About 18 months into the AI journey in Finance and FP&A, the proven business case and value-add of AI is very limited. While AI capabilities are deployed and used at reasonable scale, the impact of AI as a transformative force in FP&A is still very low. It is far below the hype promoted by various vendors and consulting firms in the market.
Finance and FP&A teams use AI tools and models extensively. A lot of this use is Copilot, we see rapid adoption of Claude and many other models. However, majority of these use cases represent “personal productivity”. They help streamline the work of an individual or perhaps a small team. A large-scale application of AI that would meaningfully automate Enterprise Performance Management in a way that would justify the business case is still pretty far out.
Companies now see the results of their early investment. They are getting a better picture of what works and what does not. The initial hype around AI subsided and gave way to more realistic expectations. Companies are settling into the reality of what AI can deliver and are adjusting their plans and priorities accordingly. I think this is a good thing. It leads to more constructive and reasonable client discussions about possibilities, which should lead to better and more tangible outcomes.
I wrote this document as a summary of pieces from various presentations, briefing and engagements with our customers. My objective is to try to separate hype from reality. There is still a very significant FOMO in the market. Customers have unreasonable expectations of AI based on narratives that are being pushed by vendors and consultants. Most of these narratives are way too optimistic and create a perception for users that they are behind somehow.
They are not. There are still significant challenges with AI capabilities being deployed over finance data on a scale. Technical integration, information security, applications and especially the business case are still being defined. We are in very early stages of this journey.
I fully expect the capabilities to evolve and improve as significant investment in AI continues. We will update our analysis accordingly. Any direct feedback from the FP&A community is always welcome.
This paper is looking at Finance from an Enterprise Performance Management perspective. It is primarily focused on the FP&A function. It does not relate to, or capture, the benefits that may exist in other areas of finance, including accounting, tax, finance ops, corporate finance and M&A, investment management and other.
The write-up is based on our engagement with over 20 enterprise customers in various stages of investment into AI and spans a wide variety of platforms and technologies. The document addresses the following areas:
Two Primary AI Application Areas
Challenge with the Applicability of AI in Finance and FP&A
Language and Operational Focus of AI
Ability to Navigate Multi-Dimensional Data Stores
Navigating Application Complexity
High Barriers of Entry and Difficult Business Case
Typical AI FP&A Use Cases
Natural Language Interactions
Executive Reporting
Management Reporting and Dashboarding
Variance Analysis and Commentary
Data Validation and Reconciiation
"AI Forecasting" - Budget and Forecast Automation
Most Common AI Deployment Methods for Finance and FP&A
Personal Productivity
Embedded AI Agents in EPM Tools
AI at the Data Layer
Custom AI Agents and Capabilities
At a very high level, the AI application in FP&A have bifurcated into two distinct focus areas in terms of productivity and value add:
The focus here is to generate efficiencies for the business and FP&A area and transform the FP&A function. The processes targeted here are budgeting, forecasting, reporting and analytics. The hope is that AI can help automate many of the tasks and create significant efficiencies for the function.
Diving one level deeper, this can be further broken down into two focus areas:
IT and Operations is an area where we continue to see rapid and effective adoption of AI at-scale. After the initial wave of investment, companies observed slow progress and efficiency on the finance / business side and very rapid ROI on the IT side. This is where a lot of the investment is paying off.
The use of AI in the IT space is very prevalent. It is used extensively for coding, testing, various operational and administrative tasks and the use cases keep growing. This is a natural development. IT tasks are much more defined and specific / repeatable. Many AI vendors, Anthropic most notably, focused on the IT community from the inception. Therefore it is natural that a lot of traction and value is extracted from this space.
We will not be spending any more time on this topic in this paper. This write-up is focused on Finance and FP&A, and specifically on the large-scale transformation objectives. We can cover the various IT use cases in separate discussions.
I think there is a fundamental and structural difficulty for AI to deliver a strong business case in FP&A. This has been described in my previous post HERE. The key reasons and drivers for this are the following:
At its very core, Artificial Intelligence is not very well aligned with the needs of Finance and FP&A.
AI capabilities, as they were designed and exist today, are centered primarily around language and they are “operational” in nature. The Large Language Models (LLMs) are just that – they are built to understand language and be able to interact / converse with humans in a way that we are used to. They are extremely useful in filtering large amounts of dispersed and unstructured data found in the digital domain. They parse out the relevant information and serve it up to users in a concise and structured manner. They learn and are increasingly more sophisticated as they mature.
This is super useful in terms of doing broad research, answering specific questions about relevant topics etc. This is why we all use ChatGPT or the other tools to help with many operational tasks.
FP&A represents the opposite of this universe. It focuses on numbers (vs words) and is analytical (vs transactional). The numbers it relies on are organized, stored and structured in various forms (database, XLS etc). The scope of the problem is well defined. It is difficult for AI to deliver a lot of value-add at an acceptable cost over the tools and capabilities that are already broadly available.
Most data in FP&A lives in OLAP or some combination of data stores. OLAP (multi-dimensional) data stores represent significant portion of where data resides in any FP&A platform or infrastructure. AI tools were not developed, and are not very good at, reading from them. They are getting progressively better at it, but it is not a core capability / part of initial design.
OLAP can be difficult to navigate. It contains large volumes of data that is organized across dimensions with multiple hierarchies, attributes, variables. It is critical to define and understand the proper context to ensure retrieval of the correct information.
This is one of the key reasons why the EPM vendors struggle to deliver robust AI capabilities directly in the EPM platform. They are significantly behind what is available at the data layer (more on that below) and with unstructured data in general.
This is probably the biggest challenge. In any meaningful size organization, the data that FP&A needs to conduct proper analysis lives in many data stores across a potentially complex infrastructure. These include OLAP, relational, data files or other. There are rules and calculations that drive outcomes, such as resource, cost and capital allocations to measure profitability etc. These may rely on data and assumptions that span multiple data stores. There are various data transformations, mappings, adjustments etc. The applications are fluid with frequent updates of forward-looking scenarios (plan, forecast) by users, which can be pervasive in case of driver-based or bottom-up planning.
Unless an AI agent can access all this data, understands the information flow and is able to navigate the application in all its complexity, its value will be extremely limited. This is what we see today. Yes, you can retrieve data from the system using natural language. Yes, you can update forecast using a simple language command. However, all these commands are typically executed at a high level and represent only a very small and simplistic subset of the overall model. The ability to affect drivers across the system at scale is extremely limited given the significant complexity involved.
Until this is resolved and addressed somehow, AI will not be able to provide the transformational value that is needed to validate the business case in FP&A.
Even if all the above technical / structural challenges above could be addressed, the business case for AI is very challenging.
There are many proven tools and technologies that support the Finance and FP&A space that work extremely efficiently and deliver value already. These include various OLAP, BI or personal productivity tools. They are all relatively cheap, thoroughly proven and widely available. They can serve up numbers almost immediately with great data integrity and reliability. The models are transparent and ensure that information that is retrieved or analyzed is correct.
The challenge for AI is to prove its value above and beyond what is already there and at a cost that would justify the business case. Application of AI at a meaningful scale can be expensive. It includes ongoing cost of tokens and infrastructure in addition to other operational overhead. The compliance and information security implications are still not well understood or vetted. Given some of the limitations described here, the business case for a large-scale AI deployment has been a difficult one to prove thus far.
Many companies implement AI to close gaps that exist with their current systems and applications. However, most issues with FP&A tools and capabilities today do not relate to technology. They exist because most applications are poorly designed and implemented. This should be addressed through proper design and execution, which emanates from well-articulated set of requirements, vision and business strategy. There is a tremendous opportunity to capture value and transform the FP&A function using existing tools. This should be done while continuing to explore and incorporate AI capabilities as they evolve. However, the business case here is significant and it is not dependent on AI.
The initial application of AI in Finance and FP&A was conversational. Instead of running a report or inputting a value / assumption, a user can just tell the agent what they want, and it will do it for them. It feels natural, it is conversational and makes for a more fluid interaction.
All these capabilities exist and work well across multiple use cases and specific applications. They are being used and are widely available. However, beyond flashy demos and some simple use cases they did not transform the way we do business just yet. The key reasons are
An often cited second requirement of AI is for executives to get information in real time using simple natural language queries. “A CFO goes to a meeting and needs some key data points – they simply ask the agent and AI will provide the answer”.
While the capability is there, the level of deployment is lacking. For this to be successful there are two requirements:
The tools generally do not have access to all the information which you can serve up, as described above. And our ability to blindly trust AI answers is pretty far away still.
One way this requirement is being addressed is to create a summarized data repository inside the application specifically designed to answer a defined set of questions. Almost like an Executive Layer inside the application with a defined scope and dataset. This provides some level of control over data quality and reduces risk of issues, however it limits the depth and breadth of information that is available.
The application provides a “first line of defense” in terms of asking all the basic questions. A CFO doesn’t have to call someone or open a deck / login to system if they need something. The agent provides some basic insight and the CFO can still call someone to get deeper insight if needed. This is beneficial because it moves the discussion into more value-add category. It is a small step, however, and the return on that is small.
As of today, though, we did not see any company deploy this capability to their executives at any meaningful scale. The ones that have, did so with specific guardrails or handholding in place. They just do not trust the answers. A simple question that is asked just slightly differently could produce different answers and that is not a risk that is acceptable at any level in this domain.
There are many examples and use cases of AI for dashboarding and reporting. I put these in the category of ‘Personal Productivity”. There are many tools out there today for dashboarding and reporting. To me having an AI agent create a chart is not deeply transformational.
The differentiating factor in this category is the ability of AI to handle the application complexity described above. If it could, the feature and its potential would be much more valuable.
Same as above – there are many agents and capabilities out there that can analyze variances and write commentary. The question, again, is model complexity and depth of information the agent can analyze. With most complex models remaining out of reach, any analysis will be narrowly defined or very high level and therefore of limited use.
As relates to Commentary – the AI agents can write pretty long and elaborate summaries. It is intriguing to see and think that it can help automate reporting and executive updates. The challenges we saw with this in practice are the following:
I think that the best way to think about this capability right now is to use the AI analysis as the initial pass that gets reviewed by the analyst, therefore automating the initial stage of financial analysis. The analyst can review and spend time adjusting / correcting and focusing on more value-add part of the analytics cycle. There is surely value in that.
I think there is definitely potential here and the capabilities will evolve as the models get better.
There are several applications and use cases related to automation in this domain. While obviously extremely valuable, I do not view these as core to automating the FP&A function from an EPM perspective. I look at these as technical elements that need to be covered in every finance implementation. There are tools and technologies to do that. I wanted to acknowledge this area but will skip it and perhaps return to it in the future.
The term “AI Forecasting” is nothing but a marketing catchphrase. Any automation that currently exists in the forecasting or modeling space has very little to do with AI.
Majority of the talk about this topic out there is marketing and positioning by vendors that try to ride the AI wave and sell their product. They use the term “AI” to describe basic arithmetic or perhaps statistical algorithms that existed for a long time. It creates a perception that their platform is more sophisticated than it really is and differentiates them from competition.
When people say “AI Forecasting”, they typically refer to one of the two capabilities below:
The demo goes something like this: “A manager wants to see what a 5% decrease in expenses across a defined market and product will do to profitability. They ask the AI agent the question in natural language and the system automatically creates the resulting forecast”
While this capability exists and works, there are several challenges that prevent it from being a meaningful one.
As nice as the demo looks, the capability is not practical in real life. It may actually create more work than the benefit it provides. In legacy tools and processes, users leverage specific forms and templates to enter values for various drivers and assumptions. These are clear and concise if well designed. Writing the command instead of inputting the values looks nice in a demo, in real life, however, it is much more work. Especially if you have to go back and forth, change assumptions multiple times, create different scenarios etc. I have seen many people default into using legacy views for input during demos and meetings when time was of essence and we need to get things done fast. Perhaps this will change down the road with dictation and better language recognition, but for now it is just a cute demo that does not translate to practical efficiency.
This may be different for a manager at a high-level vs a planner. At that level the model gets grossly simplified, which again limits the business benefit.
Verbal commands create a lot of space for ambiguity. When you say “increase international revenue by 5%” are you sure the tool applies it to the right data? What if you have both an International region and a business unit that has “International” in its name? How do you know the agent will choose the right entity / context? How do you know users will use the terms consistently?
Using natural language increases the risk of errors. It also creates more validation and testing effort, which may outweigh any productivity gains. I have spoken to many clients who said two users gave their agents the same command and came up with different answers.
Inputting a number into a well-defined form with exposed context ensures data integrity. It is very fast, cheap and available with any tool on the market
The biggest question is how complex of a model are you able to apply the natural language command to.
Let’s take Workforce Planning as an example. Compensation is calculated based on people’s salaries. It includes merit and promotion increases, add joiners / leavers, transfers. Apply benefits and payroll taxes that vary by jurisdiction etc. Include employee-related expenses and various allowances. Possibly run allocations of resources to arrive at the P&L view. This application is likely implemented across several cubes or data stores with information flowing across the model and calculations happening in various stages.
A simple “Reduce Compensation by 5%” command will have a ripple effect across the model. Unless you are just making a top-side adjustment, which is way too simplistic, the number of data points and permutations that a simple command like this will trigger is way too big. The AI agent needs to have access to all this data, which it typically does not. It then needs to decide which of the many combinations of levers to pull to achieve the objective, run the calculation and decide which approach is best.
The only way to accomplish this with AI is to deploy a highly customized agent that will understand complexity of the model, can access all the data points and execute the operation. This is very difficult to accomplish. It is very far from vast majority of AI deployments that currently exist in this area.
Furthermore, AI is not the most effective means of accomplishing this. Any modeling implementation of this kind will involve some statistical modeling or prescriptive, “goal seek”, function, which has nothing to do with “AI”. We have customers using Monte Carlo simulations to arrive at a combination of possible outcomes and then analyze from there.
AI can be used to manage / orchestrate the process and perhaps do some autonomous development. The model itself, however, is arithmetic and deterministic. More on that in the “Modeling” section below.
Any true “natural language” driven forecast automation will be limited to a smallish model or simplistic algorithm. This may be useful for an executive as a high-level scenario analysis, however that puts us back to the realm of individual productivity vs a large-scale business process automation.
There is almost zero “AI” in any modeling that is currently done or promoted by vendor platforms.
The modeling capabilities that are currently presented as “AI” are really statistical models, libraries and algorithms that are tried and proven and existed for a long time. AI may help with certain aspects of the model, however the modeling itself and our ability to define and execute a highly efficient and scalable model is nothing new.
Most finance forecasts are a function of historical data / performance to the extent these are available at the right level of detail. These can be enriched by various business assumptions (growth, investment objectives, business strategy, pricing etc), macroeconomic factors or other variables that impact the business (weather, consumer sentiment, health trends etc). A forecast will typically use some level of historical data as the basis and then apply a set of the above factors and variables to arrive at the possible outcome or perhaps create several scenarios.
In most tools today, this is done using simple driver-based calculations or perhaps statistical models and libraries. Driver based models combine financial and operational data and include various business assumptions. Statistical models typically include time series analysis that incorporate trending, seasonality and cyclicality. It looks at correlations of various variables, uses moving averages of exponential smoothing to remove noise and more. There are hundreds of tools and libraries available for this, depending on the level of depth and sophistication that the company wants to get to. These are all proven and tested methodologies that have been around for a very long time and have nothing to do with AI.
There is a tremendous benefit and value to be derived from properly designed and deployed forecast model that is driver-based or relies on statistical methodologies based on relevant variables. Combining financial and operational data into a fully integrated model opens a world of opportunity to streamline, automate and improve the accuracy and depth of the forecast. There is tremendous business value that could be unlocked.
The areas where AI can assist here include orchestrating and managing the workflow, or perhaps bring these capabilities closer to the users that have a different way to interact with the system. Another opportunity is to mine the digital domain and use the findings as input into the statistical model that may not be available otherwise. Ai can help source data on a variety of topics that may not otherwise be readily available. Again, the users have to consider data accuracy and integrity, but the capability is there.
There are four main approaches that we have observed and will discuss here in more detail:
As of now, almost all companies and finance teams are using Copilot, Claude Desktop or other AI tool for their day-to-day work. These tools are well embedded in everyday workflow. They are widely available, integrate with MSFT Office applications (XLS, Word, PPT) and are accepted / deployed in organizations broadly.
However, while these are great tools and certainly provide benefits to finance, they are really addressing “personal productivity” objectives. They help the individual be more efficient but will not transform the function more broadly.
Pretty much all the use cases with these tools involve the same approach: loading data from an XLS file and having the AI generate financials, a summary, a projection, a dashboard, report or presentation. Access to these capabilities is easy given the XLS add-in and general integration into the MSFT stack (PPT mostly).
What makes this “personal productivity” vs large scale function automation is primarily the following:
The typical use case assumes that data already exists in XLS form, ready to upload to AI. But where does that data come from? In most companies, the data sits in a reasonably complex finance infrastructure that may include one or multiple databases, often a combination of multi-dimensional and relational data stores as part of the EPM infrastructure. There are data loads, mappings, calculations, business rules etc. Getting this data into a clean XLS file is not a trivial task. It often requires a lot of manual work: sprawling XLS files with vlookups, pivot tables—which is exactly where all the efficiency, risk and control challenges are. Unless we automate this part of the process— the access and retrieval of data— any benefit we gain from AI will be marginal. We will fail to capture the real opportunity in the long run.
The second issue is one of depth: there is only so much detail a standard XLS file can hold.
Spreadsheets will work fine for simple modeling and reporting. But the real meat and opportunity in FP&A isn’t the “WHAT”— it’s understanding and explaining the “WHY”. We need to understand the drivers of performance and the causes of variances, not just report what happened. And to do that effectively, we need to slice and dice the data along available dimensions and attributes, understand the rules or calculations (drivers of cost allocations, for example, or changes in HC, comp and benefits to understand impact on compensation). Often, we need to drill down to transaction-level detail to truly uncover the root causes and get appropriate insight.
This requires access to data that is well beyond the scope of a standard XLS file. Furthermore, the standard XLS file is static; it won't account for the fact that the drivers of performance shift with changing market conditions, so the data we need changes month to month. Without AI that can understand and access data in our entire finance ecosystem, pulling the relevant data points as needed, any upside we gain from AI will be very limited.
Every EPM tool or platform at this stage comes with a number of embedded AI capabilities. At least that is what is advertised. The capabilities vary and generally fall into one or multiple categories described in this document.
Bottom line here is that the capabilities, while they exist and work, are still limited and not very transformational. There are many different AI tools that provide a benefit in various tasks and areas, and they vary quote significant by vendor. However, these are still reasonably small and do not provide cohesive and game-changing automation.
This is a very fragmented space; it is difficult to provide a complete overview or generalize. But with every customer we work with or spoke to – most tried the EPM AI capabilities, they were not sufficiently impressed and that is why they default back into Copilot, Claude or perhaps supplement their work with AI at the Data Layer (next section).
As with everything here, this is evolving rapidly and I am sure will change in the future. Although I do believe some of the structural headwinds will remain.
We have seen significant progress in application of AI at the Data layer. This is the layer that most people refer to as Data Warehouse or Data Lake. It is where the data resides and is accessed / analyzed using various BI tools. The data tier also serves as the central repository and makes data available for consumption to downstream business applications, including EPM.
Perhaps because most data is here and readily available, it is the most natural point of entry for AI. The various vendors in this space made significant advances in enabling AI at this level. And since the tools are already approved and passed through the information security and compliance barrier, the AI tools are much easier to deploy.
Each vendor / platform has their own set of tools and frameworks. These include AWS Bedrock, MSFT Foundry, Google Gemini Agent etc. They are not quite apples-to-apples but generally all vendors provide their own ecoystem of tools. We see a significant use of Snowflake and Databricks in enterprise customers, which have their own AI agents and frameworks (Cortex and Genie, respectively). The AI tools at this level are used extensively by IT and various teams for operational reporting and the frameworks and capabilities are evolving rapidly.
Despite the widespread deployment, this is NOT the right approach or solution for FP&A due to a number of challenges:
While a lot of data is here, it is organized kind of mechanically, it is stored in tables or files with views available to users. While close, this is not how information is stored or used in FP&A. Data here is not organized to facilitate ease of reporting and analysis. As a result, any FP&A type use will be simplistic, complimentary to other use at the EPM level (drill down to detail etc) and will not provide the full picture
Most of the information and structures that are used in FP&A are not available at the Data Layer. This could include various hierarchies that are used for reporting and analyses. There are alternative hierarchies or custom views and aggregations defined and managed by users or groups. Calculations of any ratios and KPIs, cost or capital allocations for profitability management, assumptions that result in various views and scenarios will not be here. Therefore, any analytics at the Data layer will be sub-optimal and not be effective to satisfy the broader needs of FP&A.
Plan and Forecast will typically not be at the Data layer. These are done in the EPM tools, that can be reasonably complex and are dynamic (in that they may change often) to be copied over the Data lakes consistently. Thus, any comparison of actuals to plan / forecast, variance analysis at the Data layer will be very simplistic.
Despite the above, companies continue to try leverage the Data Layer AI capabilities to address the needs of FP&A. It is an obvious candidate given that the AI tools are there, they work and are available. A typical approach / progression we see is as follows:
This last item is problematic and fails to address the need for a number of specific reasons:
Any copy of data from EPM back into the data lakes will take time. It could be a long process the more complex and sophisticated the EPM model is. This will result in delays for FP&A to get the data, it will also affect the ability to copy the data frequently (budget / forecast updates and revisions etc), losing any type of real time access
Replicating data across data lakes and EPM leads to risks and issues from data integrity, it leads to reconciliation issues and breaks data governance rules. We saw many users get different answers to the same question from the infrastructure.
While the data can be copied into the Datalakes, it will miss all the relevant context – any assumptions, rules, calculations or other logic will not be there. This will negatively influence the ability to analyze variances and understand drivers of business performance. The data will sit there standalone, making any in-depth analysis of business performance impossible.
Any security constraints related to various data, versions, scenarios based on combination of dimension values will be very hard or impossible to replicate – which will impact how widely the AI agents at the Data layer can be deployed and used.
To create an effective AI Agent of Analyst at the Data layer, companies would have to essentially replicate the EPM application in the database. This is extremely inefficient, expensive and will likely yield negative returns. It is not what the data stores were designed to do and it will not be a business case that will ever work at a meaningful scale
A custom AI agent can address some very specific requirements and objective based on each company’s own process, technical stack and specifications. There are infinite ways to approach this. It is impossible for us to go into the depth of this here so we will not spend a lot of time here.
However, as of today, we have not seen anyone successfully deploy AI for Finance in a truly transformative way as it outlined at the outset of this article. Most companies we work with have live projects in flight, are at various levels of deployment and continue exploring various options. Most of them are tackling the challenges mention in this article and encountering the same roadblocks.
The investment here continues and we will revisit as things progress.
At ACG, our attempt at addressing the need of FP&A resulted in launch of our AI Financial Analyst, “whyFA.ai”. We are in various stages of implementing the AI Analyst at various companies. We are not far enough with the rollout at scale to declare victory, however we have seen the agent work successfully across a number of reasonably complex applications. We believe this is an approach that has the potential to transform some of the FP&A functions and processes. We will continue evolving the system as our efforts continue.