoption requires more than asking a chatbot questions. Finance professionals need to break business problems into tasks, provide context, select appropriate tools, test results and trace conclusions back to source material.
Dr Chen highlights uneven readiness across APAC teams. Some may be equipped to move quickly, while others require structured learning, simulation and mentoring. Regional deployment therefore cannot assume a uniform starting point.
Chan, meanwhile, emphasizes AI literacy, data stewardship and governance fluency. Finance teams, she says, need enough understanding to ask the right questions, interpret AI outputs confidently and explain their use to boards and regulators.
Capability, however, is only half the challenge. Employees also need a reason to share their knowledge and change how they work. Law describes an approach to developing that confidence. “We encourage people to start with a real problem from their own work, then build and test solutions hands-on,” she says.
The company encourages employees to share their experiences during learning sessions. Promising ideas can then be taken further, while builders and early adopters receive recognition. The purpose is not only to teach employees to use a tool, but to help them rethink how work gets done.
Lo warns against presenting AI as a substitute for human expertise. “You cannot expect people to embrace a technology they believe will replace them,” he says. He argues that participation needs a genuine upside: better work, greater autonomy, creativity and opportunities to grow.
A more demanding role
“The hard part everywhere is turning that enthusiasm into a change in how the organization actually works.”
APAC is now advancing quickly with AI adoption and, in some dimensions, faster than Europe and North America, notes Chan. She sees Hong Kong’s regulatory frameworks as a potential advantage when organizations engage with them effectively. In her assessment, leadership commitment and clear regulatory dialogue support progress, while talent scarcity and governance uncertainty slow it.
On the other hand, Lo resists treating the adoption challenge as uniquely local. “Every market has companies talking about AI, running pilots and showing demos,” he says. “The hard part everywhere is turning that enthusiasm into a change in how the organization actually works.”
Dr Chen’s regional perspective adds another complication: differences in infrastructure and digital maturity. “While cloud-ready locations can scale AI rapidly, markets dependent on legacy platforms face fragmented rollouts and inconsistent user experiences,” he says.
Experts propose different starting points. Chan favours sequencing, beginning with lower-data-risk opportunities and moving towards production rather than another pilot. Dr Chen believes finance leaders should focus on governance, strategy and performance measurement, while Tam stresses staged funding and the discipline to exit. Lo emphasizes hands-on leadership and opportunities identified by people closest to the work.
Law adds the importance of a learning loop: “capturing what works, governing it properly, and reusing it across teams and workflows. Over time, those learnings become part of the organization’s institutional knowledge and contribute to AI becoming a business asset rather than a series of disconnected pilots.”
Together, they describe a CFO role that is both more enabling and more demanding. Finance must create room for experimentation without allowing it to become an exemption from accountability.
AI-focused CFOs are therefore not simply the finance leaders with the most advanced tools. They are the leaders who can connect those tools to a business problem, organize the conditions for reliable use and distinguish genuine progress from additional output. “AI does not make the CFO less necessary. It makes the CFO’s judgement more valuable,” says Lo.
In recent years, 62% of CFOs have taken on new enterprise technology or artificial intelligence strategy leadership responsibilities, according to IBM’s 2026 CFO Study.














