“AI Forecasting” comes up in every discussion with clients in FP&A and finance, more broadly. The expectation is typically to automate and streamline the planning and forecasting process, reduce the burden on their people and improve the quality and accuracy of the outcomes. Most clients react to the vibe around AI and do not want to be left out, many are exploring the options and trying to understand what exactly AI can do for them.
However, at this stage in the AI lifecycle, the use cases that leverage AI in FP&A are not very well established or articulated. We find that the hype often exceeds reality, most companies do not have a clearly defined or articulated objectives of what they want AI to do for them. Many clients do not really understand all the components of AI and the associated risks, costs or benefits associated with its implementation.
Purpose of this post is to break down the AI capabilities in Finance and FP&A into some more manageable components. Not all AI is created equally. It is important to understand the different features and capabilities that exist, what they do and how they can be best applied. It would make for a better discussion about requirements and solutions
The sections below outline the various areas and key characteristics. At the end of the article we explore the general applicability of AI in Finance and especially FP&A that is analytical in nature and works predominantly with structured data.
AI in Support of the Plan Process
Most of the “AI” capabilities that currently exist relate to Gen AI, AI Agents or AI Assistants. These capabilities essentially offer a way to interact with models and data in natural language. They return answers in human speak and make data more accessible to users. Gen AI can provide written summary of findings or narrative around results, analyses, content of dashboards etc.
Some of the use cases we have seen end explored were to provide a commentary on variances based on data extracted from the system, description of results of a statistical or other model etc. However, most of the narrative is explaining the "WHAT" and not the "WHY". The systems now can do a decent job at describing what the numbers are (the quality of the output varies by AI platform and generally requires a fair amount of query tuning and model training to produce meaningful results). The AI tools are not very effective at explaining the driving factors. They provide the analyst a starting point but the explanation still has to be done by a human.
While the language capabilities are very impressive and rapidly improving, they apply to the PROCESS of planning and not the plan itself. They accommodate a preference of a new generation of users to interact with tools and systems. This will improve system adoption and process participation by users, which is a good thing. But they do not directly influence the quality and accuracy of the forecast, at least not yet..
AI in Support of Data Readiness
The above agents and assistants will only be as effective as the data that they work with. The effort of collecting and normalizing the data on the back end remains the same as it is today. This is the biggest challenge that companies face to ensure successful implementations.
AI can be helpful here, especially if some of the data is in written / unstructured format. Some AI tools will streamline the efforts around processing and normalizing structured data, learn from patterns etc and also add value in terms of including data that exists in unstructured format. But again - there is a fair amount of work required to "teach" the model how to get and return the data - we tested a number of tools and queries and often times the systems return data that is incomplete or just plain inaccurate.
Assuming you can get the query optimized to return reliable information, the completeness of data will indirectly improve the quality of the outcome and again that is a value-add component.
Despite all the proliferation of AI, most plans and forecasts continue to be constructed using either i) direct input, ii) set of calculations or arithmetic rules (which can be driver based) or, iii) statistical models. Not a lot of AI involved at this stage.
Input is pretty common. Users run models and perform calculations offline, input results into the system for consolidation and analysis. The calculations can be manual or automated, or a mix. The opportunity is to bring everything online.
Driver based modeling can be extremely effective, especially if companies include operational data and use them as assumptions to drive calculations. These models are very effective as they break through siloes, make use of data that exists across the enterprise and connect the dots within an integrated system / model. The calculations are usually simple arithmetic, with clear visibility into the logic and core drivers. There is a huge opportunity here that remains unexplored in most corporations and it is not really AI related.
Statistical models largely include well-established and widely used deterministic models such as ARIMA, BATS, Holt-Winters and others that have been in use for a very long time. They incorporate the effects of trends, seasonality or even impact of external variables (multivariate) to a time series of data. Modern tools allow users to effectively manage inputs and assumptions as well as model outcomes. This is making the methodologies more readily accessible to the average user. If we chose to call this “AI” then fine, but it is important to understand what the models are and the limitations that they present. If you let AI make up a forecast it will, however there is no way anyone can rely on it or substantiate it without understanding the drivers and math behind it.
Fundamentally, I do not think that AI naturally lends itself to the world of FP&A. AI is transactional in nature, most AI applications are language based or driven. It may be more effectively applied to accounting or finance operations with greater process intensity and repeatable tasks. In FP&A, once a problem is quantified, a performant, flexible and scalable multi-dimensional platform will provide the ability to effectively analyze the data and model outcomes with a high level of flexibility and precision as well as deep insight into the drivers of results. At this rate of focus and investment I am sure things will change and improve. However, at least for now, the ability of AI to materially add value in the finance and FP&A space is marginal at best.
Current vs Future Opportunities
There are many opportunities to significantly improve the planning and forecasting process as well as outcomes without relying 100% on AI to do this for us. We find that in any corporation we work with, no matter how big and sophisticated, there is a huge amount of manual processing, disconnected data, etc. Most problems can be easily addressed with well-established and proven tools, assuming proper governance, focus and execution. I think a lot of value can be generated today by focusing on fundamentals and using capabilities that are easily available to us today, while we continue exploring the future potential of AI in parallel.
Please reach out if you have comments or would like to discuss in more detail.