Generative AI Adoption Strategy in FP&A
Merging the concepts from The Friction Project, The Crux and Ethan Mollick
It’s been a while since my last blogpost. Having a series to complete was good while it lasted, but then came Christmas and a new year and questions where to start again. That’s the thing with having a goal, achieving it and then trying to restart anew.
From fs.blog: Amateurs have a goal. Professionals have a system.
Of course there’s plenty of things happening that are good reasons why not to take the time to write, but it’s time to process some materials I’ve gone through into possibly one coherent new framework.
There’s four resources that I’ve explored in the coming weeks, which I’ll try to combine into one potential framework for generative AI adoption in FP&A. The resources were:
The Friction Project book by Bob Sutton
The Crux: How Leaders Become Strategists book by Richard Rumelt
Ethan Mollick’s article on 20 February 2024 titled Strategies for an Accelerating Future
HBR article on 26 February on Your Organization Isn’t Designed to Work with GenAI
Let’s first see what resonated with me from the recent resources and where do we end up by the end.
Concepts from The Friction Project by Bob Sutton
In "The Friction Project," Robert I. Sutton and Huggy Rao highlight both the beneficial and detrimental forms of friction. They advocate for the role of leaders as "friction fixers," who are able to diagnose and mitigate negative friction while fostering positive friction that encourages innovation and efficiency.
The book focuses on the concept of friction within organizations, distinguishing between positive and negative friction. Positive friction is seen as a necessary and beneficial force that can stimulate creativity, foster critical thinking, and drive progress. On the other hand, negative friction is identified as a destructive force that can stifle innovation, drain resources, and lead to employee dissatisfaction. What resonated me most here was the thinking around positive friction, as the common conception seems to be that all friction is bad or that we need to minimize the friction as an overarching goal. It’s not either or, it’s a balance that should be consciously managed.
The authors emphasize the role of leaders as "friction fixers", tasked with the responsibility of identifying and managing these forces within their teams and organizations. The goal is not to eliminate all friction, but rather to strike a balance, ensuring that positive friction is harnessed for growth and negative friction is minimized to prevent harm. This delicate balance can be the difference between a thriving, productive organization and one that is stagnant and unfulfilling. And like with all “balances”, there is no one true answer. In certain environments and circumstances we need to have more of the former and at other times the latter. Organizations should be able to learn and adapt not to become dogmatic in their search for the ultimate single outcome.
One of the intriguing terminologies introduced in the book as source of friction is "jargon monoxide". This term refers to the overuse of industry-specific language or jargon, which can often act as a form of negative friction. As a summary from the discussion, jargon monoxide can create barriers to communication, hinder understanding, and ultimately slow down progress. The key takeaway was that we all should be mindful of jargon creep, call it out and strive for clear, accessible communication to ensure everyone is aligned.
Through examples and methodologies like friction maps and root cause analysis, Sutton and Rao proposed frameworks to transform organizational friction into a force for good. Emotional intelligence, open communication, and a culture of feedback were brought out as keys to fostering an environment where positive friction thrives. Leveraging friction was the final idea that stuck with me to be used to benefit both the employees as well as the broader organizational goals.
In the context of FP&A, aside from understanding how friction in the organization can be both a force for bad and good, from the perspective of FP&A itself we could think here about inefficient processes and data silos that hinder our daily work as well as decision-making. Some friction might be good to step back for a minute and think, but its more prevalent that there’s more friction than optimal in our daily interactions.
Concepts from The Crux by Richard Rumelt
In "The Crux: How Leaders Become Strategists," Richard Rumelt develops his thesis on strategic thinking from the foundational book “Good Strategy Bad Strategy”. The central premise of this new book is that effective strategists and leaders focus on challenges rather than goals. This shift in perspective allows leaders to identify and address the root causes of problems, rather than merely dealing with their symptoms or chasing after objectives.
Rumelt introduces the concept of the “crux” of a challenge (from his background and interest in climbing), which is the pivotal point where action will have the most significant impact and promises the most progress. By identifying and focusing on the crux, leaders can make strategic decisions that lead to meaningful progress and success. This approach requires a deep understanding and acknowledgement of the challenge at hand, a clear vision of the desired outcome, and the ability to make coherent and decisive actions.
Rumelt brought out a three-part strategic skill set for mastering the crux. It includes i) the ability to discern which issues are of paramount importance and which are secondary, ii) understanding the difficulties inherent in addressing these challenges, and iii) maintaining a focused approach to avoid diluting efforts across too many fronts.
These skills are critical in identifying the strategic issues that require immediate attention and devising policies and actions to overcome them. He warns against common pitfalls such as mistaking symptoms for root causes, failing to prioritize effectively, and overlooking external factors, which can detract from a strategy's effectiveness.
Moreover, Rumelt underscores the importance of a systematic approach in identifying strategic issues, advocating for strategies that include data collection, cross-functional analysis, and both external and internal evaluations. This approach is essential to formulate precise problem statements that facilitate exploring viable solutions. By understanding their strategic landscape paves the way for more informed decision-making.
Takeaways from HBR’s Your Organization Isn’t Designed to Work with GenAI
The recent article from Harvard Business Review, "Your Organization Isn’t Designed to Work with GenAI," emphasizes the need for a strategic shift in how businesses integrate generative AI (GenAI) into their operations. Here are five main takeaways for an FP&A organization:
Redefine GenAI Integration: Unlike traditional automation technologies, GenAI should be approached as an assistive agent that enhances human capabilities over time, demanding a reimagining of business processes for full value extraction. Also here the adage of new technology on bad processes just mean faster bad outcomes.
Design for Dialogue: Adopting a new paradigm where technology and humans dynamically share responsibilities, with the capacity for GenAI to act more like a coworker than a static tool. This concept resembles the ideas about needing to treat the generative AI tools as probabilistic, not deterministic solutions. With IT we have been used to getting defined outcomes for our detailed instructions, but generative AI is more about a dialogue than expecting to get a defined repeatable single answer.
Implement Continuous Improvement and Efficiency Gains: By fostering a symbiotic relationship between humans and GenAI, organizations can organically evolve more efficient processes, capturing and analyzing each improvement for potential future automation. Feedback loops are key. Even more than with traditional software development, the final solution starts to decay from day 0 and faster than ever. Without constant feedback and learning built to the foundations, the outcomes can become obsolete or useless pretty fast.
Overcome Traditional Reengineering Limitations: Traditional business process reengineering methods, which rigidly assign tasks to humans or technology, are less effective with GenAI due to its flexible and iterative nature. A bit similar to the second takeaway of designing for dialogue, the solutions are more about the how and the why, not the what.
Practical Application Example: The case of Jerry, a customer service model, illustrates how breaking down tasks into knowledge domains and using GenAI can lead to significant improvements in response times and operational efficiency. As users get acquainted with generative AI capabilities in their daily life, they expect the same useability from the customer service solutions enterprises give them. The tolerance for old school chatbots is getting lower by the day.
Overall, FP&A organizations can leverage these insights by e.g. adopting a Design for Dialogue approach and through this improve forecasting accuracy, enhance decision-making processes, and increase operational efficiency. Sounds good, but generic, but that’s my feeling about the current phase of development - everybody needs to try out and learn what are their actual use cases. Embrace continuous improvement and adaptability so that FP&A teams can stay ahead in a rapidly evolving business environment so that we continue to work on the most accurate and up-to-date information available.
FP&A Generative AI Adoption Strategy
Ethan Mollick’s “Strategies for an Accelerating Future” brought out the following framework “Four Questions to Ask About Your Organization” that felt like a tangible item to continue practical work on.
So how can leaders start to think about the rapidly advancing nature of AI? The first thing they should do is use it. No amount of reading and research can substitute for spending 10 hours or so with a frontier model, learning what it can do. After getting familiar, companies should think about the following four questions:
What useful thing you do is no longer valuable?
What impossible thing can you do now?
What can you move to a wider market or democratize?
What can you move upmarket or personalize?
Combining this framework with the ideas from Friction Project and The Crux, the four questions in FP&A GenAI adoption could be interpreted as:
What useful thing you do is no longer valuable?
In light of Sutton's focus on removing friction, FP&A teams should first identify processes that are no longer valuable due to AI's capabilities. For instance, manual data consolidation and routine financial analysis may no longer be an essential part of FP&A's identity, as AI can automate these tasks with higher efficiency and accuracy.
This step involves critically assessing which parts of the traditional FP&A workflow are rendered obsolete by AI, such as certain types of financial modeling and reporting that can now be generated more creatively and insightfully through AI tools.
One should try out what work is really replaceable in FP&A, as “generating financial reporting with AI” may sound appealing, but today seems not realistic. Perhaps automating the method of getting faster to the outcome can be facilitated by generative AI, but actual report generation as a black box cannot be taken on. Hence, think more about the non-core tasks that one needs to complete that do not provide additional value (documentation). And be mindful about what the defines the value of you as an FP&A professional.
What impossible thing can you do now?
Reflecting on Rumelt's focus on addressing the crux, FP&A departments should explore new possibilities unlocked by AI. This includes leveraging AI for real-time scenario analysis, predictive forecasting, which were previously unattainable due to resource and technology constraints. I do want to see what tools will get us there, but it’s not “impossible” unless it seems out of reach or imcomprehensible.
The "impossible" now becomes feasible, such as providing every FP&A analyst with the equivalent of an "infinite number of interns" for data analysis, or equipping each team member with a virtual advisor for strategic decision-making. These all do need that FP&A professional step up their game and think how they can speak to the AI tools in a human language and not be as deterministic as they’re used to in Excel or SQL.
What can you move to a wider market or democratize?
Mollick's question about democratization is particularly relevant for FP&A. With AI, financial insights and analytics can be made accessible across all levels of the organization, empowering decision-makers and enhancing the overall agility of the business. This could mean deploying AI tools that provide customized finance advice to different departments, thereby enabling more informed budgeting, forecasting, and investment decisions without the bottleneck of centralized FP&A teams. FP&A should not and those need not be the gatekeeper of information in the organization. Be the facilitator, not the controller.
What can you move upmarket or personalize?
Finally, AI allows FP&A services to be personalized and extended to more sophisticated, high-value tasks. For example, AI can help small and medium-sized enterprises (SMEs) offer financial consulting services that were once the domain of larger firms, by providing deep, personalized insights into financial health, risk management, and growth opportunities. Accounting firms can step up the value ladder. FP&A teams in larger organizations can up their game in becoming more competent in the real business understanding. Time freed up from the manual work of before can be used to upskill and think how to partner your business stakeholders better. If you don’t move upmarket, somebody else will and disrupt you. Your stakeholders will demand more from you one way or the other.
Generative AI Fundamentals for FP&A
Before going to the conclusion on the AI adoption strategy, let’s recap what are some of the main concepts that I’ve recognized from myself in generative AI are about.
Large Language Models (LLMs)
LLMs like GPT (Generative Pre-trained Transformer) have revolutionized the way we process and analyze textual data. In FP&A, LLMs can automate the generation of financial narratives, interpret complex regulatory documents, and provide summaries of financial reports. This not only speeds up the analysis process but also ensures consistency and depth in understanding financial narratives. Be concious of data privacy, therefore learn how you can get what you need directly from the LLM or what you can build with the help of LLM to get to your desired outcomes faster.
Small Language Models (SLMs)
Small language models (SLMs) are compact versions of larger AI language models, designed to offer faster processing speeds and require less computational resources for operation. They excel in specific, constrained tasks, providing efficient and effective language understanding and generation capabilities while maintaining a balance between performance and resource utilization.
As LLMs seem to get expensive quite fast when used at scale, by customizing an SLM to understand the specific financial terminology and data structure of the company, these models can produce accurate and contextually relevant reports with less cost than an LLM.
Below is an illustration of language models by date of release and the size (vertical). The small language model threshold seems to be somewhere around low double digit billion parameters with some more known SLMs being Llama 2 7B, Mistral 7B & 13B, Microsoft Phi-2 13B, Zephyr 7B.
Fine-Tuning
Fine-tuning allows organizations to tailor generative AI models to their specific needs. In the FP&A context, fine-tuning LLMs or SLMs on proprietary datasets can enhance the models' ability to understand and generate reports or forecasts that align closely with the company's financial language and metrics. This customization might at the same time require significant engineering resources, therefore a ROI remains possibly hard to assess.
Retrieval-Augmented Generation (RAG)
RAG combines the power of generative models with external knowledge bases or databases, enabling the AI to pull in relevant information when generating text. For FP&A, this means financial forecasts, risk analyses, and scenario planning can be enriched with the most current data, ensuring that reports are both comprehensive and up-to-date. However, this means that there is a database to retrieve information from, which in turn means more training cost as well as longer time to retrieval.
Retrieval-Centric Generation (RCG)
RCG focuses on generating answers or content that is directly based on retrieved documents. In FP&A, this approach can be invaluable for conducting competitor analyses, market trend reviews, or regulatory compliance checks, where answers need to be directly traceable to source materials. RCG should enable better traceability and adherence to the source materials, which may lose some of the creativity of the language model, but enables better traceability and precision.
Generative AI Use Cases in FP&A
A practical application of Generative AI in FP&A could involve the automated generation of narratives in the regular reporting. The solution can draft narrative sections of the report, summarizing performance, highlighting trends, and identifying areas of concern. Fine-tuning ensures that the AI’s output aligns with the company’s financial terminology and reporting standards. RAG and RCG elements can pull in real-time market data and competitor information, providing a rich, contextual backdrop to the analysis. This approach could both improve time-to-market as well as enhance the strategic value of FP&A outputs, enabling more informed decision-making at the executive level.
Through a thoughtful application of Generative AI technologies, FP&A departments can be the key strategic advisor that has the timely, data-driven insight supporting the organization's financial health and strategic direction.
A GenAI Adoption Strategy for FP&A
Aiming to summarize the above into one single narrative, I’ve brought out three pillars to the strategy and then a short conclusion to wrap up the ideas for today.
Redefining Value through AI: The initial step involves a critical examination of existing FP&A processes to identify areas where traditional approaches no longer add significant value. This introspection, inspired by Sutton's focus on mitigating negative friction, enables teams to pinpoint tasks ripe for AI-driven automation or enhancement—such as predictive analytics and real-time scenario simulations. The objective here is not to replace human effort but to elevate the strategic role of FP&A by reallocating our resources to more complex, value-adding activities.
Embracing the 'Impossible': Echoing Rumelt's advocacy for focusing on the crux, FP&A teams must leverage AI to unlock new capabilities and address previously insurmountable challenges. This could be about democratization of our insights across the organization, thus fostering a more inclusive and informed decision-making culture. Think about access, not a need to present. The aim is to transform FP&A from a back-office function into a strategic curator that empowers stakeholders at all levels with financial intelligence (from BI to FI?). Do we think ourselves as needing to become closer to business or should we instead find ways how business can easily become closer to finance (or will we lose our identity then).
Strategic Democratization and Personalization: Mollick’s framework prompts a forward-looking approach to FP&A, advocating for the expansion of services both horizontally, by making sophisticated financial analysis accessible to a broader audience within the organization, and vertically, by offering bespoke, high-value financial consultation that leverages AI's predictive and analytical capabilities. Perhaps this learning will enable both the improvement in operational efficiency as well as that FP&A's contributions are tightly aligned with the organization's overarching strategic goals.
Navigating the Path Forward
The journey towards integrating generative AI into FP&A is iterative and demands a culture that values continuous learning, experimentation, and adaptation. It requires FP&A leaders to be not just technologically adept but also visionary in their strategic outlook—capable of seeing beyond the horizon of current capabilities to imagine what could be achieved with AI as a strategic partner. Coming back to the start of the paragraph, don’t worry about finding the perfect vision, just start and learn and define the vision on the go. Generative AI does not just enhance the efficiency, but also enables stepping up the strategic value of FP&A. A culture of innovation and continuous improvement will be the key enabler in the journey.
P.S. As a parting thought, the latest news on Croq providing answers at a pace not seen so far, seems to enable overcoming one of obstacles of AI adoption (slow responses) while offering this at a relatively good cost base. When the preferred model seems to be one in one day and another the next, how should the AI adoption strategy in FP&A take these developments into account? First, just start. Second, know that you’ll have to redo what you’re doing. Third, in one year you wish you would have started one year ago.
Thanks for now, I hope to get back to merging concepts in the near future!