AI or RPA: Decision Matrix for CFOs

Thesis

Portuguese CFOs face a recurring choice: automate financial processes with Robotic Process Automation (RPA) or invest in Artificial Intelligence (AI) solutions. The issue is not technological—it's operational. RPA automates repetitive tasks based on fixed rules; AI learns patterns, adapts to variability, and supports decisions in ambiguous contexts. The wrong choice wastes capital, creates operational rigidity, or introduces opacity into critical processes.

Portugal ranks 17th among the 27 EU Member States in the DESI 2025 index, with 56% of the population possessing basic digital skills—close to the European average of 55.6%. SMEs represent 99.9% of the Portuguese business fabric, with limited resources for technological experimentation. In this context, the decision between AI or RPA requires rigor: mapping process structure, assessing data quality, quantifying operational risk, and diagnosing internal capacity before selecting the tool.

This article develops a decision matrix based on five critical dimensions: process structure, data quality and volume, operational risk, internal capacity, and value horizon. It presents use cases by financial function, identifies common implementation errors, offers diagnostic questions for CFOs, and outlines the path from decision to implementation. The aim is to provide financial decision-makers with a defensible framework for choosing between AI or RPA—or recognizing when neither should be the priority.

Concept Genealogy

Business process automation is not new. Its genealogy dates back to the 1950s, when data processing systems began to replace manual accounting and payroll tasks. What has changed in the past two decades is the democratization of two distinct technologies: RPA, which replicates human actions in software interfaces without altering underlying systems, and AI, which learns patterns from data and generalizes to new contexts.

RPA emerged as a commercial category in the early 2000s, with vendors like Blue Prism, Automation Anywhere, and UiPath popularizing the idea of "bots" executing repetitive tasks in legacy systems without the need for API integration. The promise was simple: reduce operational costs and human errors in structured processes, with rapid implementation and no system reengineering. Adoption accelerated after the 2008 financial crisis, when companies sought operational efficiency without major capital investments.

AI applied to business processes has older roots—academic machine learning dates to the 1950s—but its practical application in finance has only scaled in the last decade, driven by three factors: availability of cloud computing power, maturation of open-source frameworks (TensorFlow, PyTorch, scikit-learn), and accumulation of sufficient digital historical data to train models. Unlike RPA, AI does not replicate fixed rules; it learns patterns from examples and adapts to variability in inputs.

The conceptual distinction is critical. RPA is deterministic automation: if X, then Y. AI is probabilistic automation: given X, Y is likely with confidence Z. This difference has profound implications for audit, internal control, and risk management. An RPA bot failure is easily diagnosed; an AI model that biases decisions can be opaque even to its trainers.

The management literature has treated these technologies in a fragmented way. Kotter (1996) established the 8-step model for organizational change management but did not anticipate the speed of automation adoption or the specific resistance that RPA and AI generate in finance teams. Kaplan and Norton (1992) proposed the Balanced Scorecard as a framework for financial and non-financial metrics, but the inclusion of automation KPIs—robotized process rate, predictive model accuracy, post-automation cycle time—is recent and still not standardized.

What has changed in the past five years is the convergence of RPA and AI in hybrid platforms. Vendors such as UiPath, Automation Anywhere, and Microsoft Power Automate have integrated machine learning capabilities into RPA tools, enabling bots to execute structured tasks and invoke AI models for decisions requiring classification or prediction. This convergence complicates the decision: the boundary between pure RPA and applied AI has become blurred, and CFOs face commercial proposals promising both without clarifying when each is appropriate.

International Evidence

Empirical evidence on RPA and AI in financial processes is fragmented, but three patterns emerge from studies published over the past 10 years: efficiency gains are real but heterogeneous, implementation risk is underestimated, and ROI depends critically on organizational maturity.

Efficiency Gains and Heterogeneity of Results

Case studies published by global consultancies report reductions of 30% to 70% in processing time for tasks such as bank reconciliation, invoice posting, and financial data consolidation after RPA implementation. McKinsey Global Institute (2017) estimated that 45% of labor activities could be automated with existing technology, including 69% of tasks in accounting and auditing. However, these estimates are based on technical potential, not actual adoption or verified ROI.

The heterogeneity of results is significant. Deloitte (2020) reported that only 3% of organizations managed to scale RPA to more than 50 bots, and that 30% to 50% of initial RPA projects failed to deliver expected ROI. The most cited causes include poorly documented processes before automation, lack of bot governance, and unanticipated organizational resistance. The lesson is clear: RPA automates processes as they exist; if the process is inefficient, automation perpetuates inefficiency at greater speed.

AI and Predictability in Finance

AI application in finance focuses on three areas: cash flow forecasting, anomaly detection, and credit scoring. Evidence of effectiveness is more robust in fraud detection—machine learning models outperform heuristic rules in accuracy and false positive rates—but less clear in cash flow forecasting, where the quality of historical data and stability of business patterns determine model usefulness.

A PwC (2019) study with 500 global CFOs reported that 54% considered AI a strategic priority, but only 20% had implemented use cases in production. The main barrier cited was not technological, but organizational: lack of clean data, absence of internal data scientists, and difficulty explaining model decisions to auditors and regulators. This last point is critical in finance, where auditability and transparency are regulatory requirements.

Operational Risk and Model Opacity

The academic literature on AI risk in business contexts is recent but growing. Machine learning models can incorporate bias present in historical data, amplifying discrimination in credit or expense approval decisions. The opacity of complex models—especially deep neural networks—makes audit and compliance validation with accounting and tax standards difficult.

The European Central Bank (2020) published guidelines on the use of AI in financial institutions, requiring explainability of automated decisions, robustness testing, and model governance. For CFOs of non-financial companies, these guidelines are instructive: even if not legally binding, they set a diligence standard that external auditors and investors expect.

Consensus and Dissent in the Literature

There is consensus on three points: (1) RPA is effective for repetitive, structured, high-volume processes; (2) AI is superior to fixed rules in contexts with variability and ambiguity; (3) both require rigorous change management to avoid organizational resistance. There is dissent regarding ROI horizon—vendor studies report 6-12 month payback for RPA, while independent analyses suggest 18-24 months when including maintenance and process reengineering costs.

The most relevant dissent concerns scalability. Gartner (2021) predicted that 90% of large companies would have implemented some form of RPA by 2022, but that 40% would face "bot fatigue"—proliferation of unguided automations, creating operational fragility. The implication for Portuguese CFOs is clear: automation without governance creates risk, not efficiency.

The Portuguese Case

Portugal presents a paradox: significant progress in digital public services and infrastructure coverage, but business adoption of automation still concentrated in large companies and specific sectors. According to the State of the Digital Decade 2025 report by the European Commission, Portugal ranks 17th among 27 Member States in the DESI index, with strengths in digital public services and 5G coverage (65.2% of households in the 3.4-3.8 GHz spectrum), but with a persistent gap in digital skills among the population.

The Portuguese business fabric is dominated by SMEs: 532,174 non-financial companies in 2024, of which 99.9% are micro, small, or medium-sized enterprises. These companies generated aggregate turnover exceeding €319 billion in 2023 and contributed around €93.5 billion in gross value added. The concentration in SMEs has direct implications for AI or RPA adoption: limited financial resources, lower capacity to absorb technological risk, and dependence on external providers for implementation and maintenance.

Adoption data for RPA and AI in Portugal is not systematically published by INE or other official sources. Anecdotal evidence from sector associations and vendors suggests that RPA adoption is concentrated in banking, insurance, telecommunications, and utilities—sectors with high-volume processes and IT maturity. Industrial and service SMEs adopt automation in a fragmented way, often through modules in ERP systems (SAP, Microsoft Dynamics, Sage) or low-code tools like Microsoft Power Automate, without an integrated automation strategy.

The comparison with European averages is instructive. Eurostat (2023) reported that 8% of EU27 companies used AI, with significant variation between Member States: Denmark (24%), Finland (16%), Portugal (no disaggregated data, but sector estimates suggest below 10%). AI adoption in Portugal is concentrated in tech startups and large companies with internal data science capacity; traditional SMEs face barriers of skills, data, and governance.

A distinctive Portuguese factor is the availability of public incentives for digital transformation. The Portugal 2030 program provides €23 billion in European funds (2021-2027), with Compete 2030 allocating €3.9 billion for innovation and digital transition. These incentives cover investment in software, training, and consulting for automation diagnosis and implementation. However, the execution rate of European funds in Portugal has historically fallen short of potential, partly due to administrative complexity and lack of capacity to prepare applications in SMEs.

Another relevant factor is the labor cost structure. Labor productivity in Portugal is about 35% below the EU average, ranking 19th among Member States. This productivity gap creates competitive pressure to automate, but also means that the opportunity cost of maintaining manual processes is lower than in high-productivity economies. For Portuguese CFOs, the automation business case should consider not only FTE reduction, but also quality improvement, reporting speed, and the ability to scale without proportional hiring.

Finally, data maturity in Portuguese SMEs is heterogeneous. Companies that have adopted modern ERP in the past 10 years have structured transactional data and sufficient history to train AI models. Companies operating with legacy systems or partially manual processes face data gaps that make AI unviable without prior investment in data infrastructure. This heterogeneity means that the decision between AI or RPA cannot be standardized—it critically depends on each company's current digitalization state.

Five Critical Dimensions

The choice between AI and RPA is structured around five dimensions that determine technical suitability, operational feasibility, and expected ROI. Each dimension requires specific diagnosis before making a decision.

Dimension 1: Process Structure

RPA is suitable for deterministic processes: tasks with predictable inputs, fixed rules, and structured outputs. Examples include bank reconciliation (matching transactions by reference), posting recurring invoices (same structure, fixed fields), and extracting data from standardized PDFs (bank statements, delivery notes). In these cases, the logic is if-then-else, and automation replicates exactly what a human operator would do, but without typing errors and at constant speed.

AI is suitable for processes with variability, ambiguity, or the need for judgment. Examples include classifying expenses into accounting categories when the description is ambiguous, forecasting cash flow when there is complex seasonality or non-recurring events, and detecting anomalies in transactions when fraud patterns evolve. In these cases, the logic is probabilistic: the model learns patterns from historical examples and generalizes to new cases, adjusting as new data is incorporated.

The most common mistake is applying RPA to processes that appear structured but contain frequent exceptions. An RPA bot that encounters 10% of cases outside the programmed rules will fail or escalate to human intervention, creating exception management overhead. In such cases, AI may be more robust, but requires historical exception data to train the model to recognize them.

Dimension 2: Data Quality and Volume

RPA does not depend on historical data—it operates on current data, in real time, following programmed rules. The only requirement is that inputs are in a format readable by the bot (form fields, Excel cells, software interface elements). This makes RPA viable even for companies with limited historical data or legacy systems without APIs.

AI requires clean, representative historical datasets in sufficient volume to train models with acceptable accuracy. The empirical rule varies by model type: linear regression may work with hundreds of observations; deep neural networks may require thousands or millions. For finance, the practical requirement is at least 12-24 months of structured transactional data, without significant gaps, and with correct labels (e.g., transactions classified as fraudulent or non-fraudulent if the goal is fraud detection).

Data quality is more critical than volume. A model trained with biased data—for example, a credit approval history reflecting past discrimination—will perpetuate and amplify that bias. For CFOs, this means auditing data before investing in AI: identifying gaps, correcting inconsistencies, and validating that historical data reflects patterns to be replicated, not anomalies to be eliminated.

Dimension 3: Operational Risk

RPA introduces rigidity risk: bots execute fixed rules and fail when the context changes. If a supplier changes invoice format, the bot extracting data from that invoice stops working until reprogrammed. If an ERP system is updated and the interface changes, all bots interacting with that interface need adjustment. This risk is manageable through bot governance—centralized inventory, regression testing after system changes, and fallback processes for human intervention when bots fail.

AI introduces opacity risk: complex models make decisions that may be difficult to explain, even to their trainers. A credit scoring model that rejects a client may not provide a clear justification, creating legal and reputational risk. A cash flow forecasting model that underestimates liquidity needs can lead to wrong treasury decisions. This risk is manageable through model governance—documentation of assumptions, robustness testing, validation by independent teams, and preference for interpretable models (regression, decision trees) over opaque models (deep neural networks) when explainability is critical.

For processes subject to external audit or regulation, the opacity risk of AI may be unacceptable. Financial and tax auditors expect full traceability of accounting entries; an AI model classifying expenses without clear justification may not pass audit. In such cases, RPA with explicit rules or AI with interpretable models are preferable to black-box solutions.

Dimension 4: Internal Capacity

RPA can be managed by generalist IT teams or business analysts trained in low-code/no-code tools. Platforms like UiPath, Automation Anywhere, and Microsoft Power Automate offer visual interfaces that allow automation flows to be designed without advanced programming. The learning curve is weeks to months, and bot maintenance can be decentralized to business users with basic training.

AI requires data scientists, data engineers, or specialized external partners. Training machine learning models demands knowledge of statistics, programming (Python, R), ML frameworks (scikit-learn, TensorFlow), and the ability to diagnose when a model is overfitting or underfitting. The learning curve is months to years, and model maintenance—periodic retraining, drift monitoring, hyperparameter tuning—requires technical skills that few SMEs possess internally.

For CFOs of Portuguese SMEs, this implies a choice: build internal capacity (hire or train), outsource to partners (consultancies, software vendors), or limit automation to RPA until internal capacity is available. Outsourcing has cost and dependency risk; internal building has time cost and risk of technical talent turnover in a competitive market.

Dimension 5: Value Horizon

RPA delivers quick, incremental gains. Implementing a bot for bank reconciliation can be completed in weeks, with visible ROI in months: reduction of manual work hours, elimination of typing errors, acceleration of month-end closing. The value is tangible and easily quantifiable. However, RPA does not transform processes—it optimizes them. If the underlying process is inefficient, RPA merely makes the inefficiency faster.

AI delivers structural transformation in the medium term. A cash flow forecasting model that reduces treasury uncertainty enables more aggressive investment decisions and reduction of liquidity buffers. A fraud detection model that identifies non-obvious patterns protects the company from significant losses. However, AI's value is probabilistic and diffuse: it does not manifest as direct FTE reduction, but as improved decision quality, risk reduction, or the ability to scale without proportional overhead growth.

For CFOs under short-term pressure—operational cost reduction, margin improvement—RPA is more defensible. For CFOs with a mandate for structural transformation—preparation for growth, governance improvement, building analytical capacity—AI is justified, but requires patience and clear communication of value to stakeholders.

Use Cases by Financial Function

The application of these technologies varies by financial function. Accounts payable and receivable, financial reporting, treasury, and internal control each present distinct profiles of structure, data, and risk, requiring differentiated choices.

Accounts Payable and Receivable

RPA is suitable for invoice matching with purchase orders and delivery notes (three-way match), posting recurring invoices, and automatic sending of payment reminders. These processes are high-volume, structured, and based on fixed rules. A bot can process hundreds of invoices per hour, error-free, freeing AP/AR teams for exception management and supplier or client negotiation.

AI is suitable for predicting payment delays (identifying clients with a high probability of default, enabling proactive action), prioritizing collections (classifying debts by recovery probability and required effort), and detecting duplicates or invoice fraud (identifying anomalous patterns that fixed rules do not capture). These cases require historical data on payment behavior but deliver greater value than RPA when client behavior variability is high.

Financial Reporting

RPA is suitable for consolidating data from multiple sources (ERP, CRM, POS systems), automatic generation of periodic reports (balance sheets, income statements, cash flow statements), and distribution of reports to internal stakeholders. The logic is deterministic: extract data, apply templates, format outputs. The value is reduced month-end closing time and elimination of manual copy errors.

AI is suitable for automatic variance analysis (identifying drivers of deviations from budget or forecast), automatic narrative (generating written commentary explaining movements in financial metrics), and anomaly detection in consolidated data (flagging accounting entries outside the norm that may indicate error or fraud). These cases require models trained with a history of variances and approved narratives but deliver insights that RPA cannot generate.

Treasury

RPA is suitable for daily bank statement reconciliation, execution of scheduled transfers, and updating cash positions in treasury dashboards. These processes are critical for liquidity control but are repetitive and do not require judgment. RPA automation reduces manual error risk and accelerates information availability for treasury decisions.

AI is suitable for cash flow forecasting (incorporating seasonality, non-recurring events, and correlations with external variables such as exchange rates or commodity prices), working capital optimization (identifying opportunities to reduce inventory days or extend payment terms without impacting supplier relationships), and interest or exchange rate risk management (simulating stress scenarios and recommending hedges). These cases require historical cash flow data and financial modeling capacity but deliver strategic value beyond RPA's reach.

Internal Control and Audit

RPA is suitable for automatic compliance tests (verifying that all expenses above a certain amount have approval, that accounting entries follow the chart of accounts, that segregation of duties is respected) and sample extraction for audit (selecting transactions by defined criteria for manual review). These processes are based on fixed compliance rules and benefit from continuous execution, not just audit cycles.

AI is suitable for fraud detection (identifying transaction patterns deviating from historical norms, such as expenses submitted outside working hours, duplicate amounts, or fictitious suppliers), third-party risk analysis (classifying suppliers or clients by compliance risk, based on public data and transaction history), and continuous audit (real-time monitoring of risk indicators and flagging anomalies for investigation). These cases require models trained with examples of fraud or non-compliance but deliver proactive detection capabilities that rule-based compliance tests cannot achieve.

Common Errors and How to Avoid Them

Implementation of the technological choice in finance often fails due to four avoidable errors, each with a distinct implication for CFOs.

Error 1: choosing AI due to technological hype when RPA would suffice. AI is perceived as more advanced, and vendors promote machine learning capabilities even when the process only requires rule execution. The result is over-engineering: longer, more expensive projects, higher risk of failure, and no value gain over RPA. The correction is rigorous diagnosis of process structure before choosing technology: if the process is deterministic, RPA is sufficient.

A practical example is bank statement reconciliation. Vendors may propose AI solutions to "learn" matching patterns, but if the matching rules are fixed (reference number, amount, date), RPA is faster, cheaper, and more auditable. AI only adds value if there is significant variability requiring probabilistic classification—for example, matching transactions without a unique reference, where the model must infer correspondence based on multiple fields.

Error 2: implementing RPA without documenting processes, automating inefficiencies. RPA replicates processes as they exist. If the process includes redundant steps, unnecessary approvals, or rework due to poor upstream data quality, the bot will automate these inefficiencies. The result is speed gain without real efficiency gain, and organizational resistance when teams realize automation did not solve underlying problems.

The correction is process mapping and reengineering before automating. Tools such as structured project management and Lean methodologies can identify waste and simplify flows before RPA is applied. This step adds time to the project but multiplies ROI: automation of an optimized process delivers more value than automation of an inefficient process.

Error 3: deploying AI without data governance, resulting in biased or unauditable models. AI models learn patterns present in historical data. If this data contains bias—for example, a credit approval history that discriminates against certain customer segments—the model will perpetuate that bias. If the data has gaps or inconsistencies, the model will learn spurious patterns that do not generalize. The result is automated decisions that fail in production, create legal risk, or do not pass audit.

The correction is data audit before training models: identify bias, correct inconsistencies, validate dataset representativeness. This requires data engineering capacity—cleaning, transformation, validation—that many SMEs lack. For CFOs, the implication is that AI is not just software purchase; it is investment in data infrastructure and governance, with costs and timelines that must be included in the business case.

Error 4: underestimating the need for change management and team training. Automation changes roles, responsibilities, and workflows. Finance teams operating manual processes may resist RPA or AI due to fear of redundancy, loss of control, or inability to manage new tools. Without rigorous change management—clear communication of objectives, training in new skills, redesign of roles to focus on analysis rather than execution—automation projects generate organizational friction and fail to scale.

The correction is to include change management as a formal workstream in automation projects, with defined budget and responsibility. Kotter (1996) established the 8-step model for change management, which remains applicable: create a sense of urgency, form a leadership coalition, communicate the vision, remove obstacles, generate short-term wins, and consolidate gains. For CFOs, this means automation is not just an IT project—it is an organizational transformation project, with implications for organization and culture.

Implications for Decision-Making

CFOs evaluating intelligent or rule-based automation face five critical questions, each with a distinct implication for the decision.

Question 1: Is the process repetitive and based on fixed rules, or does it require judgment and adaptation to context? If the process is deterministic—predictable inputs, explicit rules, structured outputs—RPA is sufficient and faster to implement. If the process has variability, ambiguity, or requires probabilistic classification, AI may add value but requires historical data and modeling capacity. The implication is that the technology choice should follow process structure diagnosis, not technology preference.

Question 2: Do we have clean, representative historical data to train AI models? AI requires datasets with volume, quality, and correct labels. If the company operates with legacy systems, partially manual processes, or fragmented data across multiple sources without integration, AI is not viable without prior investment in data infrastructure. RPA, by contrast, operates on current data and does not depend on history. The implication is that the choice between technologies may be determined by data maturity, not just process suitability.

Question 3: What is the operational risk of error—greater in the rigidity of fixed rules or in the opacity of automated decisions? RPA fails when the context changes; AI can make biased or inexplicable decisions. For critical processes subject to audit or regulation, opacity risk may be unacceptable. For high-volume processes with low tolerance for downtime, rigidity risk may be unacceptable. The implication is that the choice should consider the process risk profile, not just expected efficiency.

Question 4: Do we have internal capacity to manage and audit the solution, or do we need an external partner? RPA can be managed by generalist IT or business analysts; AI requires data scientists or specialized partners. For SMEs without internal capacity, outsourcing has cost and dependency risk. The implication is that the choice of automation model may be determined by talent availability, not just technical suitability. CFOs should assess whether the company has or can build internal capacity, or whether automation should be limited to RPA until that capacity is available.

Question 5: What is the expected value horizon—quick, incremental gains or medium-term structural transformation? RPA delivers tangible ROI in months; AI delivers strategic value over years. For CFOs under short-term pressure, RPA is more defensible. For CFOs with a transformation mandate, AI is justified but requires clear communication of value to stakeholders and patience to allow models to mature. The implication is that the choice should align with the company's strategic horizon and board or investor expectations.

Diagnostic Questions for the CFO

Before deciding between AI and RPA, CFOs should answer five internal diagnostic questions:

• Have we mapped financial processes by volume, frequency, variability, and error impact, identifying which are candidates for automation?
• Have we assessed the quality of historical data—completeness, consistency, absence of bias—and the data engineering capacity to prepare datasets for AI?
• Have we quantified the operational risk of rigidity (RPA) versus opacity (AI) for each critical process, and validated that the chosen solution will pass external audit?
• Have we diagnosed internal IT and data science capacity, and decided whether to build, hire, or outsource?
• Have we defined success KPIs—cycle time reduction, error elimination, forecast quality improvement—and established automation governance with clear responsibility?

These questions structure the decision and reduce the risk of choices based on technological hype or vendor pressure. For companies unable to answer confidently, the next step is not to choose technology—but to invest in digital maturity diagnosis and process mapping, with support from specialized management consulting.

Next Steps: From Decision to Implementation

The transition from decision to implementation of these technologies is structured in four steps, each with a clear deliverable and success criterion.

Step 1: Map financial processes by volume, frequency, variability, and error impact. The goal is to create an inventory of processes eligible for automation, classified by suitability for RPA (structured, repetitive, low variability) or AI (variable, ambiguous, need for prediction or classification). The deliverable is a process matrix with priority scoring, based on business impact and technical feasibility. This step can be executed internally or with support from consulting specialized in digital transformation in finance.

Step 2: Assess data quality and internal IT and data science capacity. The goal is to diagnose whether the company has sufficient historical data for AI, and whether it has internal talent to manage RPA or AI in production. The deliverable is a data maturity report and competency gap analysis, with recommendations to build, hire, or outsource. This step is critical to avoid projects that fail due to lack of data or maintenance capacity.

Step 3: Pilot the solution in a low-risk, high-visibility process. The goal is to validate technology, test governance, and generate a short-term win that builds organizational momentum. The deliverable is an RPA bot or AI model in production, with monitored performance KPIs and documented lessons learned. Typical pilot processes include bank reconciliation (RPA) or short-term cash flow forecasting (AI). This step should last weeks to months, not years, to maintain focus and avoid scope creep.

Step 4: Establish automation governance, including audit, change management, and performance KPIs. The goal is to scale automation in a controlled manner, avoiding proliferation of unguided bots or opaque AI models. The deliverable is a governance framework with centralized automation inventory, approval processes for new bots or models, regression testing after system changes, and performance dashboards. This step is often neglected but is critical for medium-term automation sustainability.

For CFOs of Portuguese SMEs, these steps can be supported by public incentives. The Compete 2030 program, with €3.9 billion allocated for innovation and digital transition, covers investment in software, training, and consulting. Macro Consulting supports CFOs in assessing digital maturity, process mapping, and designing automation roadmaps, including preparing applications for incentives and funding for eligible projects.

Where the Topic Is Fragile

The decision matrix presented assumes three conditions that do not always hold, limiting the applicability of the argument.

First, it assumes that processes are mappable and stable. In fast-growing companies, with frequent changes in product, market, or business model, financial processes may be too volatile to justify automation. In these contexts, investing in flexibility—modular systems, multifunctional teams—may deliver more value than automating processes that will be redesigned within months.

Second, it assumes that historical data reflects patterns to be replicated. In companies that have undergone mergers, ERP system changes, or restructurings, historical data may not be representative of future operations. AI trained with pre-transformation data may generate forecasts or classifications irrelevant post-transformation. In such cases, RPA may be more robust, but the value of automation is limited until operations stabilize.

Third, it assumes that CFOs have the mandate and budget to invest in automation. In companies under financial pressure, with restricted liquidity or ownership focused on short-term survival, automation may not be a priority—even if technically justifiable. In these contexts, the ROI argument for automation is academically correct but operationally irrelevant.

Open Questions

The literature on technological choice in finance leaves three critical questions unanswered, creating experimentation opportunities for leading-edge companies.

First question: what is the balance point between automation and flexibility? Automation reduces variable cost but increases fixed cost (bot maintenance, model retraining, governance). In high-uncertainty environments, flexibility may be worth more than efficiency. The question is: how to quantify the option value of maintaining manual processes that can be quickly adapted, versus the efficiency gain from automation? The answer varies by sector, growth stage, and market volatility, and there is no standardized framework.

Second question: how to manage the transition from RPA to AI as data maturity increases? Companies starting with RPA accumulate historical data from automated processes, creating datasets that can train AI models. The question is: when and how to make this transition? Replacing RPA bots with AI models has re-implementation cost and operational disruption risk. The literature does not offer clear timing criteria or migration methodology.

Third question: how to measure AI ROI when value is diffuse and probabilistic? RPA delivers tangible gains—hours saved, errors eliminated. AI delivers improved decision quality—better forecasts, proactive risk detection—that is hard to quantify. The question is: how to build an AI business case that is defensible to CFOs, boards, and investors who demand clear financial metrics? The answer may require new valuation frameworks that incorporate option value and risk reduction, not just cost reduction.

Sources

• European Commission (2025), State of the Digital Decade 2025, annual report on digital progress of EU Member States.
• INE — Instituto Nacional de Estatística (2024), Empresas em Portugal 2024, definitive data on the Portuguese business fabric.
• INE — Instituto Nacional de Estatística (2024), Contas Nacionais Anuais, consolidated version (September 2025).
• Banco de Portugal (2025), Boletim Económico Dezembro 2025, macroeconomic projections for Portugal 2025-2028.
• Agência para o Desenvolvimento e Coesão (2026), Portugal 2030, data on execution of European funds (updated April 2026).
• McKinsey Global Institute (2017), A Future That Works: Automation, Employment, and Productivity, analysis of automation potential by sector and function.
• Deloitte (2020), The State of RPA: Scaling Robotic Process Automation, study on RPA adoption and scalability in large companies.
• PwC (2019), AI Predictions 2019, global survey with 500 CFOs on AI priorities.
• European Central Bank (2020), Guide on the Use of Artificial Intelligence in Credit Institutions, guidelines on AI governance in financial institutions.
• Gartner (2021), Predicts 2022: RPA Renaissance Driven by Automation Fabric, forecasts on RPA evolution and convergence with AI.
• Eurostat (2023), ICT Usage in Enterprises, statistics on digital technology adoption by EU27 companies.
• Kotter, J. P. (1996), Leading Change, Harvard Business School Press, 8-step model for organizational change management.
• Kaplan, R. S. & Norton, D. P. (1992), The Balanced Scorecard: Measures That Drive Performance, Harvard Business Review, balanced metrics framework.
• IAPMEI (2024), Edição PME Líder 2024, characterization of leading SMEs in Portugal.
• AICEP Portugal Global (2024), Dados de Investimento Directo Estrangeiro, statistics on FDI in Portugal.