Waiting for impeccable data before activating AI in finance is often a mistake in timing. The real lever consists of starting from business frictions, exploiting what already exists and building a layer of financial intelligence capable of giving meaning to the figures.
CFOs keep asking themselves the same questions about AI: what to do if the data is imperfect…? And even worse, if they are not connected to the company’s financial system…? With the hindsight we have today, the answer is clear: in the majority of cases, existing data and systems are more than sufficient to get started.
The key to success: rely on what already exists, identify truly strategic data, then strengthen governance as uses evolve.
First step, identify friction
To get started, it is necessary to assess the friction points to understand where AI is having an immediate and measurable impact. We must identify a decisive moment where the finance function bears responsibility for results, even though it is dependent on signals detected upstream. This point of convergence, between responsibility and anticipation, often constitutes the most relevant use case for AI. The mission of the financial director is based on the analysis of the upstream factors which determine demand forecasts, commercial pipeline anticipations and, more broadly, overall financial performance, factors which he very rarely controls. In other words, finance is responsible for results, but remains dependent on elements that it does not control.
Artificial intelligence and machine learning make it possible to transform these often opaque and time-consuming discussions about key factors into clear, transparent and measurable indicators, without the need to wait for a perfect data structure. Rather than waiting for the complete modernization of the data ecosystem, the question is whether we can already draw relevant lessons from the historical data associated with these levers. For the vast majority of organizations, the answer is positive.
The current capabilities of artificial intelligence and machine learning make it possible to act without delay. Drawing on a sample of structured data, it is possible to generate data-driven demand forecasts or financial projections. This exercise makes it possible to compare these new models with historical plans and thus to reveal structural biases or areas of volatility while precisely measuring the gain in reliability. These first concrete successes make it possible to substitute tangible evidence for intuitions and to clarify the roadmap even before the target architecture is fully defined.
This pragmatic approach has already proven itself in the field. Thus, a large organization had to anticipate consumption volumes between its operational units to adjust its pricing strategy. Although the data was available, its dispersion between a data warehouse and a manual management model caused significant blockages. Despite a still partially manual data pipeline, the company decided to move forward, demonstrating that immediately leveraging existing resources takes precedence over waiting for perfect automation. Using extracts from historical volumes, weather data, demographic trends and other regional indicators, the team developed a forecast driven by Finance and based on AI in just a few weeks.
The result: a production-ready model with over 90% accuracy, over 30% reduction in forecast errors compared to the previous model, and approximately 90% reduction in cycle time.
The historical data pipeline was automated a few months later, without significant slowdown or rework during implementation.
The crucial question of preparation
Executives are being understandably cautious: many believe that investing in early, targeted, progressive applications of AI is futile without a modern, robust data ecosystem. This distrust is often fueled by the belief that poor quality input produces poor quality results. However, this is no longer entirely true.
AI models aren’t just about numbers. They also take context into account. A machine learning-based forecast should result from a real understanding of what the numbers mean to the business. This involves identifying the entity, product family, cost center, and reporting point that the data represents, as well as the hierarchies to which these dimensions belong and their relationships to each other. It also requires understanding what “granularity” actually means within the organization: how reporting is done, how analysis is done, and how decisions are made. This includes mastering the chart of accounts, consolidation structure, legal entities, reporting hierarchies, key planning points and data traceability. This is the financial intelligence layer, the structured representation of the financial and legal functioning of the company. It is this layer, more than a perfectly designed process, that guarantees the precision, efficiency and value of analyzes in concrete deployments.
The development and management of this layer is the responsibility of the department which understands the downstream financial process and its subtleties: finance.
Adjusting demand based on prices to generate revenue, distributing transaction volume between legal entities, evaluating business units against company objectives: these processes fall under finance, and for a specific reason. It’s in the feedback loop between AI model results and financial decision-making, and vice versa, that true success happens. Without this layer of financial intelligence and without the right experts to manage it, one can only feed meaningless numbers to a model.
What does this look like on a large scale?
The distinction between AI capable of relevant experiments and AI producing usable results for decision-making resides almost systematically in the context of governance. On a large scale, the accuracy of a model does not depend solely on its algorithm. It depends on the characteristics and predictive variables accessible to the model, the institutional knowledge emanating from planners, and the organizational structure which defines the performance indicators. As operators share their observations and feedback within the organization, the system becomes more accurate to go beyond forecasting capability. This same principle is applicable to analytics and data.
The level of detail at which information is revealed is crucial. Managers do not need to analyze variations at the level of each product reference. They also don’t need the performance details of a single account manager to evaluate sales. They need information that matches their vision of the business. This level is defined by the financial intelligence layer. Without it, we drown in details or lose essential information due to too much aggregation.
This layer of intelligence becomes all the more essential as the financial sector adopts autonomous AI systems that, beyond simply revealing information, make decisions and act without human intervention. Companies that establish well-governed layers of intelligence now will position themselves advantageously when this transition to autonomy becomes more widespread.
How to invest team time well?
Employee time is limited, and chasing perfect data at the source is not the most cost-effective way to use it. The organizations that are moving the fastest with AI in Finance are focusing their efforts on two areas: identifying data that is ready to be exploited now and governing the data as soon as it is integrated into the financial intelligence layer, rather than trying to clean each system upstream beforehand.
This distinction is important. If, for example, forecasting is done at the business unit and product line level, this is where accuracy and traceability should be optimal. Not all source data needs to be perfect, only those corresponding to the decision-making level need to be.
Organizations that understand this approach will not only improve their processes. They will build a layer of governed intelligence capable of understanding their activity. It is this layer, and not the data itself, that will make all the difference between which companies will be able to use AI most effectively.
____________________________
Par Eric VidalVP of Sales chez OneStream
____________________________
