The AI-Powered CFO

Balancing Human Insight with Machine Intelligence in 2025

In 2025, the role of the modern CFO stands at a pivotal crossroads as artificial intelligence (AI) reshapes the landscape of financial leadership. While CFOs increasingly recognize AI’s transformative potential, many grapple with practical implementation questions: How do we start? Where will it create real value? What are the risks? The focus shifts from whether to adopt AI to determining the right approach.

For forward-thinking CFOs, this transformation presents both opportunities and uncertainties. The promise of AI’s analytical power and real-time data processing capabilities is clear, but the path to successful implementation remains complex. Success requires not just understanding AI’s technical possibilities, but also identifying which use cases will deliver meaningful ROI, managing implementation risks, and ensuring AI initiatives align with broader business strategy. The challenge lies in moving from AI awareness to actionable plans while maintaining the strategic vision and leadership qualities that have always defined exceptional financial leadership.

However, before diving into technical implementations, it’s crucial to address a fundamental question that every organization must consider.

AI or No AI

First, it’s important to recognize that in many cases, you might not need AI—at least not yet—because you can unlock immediate value simply by providing users with higher-quality and more timely data. At the same time, it should be noted that without timely, high-quality data, proceeding with AI implementation isn’t feasible.

Below we will examine the available types of AI, their limitations, and from a practical viewpoint how AI will be incorporated into your business systems. We will look at where it is currently being used, and legislative changes that you need to be aware of before its use. Lastly, and more holistically, how it fits into the bigger picture of business operations.

With this foundational understanding of when AI might be appropriate, let’s examine the specific types of AI that can be deployed within and across finance, operations, and HR.

Types of AI for Finance, Operations and HR

Different types of AI models will be integrated simultaneously into your business processes. A single operation may involve multiple algorithms working either sequentially or in parallel to achieve your desired outcomes. These models generally fall into two categories: pre-built solutions and custom models. Pre-built solutions, like document recognition, can be readily adapted to your business. Custom models are developed specifically for your unique needs.

At a high level, AI models can be classified into three broad categories: i) Bottom Up AI; ii) Top Down AI; and iii) GenAI.

Bottom Up AI

Large Language Models (LLMs), such as GPT, are the most widely known form of bottom-up AI. Users encounter their outputs in everyday applications, making them highly recognizable. These models learn by consuming vast amounts of data from the ground up, processing and identifying patterns without pre-programmed rules.

Models start with raw data and learn patterns autonomously, making decisions without predefined rules. This approach, fundamental to machine learning and neural networks, relies on training data to learn from examples rather than following explicit instructions.

The resulting systems can develop sophisticated behaviours that emerge naturally from data patterns, making them highly adaptable to new situations. This flexibility allows them to handle diverse tasks more effectively than traditional rule-based systems, as they learn from experience rather than following rigid guidelines.

Examples: Document recognition and routing; employee face recognition for attendance, noting that in many jurisdictions that employees must have access to alternative sign-in / sign –out processes; purchase contract document recognition to identify terms, obligations, and risks using multiple data points; prioritised fraud detection by analysing data outliers.

Top Down AI

This is used by those who prefer applying previously acquired knowledge to focused problem solving. Top-down AI’ = ‘Data-efficient AI’, because operationally it needs less data. The top-down approach involves starting with a high-level understanding of the problem and then breaking it down into smaller, manageable components.

This method relies on predefined rules and knowledge to guide the AI’s behaviour. Top-down AI can utilize rule-based systems and probabilistic formulae. These combined dictate the AI’s actions. This can include expert systems that apply logical reasoning to solve specific problems.

The learning process is structured and guided by human-defined parameters, making it easier to control and predict the AI’s behaviour. In other words decision points in a process can be executed by the systems as if it was human, specifically noting the limited scope of applied “thought” used in this action.

Example: Any process automation step, where decisions are made by the system based on algorithms that have been trained for this purpose. For example, consistently prioritising leave applications based on your specific corporate criteria etc; or in a consulting organisation to pinpoint staff who have not taken their holidays, yet are over-billing, so that actions can be taken by managers to avoid their burn out and churn; orchestrating purchase order approval workflows based on routing criteria (value, type of purchase etc) ; routing payment approval workflows to avoid Business Process Compromise (BPC) – this is where threat actors aim to divert your cash to bank accounts under their control; budget variance analysis to your materiality levels ranked by entity, segment, by consolidated numbers etc; automated reconciliations of various types; regulatory reporting checks including banking covenants; risk management ie FX, credit exposure, identifying duplicated payments, identifying material assets that are not fully insured to market value (insurance to value – ITV) etc; suggested optimised resource allocations based on current conditions and past metrics etc. Opportunities are endless.

Hybrid Approaches

From a practical perspective, the most effective AI implementations will often combine both top-down and bottom-up methodologies, effectively leveraging the strengths of each approach. A hybrid model typically uses top-down frameworks to establish clear boundaries and objectives, while employing bottom-up learning techniques to adapt and refine solutions based on real-world data. This balanced approach helps optimize outputs by combining the reliability and focus of structured rules with the adaptability of data-driven learning.

Example: Procure to payment cycles that leverage LLM’s for document recognition, as well as top down algorithms for ensuring that the correct tiered discount overrides are correctly in place.

Generative AI

Generative AI, often abbreviated as GenAI, is a subset of artificial intelligence that focuses on creating new content across various media, including text, image, audio, and video types. This technology utilizes generative models to learn patterns from existing data and produce new outputs based on user prompts or speech inputs.

GenAI can generate a wide range of content, such as stories, articles, images, music, and even computer program code, and all by responding to natural language prompts provided by users. Modern GenAI systems have user-friendly interfaces that enable all types of user to interact with the AI using simple language, making it very accessible to a broader audience.

Of course these same technologies can be used and modified by threat actors to cause havoc; for example Business Process Compromise (BPC), by using these tools to exactly mimic your voice and appearance within video feeds to divert your funds to accounts under their control. These can be sophisticated in nature and can include you on a call with your bosses, with each person, other than you, having voice and visual characteristics as per the “original’ real person.

Examples: 1) Standalone GenAI models: one can ask any question about anything; 2) Vendor GenAI integrations: these can do the same as “1)”, but with “controlled” execution across and within their applications. So, within vendor Human Resource solutions, employees and managers can probe and understand policies;  HR managers can identify operational anomalies – for example in retail operations where different brands, although operating under the same HR policies, have interpreted the specifics of policies differently**; with Sanctions databases for end user reference checks; to search product databases for best customer  specifications matches etc.

** This is an example of shrinking demarcation lines between operations in retail, where for example, latest technologies support the utilisation of staff across corporate brands, not just within them.

While these AI categories offer compelling possibilities, implementing them presents several challenges that organizations must carefully navigate.

Problematic Areas Working with AI

Many corporations struggle in their early AI implementation because of a key disconnect: While AI benefits are easy to discuss in broad terms, the actual execution requires precise, step-by-step data management. This mismatch often surfaces when companies move from boardroom discussions to practical implementation. The high-level strategic vision must translate into detailed data preparation, model selection, and technical processes. This shift demands not just technical expertise, but also a clear understanding of how data flows through the organization. Success requires bridging this gap between strategic thinking and tactical execution, particularly in data management and technical implementation.

Data Selection – Getting it Right

The first challenge in AI implementation is selecting the right data sets – those that directly influence your desired outputs rather than those that only have indirect effects. While comprehensive data coverage may seem ideal on the surface, each additional data point increases model complexity and training requirements. This translates directly into higher training costs and greater infrastructure needs. A practical approach requires balancing the scope of data against operational constraints and desired outcomes. In essence, this is why top down AI is attractive to corporates for many AI projects.

Processing Mechanisms

AI systems process data through two main algorithmic approaches or combinations thereof:

1. Deterministic Algorithms: These use hard rules to find exact matches to your criteria. However, while providing clear-cut results, they’re far less scalable and can produce false positives.
2. Probabilistic Algorithms: Work on likelihood matches to your requirements, with results given a % probability. These are more scalable and typically produce fewer false positives.

Understanding the Shape of Algorithmic Outputs

For any AI model to achieve its specific output result (accuracy and relevance for the actual task in hand), think of it as placing your desired data points within a translucent 3D ball containing many other “similar” data points. A combination of your deterministic and probabilistic algorithm(s) effectively reshapes this 3D ball to isolate and capture the most relevant data for your needs, similar to carving a sculpture from a block of marble to reveal the desired figure.

Conversely, algorithms can be utilized to identify data outliers that require further investigation – ie those with certain characteristics that reside outside of the 3D sphere, thereby enabling risk management and compliance to be integrated into processes. For example, looking at travel and expenses (T&E) for outside corporate policy expenses, duplicate claims etc.

Understanding the Limitations of AI

Data Availability

• AI will not work unless data sets are available to your algorithms for training. This means that you will have three scenarios
  • Vendors algorithms where algorithms are already integrated within their applications.
  • Packaged algorithms from Google or other providers that you incorporate into your own processes (so you save on the design costs etc).
  • Self-Developed algorithms, or those developed for you, to solve your identified key business challenges.

Data Accuracy – The Need for Verification

• One Key Measure of AI is the % accuracy of results, especially for transactional processing. Whilst higher percentage accuracy is increasingly possible, one needs to accept that you will actually need to identify which data sets are wrong.
• For identified root causes of inaccuracy, steps can be taken to improve the situation by tweaking the algorithm or introducing refined rules to remove / adjust anomalies
• As a result, you should view AI use as augmentation, not revolution. It is balance between automated processes and human judgment.

Data Privacy

• AI use is becoming more pervasive and it is therefore important to continually focus on privacy controls. As with any process you will need to understand where data is being stored and for how long, where it is being computed + processed, particularly in the context of cross border data transfers.
• With any AI provider, be proactive in understanding whether your data is being used to further train their AI model with your data, and whether you have a specific issue with this. This is particularly relevant for free services.
• Avoid scenarios of Shadow IT, so that all AI deployments are executed within your company policies for privacy and cybersecurity.

Data Outputs from AI Models

Beyond technical setup, organizations need to grasp how AI systems actually work with data to generate meaningful results. Understanding this process means following how raw information transforms into useful insights. Rather than viewing AI as a black box, companies must work to recognize it as a series of logical steps that filter, analyse, and interpret data. This understanding helps teams make better decisions about data quality, model selection, and how to apply AI outputs in real business situations. When organizations see AI as a transparent process rather than just an end result, they can better validate outputs and adjust their approach based on actual performance.

Process Transparency in Smaller Models

• Latest process technologies enable repeatable and auditable processes to be designed, with each specific step known and transparent to the process owner, but there comes a point in complex algorithmic models where outputs are not fully auditable, so it is imperative have checks and balances in place.

 “Black Box” Computations

• With bottom-up approaches and with GenAI, process outcomes are not always understandable. This is because AI models often function as “black boxes” – their decision-making processes might be difficult to trace and explain, to both the algorithmic creators and third party auditors. Sounds strange, especially when algorithms in theory produce finite outcomes, so put this down to the size and complexity of models.

Hallucinations

• Models can also produce outputs called “hallucinations” – outputs that appear totally plausible and convincing, but which are factually incorrect.

AI Legislation

• Latest AI legislation, being introduced or planned in various parts of the world, is moving slowly to introduce checkpoints. It is focusing corporate behaviour on AI use and transparency, especially when results might directly impact individuals eg health, employment, finance etc.
• Latest technologies more easily allow for corporates to test their latest algorithms on older retained data sets to evaluate how algorithmic tweaks have impacted data flows.

These technical considerations inevitably intersect with existing organizational processes and limitations.

Current Process Limitations

• Many companies are familiar with Generative AI and its potential benefits. However, they may not yet fully grasp the value of top-down and bottom-up AI methods to their operations, often perceiving them as less immediately beneficial.

A deeper investigation reveals that many users and managers in fact still rely heavily on month-end processes for reporting and problem identification. They have not yet shifted to actionable, contextual reporting throughout the month (where it makes sense). For companies, in this position, it will likely also provoke discussions as to whether AI is needed at all, as the driver for change will not be as apparent to them.

Understanding these current limitations is essential as we consider how to integrate AI algorithms into existing processes.

How Do Algorithms Fit into Current Processes

Today’s business systems are a combination of i) individually designed end-to-end processes + ii) proven applications which are also the trusted systems of record. End to end process design goes from data collection, thru all required ultra-granular transformations to produce actionable contextual reports / workflows @anytime @anywhere (with or without AI) + Simulations.

Processes are more likely to cross multiple applications to release greater value than is possible today.  Importantly, recognise that compliance can be built-in ground-up to ensure that specific data can only be accessed by approved authorised users.

Earlier we touched on human judgment, and it should be noted that in any transactional process design you can build in eye-ball reviews from a specific knowledgeable person(s) to check for obvious anomalies prior to data submission / onward processing; for example in month end reporting or budget submissions to have that final sanity check.

As a final comment here, it is important to recognise that change management is going to be an important part of any process re-engineering, recognising that existing boundaries within and across domain areas are shrinking.

With the aforementioned in mind, organizations can begin taking practical steps toward AI adoption.

Encouraging AI Experimentation

Before progressing, all recognise that AI is at a very early phase. There is certainly a major chasm to cross to make it work. However, many are recognising that there are significant opportunities and more importantly are beginning to appreciate that there are obstacles to overcome. These obstacles are not purely down to AI deployment alone, but more holistically to the broader changes required for digital enablement. The issues associated with them are interconnected: i) what processes should be prioritised first to release value; ii) what level of automation and depth of functionality is practical to implement to meet our needs; iii) how aggressive should one be in forcing change within and across domain areas for process optimisation; and iv) what skills are going to be needed to effect the identified change, and do they exist in the organisation.

Getting from the starting point to introduce AI is going to need additional skills and expertise. It will not be easy, as it is very much a “chicken and egg “scenario with structural HR limitations and opportunities for change coming together more forcefully than ever before.

Corporates should consider:-

• Employee Empowerment: Specifically stating where employees can and cannot use Gen AI. This should take into account what AI tools are already available eg via vendor integration or access to closed GenAI models (no access should be granted to public models as a default). Managed management updates (monthly, weekly etc) should be done via Employee Self Service processes from HR to keep everyone updated as to progress, and a mechanism put in place to handle employee concerns and feedbacks.
• Experimentation: Encouraging skunk projects and participation from employees to explore options, ensuring that privacy is positioned as a core component.
• Skills Enhancement: Providing opportunities for reskilling and upskilling. Ensuring that management have a short, medium and long term plan as to how their business systems will look. This will be increasingly relevant to drive long term digital enablement as i) processes become richer in functionality; ii) as process velocity increases and iii) as domain areas move closer together.
• Technical Awareness: Ensure that staff are aware of the capabilities of latest technologies, as well as their limitations.
• Holistic Awareness: Ensure that staff can see how their works fits into the bigger picture, with a specific emphasis on adjacent business domains.            

As organizations implement these experimental approaches and learning initiatives, they will inevitably discover that traditional organizational structures may no longer serve their evolving needs. This recognition of structural limitations, and the opportunities they present, becomes the next critical consideration in the AI transformation journey.

Structural Limitations and Opportunities

Organisational structures will be challenged. This is due to increased process velocity and the shortening of distances between domain areas. The difficulties will force organisations to tackle the integration of modern technologies into their HR organisation.

This will likely see them develop power users or specialist digital teams that have i) tech skills; ii) a broader more holistic picture of operations that can adjust / adapt to proactive employee feedback; iii) end to end process knowledge; as well as iv) mitigations against anxious employees.

These structural considerations complete the picture of AI transformation in financial leadership, bringing us back to our central theme.

Conclusion

As we stand on the cusp of 2025, the integration of artificial intelligence into financial leadership is no longer a distant prospect but an immediate reality reshaping the role of the modern CFO. This evolution presents a dual challenge: harnessing the transformative power of AI while adeptly balancing it with human insight and strategic acumen.

The path forward is clear yet complex. CFOs must navigate the intricacies of AI implementation—selecting the right data, understanding the limitations of different AI models, and ensuring data privacy and ethical considerations are at the forefront. Embracing both bottom-up and top-down AI approaches, as well as generative AI, offers immense potential for operational efficiency and insightful decision-making. However, the key lies in recognizing AI as an augmentation tool rather than a standalone solution.

Organizational success in this new landscape hinges on fostering a culture of experimentation, continuous learning, and adaptability. Empowering employees, enhancing skills, and breaking down traditional structural barriers will be essential. By encouraging cross-functional collaboration and holistic awareness, companies can overcome the inherent structural limitations and fully capitalize on AI’s capabilities.

Ultimately, the CFO of 2025 will be a catalyst for digital transformation, seamlessly integrating machine intelligence with human expertise. By maintaining a strategic focus on business fundamentals while embracing technological innovation, organizations can navigate the complexities of the modern business environment. Those who achieve this balance will not only enhance efficiency and decision-making but will also position themselves to thrive amid accelerated process velocities and converging domain areas. In a world where AI continues to expand the horizons of what’s possible, it’s the synergy of human and machine intelligence that will define the leaders of tomorrow.

Other Resources

• https://www.cfo.com/news/finance-leaders-are-using-generative-ai-tools-artificial-intelligence/735283/
• https://www.cfo.com/news/how-many-finance-employees-should-you-have-accounting-automation-revenue/734637/
• https://www.cfo.com/news/what-cfos-should-know-about-the-state-of-the-ai-market/734837/
• https://www.cfodive.com/news/tech-forward-finance-cfos-ai-literacy-seismic-data-investment-technology/735299/
• https://www.cfodive.com/news/bullish-ai-spenders-report-higher-return-rates-ey/735071/