How AI Tools Are Changing Python for Finance Professionals

How AI Is Redrawing the Efficient Frontier, and Why Python Has Moved the Most

By Gill Chowdhury, PyFi | March 2026
Building on research by Zachary Washam, "Should Any Finance Professional Learn Python?" April 2025

Introduction: Should Finance Professionals Still Learn Python?

This is the first in a series of articles and videos exploring how AI tools have changed the landscape for finance professionals. The series builds on our previous research, Should Any Finance Professional Learn Python?, which analysed 2,718 job postings across 42 major companies and introduced a tool map: a framework for understanding how Python and its alternatives relate to each other across two dimensions, scope of utility and interface abstraction.

The original research gave us an understanding of where Python and other tools sit relative to each other and revealed where and how Python is being used in the workforce. But it also opened up a series of new questions which have compounded as the implementation of AI in finance has become more and more ubiquitous. Now, a year later, I am returning to the original question: "Should any finance professional learn Python?" given the emergence of this new and important tool in the market landscape. This question has led me to think about what an AI implementation specifically entails, where "AI" sits on the tool map, and how it has affected the position of other tools as well as the specific instances in which someone should use one tool versus the other given this new reality. This first article is going to focus specifically on the tool map and how it has changed over the last year.

The Original Tool Map: Python vs Excel and Other Finance Tools

The original essay mapped Python and its most frequently mentioned alternatives across two axes:

Scope of utility is how broadly applicable the tool is. Tableau does one category of task very well. Python can be applied to almost any data problem a financial analyst might face.

Interface abstraction is how far the tool sits from the underlying computing machine. High-level tools (Tableau, Excel) are more convenient because they hide complexity. Low-level tools (Python, C++) are more powerful because they expose it.

Figure 1: The Original Tool Map
Figure 1: The Original Tool Map

When plotted, the tools fell along a diagonal. Tableau and Power BI sat in the upper-left (high-level, specialised). Excel and SQL occupied the middle. Python and R sat in the lower-right (general-purpose, lower-level). C++ sat furthest right and down: maximum power, maximum complexity.

This diagonal is not accidental. It reflects a competitive constraint: for a tool to survive in a market, it must occupy a unique position on the tradeoff between convenience and potential. A tool that is both more convenient and more powerful than an existing tool will displace it. The tools that persist are those that have found a defensible position on the frontier.

The original essay used the IKEA/Home Depot analogy: IKEA sells furniture kits (convenient, limited), Home Depot sells raw materials and tools (flexible, demanding). Both survive because they occupy different points on the same tradeoff. Python is Home Depot. Tableau is IKEA.

Figure 2: The Efficient Frontier
Figure 2: The Efficient Frontier

You can revisit the original article to see how and why we placed each tool where we did, here.

Aladdin's Lamp: The Upper-Right Corner of the Efficient Frontier

Over the past year as I have thought about how things are evolving, I kept coming back to the map. I was most specifically drawn to the upper right hand corner. This position represents God-like power. It is magical. You desire it and it is done. No friction, no other real constraints.

Aladdin's Lamp. That is what I called it. If you could ever find yourself here your ability would be almost limitless.

Figure 3: Aladdin's Lamp
Figure 3: Aladdin's Lamp

Part of the appeal or the hype around AI is that it represents some version of Aladdin's Lamp. And while that may sound like a hype statement itself, a closer look at AI in finance will clearly show us that it is materially moving the position of each tool, and the efficient frontier itself towards that corner, like a blackhole, sucking everything in.

How Finance Professionals Access AI Tools

This association to me is why the term "AI for finance" felt almost esoteric, or magical for a while. But I wanted to know what and how AI was actually being accessed and used in day to day operations and how it affected the other tools on the map. Would you still have to use Excel, or Tableau, or anything? Or could AI just do everything for you? To understand how AI has affected other tools on the map, I needed to first understand how AI was actually being accessed in finance. I identified three clear implementation categories, with a possible fourth to be revisited.

Natural language in an open environment

This is the most visible category: a finance professional opens ChatGPT, Claude, or Gemini in a web browser and asks a question. The interface is natural language. The environment is open, meaning the model draws on its general training data and whatever context you provide in the conversation.

This mode is powerful for exploration, drafting, and reasoning through problems. But it carries a fundamental constraint for finance: the model has no grounding in your specific data, your firm's policies, or your regulatory context. It generates plausible responses, not verified ones. Hallucination rates for leading models sit between 2% and 14% depending on the task.12 In most contexts that is tolerable. In financial reporting, where a fabricated metric or a non-existent regulatory reference can trigger compliance violations, it is not. Global financial losses tied to AI hallucinations reached $67.4 billion in 2024.13

The convenience is high. The governance is low. That tradeoff defines this access point.

Natural language via a focused, in-app deployment: AI Copilot in Excel

The second category is a natural language interface embedded within an existing tool. Microsoft Copilot in Excel is the clearest example for finance professionals. Here, the AI is scoped to the data and functions available within the application. It can write formulas, generate charts, and summarise data that is already in the spreadsheet.

This is a meaningful step up in AI governance from the open browser. The model sees what the tool sees, and its outputs are constrained by the tool's architecture. But the constraint cuts both ways: Copilot in Excel cannot reach beyond Excel. It cannot connect to an external database, enforce a firm-wide data standard, or chain together a sequence of operations across systems. Microsoft itself, upon launching the COPILOT function in Excel, explicitly warned users not to deploy it for "any task requiring accuracy or reproducibility."4 For finance teams operating in regulated environments, that limitation is significant.

The convenience remains high. The governance is better than the open browser, but the ceiling is set by the host application.

Programmatic access via Python and AI APIs

The third category is fundamentally different. Here, a finance professional (or a team) accesses AI models programmatically, through an API, using Python. This means writing (or having AI help you write) a Python script that calls the OpenAI API, the Anthropic API, or any other model endpoint.

The difference is control. With API access, you define what the model sees, the tools it can use, and the constraints on its output. You can feed it your firm's actual data. You can require it to cite sources from your internal documents. You can build validation layers, audit trails, and error handling around every response. You can run the same pipeline a thousand times with consistent behaviour.

This is not convenience-first. It requires more setup, more knowledge, and more deliberate architecture. But it is the only access point that offers genuine AI governance over what the model does and produces and as PyFi's understanding of AI evolved, this became the primary point we decided to focus on, at least for now. You can learn more about how AI can be introduced into your work programmatically by attending PyFi's new "Python Controlled AI for Finance Live Demo" here.

A possible fourth: agentic access

A fourth category is emerging: agentic AI, in which AI agents operate semi-autonomously across systems, executing multi-step tasks with human oversight at defined checkpoints. I mention it here for completeness, but I think it is more useful, at least for now, to understand agentic access as a combination of the three categories above rather than as a wholly separate mode. This will be revisited in a later piece in the series.

How AI Moves Every Finance Tool on the Map

As I thought about how AI was being accessed, it dawned on me that AI is changing the efficient frontier itself, and it is doing so unevenly. Microsoft Copilot, for example, allows someone who does not know VBA or specific functions to have them written and deployed in their work. The possibility for what Excel can do has not changed, but the accessibility has, which changes what this tool means in the market. It changes its position on the map.

AI is moving every tool toward the upper-right position of the map, which means the efficient frontier itself is shifting toward that position. AI tools do not simply occupy a new position on the original map. They do something more significant: they pull the efficient frontier toward the upper-right corner. Every tool on the map that can be augmented by AI gains scope, convenience, or both, and its position on the frontier shifts accordingly.

Figure 4: The Updated Tool Map — AI Era
Figure 4: The Updated Tool Map — AI Era

A note on how to read this figure: the tool positions on the updated map represent our assessment of where each tool now sits, informed by the evidence in this article but not derived from a direct quantitative measurement. The map is a mental model for thinking about tradeoffs, not a dataset. The direction of movement is what matters, not the precise coordinates.

Consider what has happened to each major category:

Graphical analytical software (Tableau, Power BI)

Copilot and similar integrations allow users to generate insights through natural language queries on top of existing data. The scope of these tools has expanded modestly: they can now do more than drag-and-drop visualisation. Their frontier has shifted, but not dramatically. The underlying architecture still constrains them to the visualisation and dashboarding layer.

Excel and AI Copilot limitations for finance

Excel now has multiple AI augmentation options: Copilot embedded directly in Microsoft 365, Claude available as an integration, and external AI tools used in conjunction with spreadsheet work. A finance professional who could not write a VLOOKUP can now describe the transformation they need and have it generated. Microsoft's own finance teams report that Copilot has compressed their FP&A weekly reconciliation from one to two hours down to roughly ten minutes, and saves approximately twenty minutes per account on accounts receivable reconciliation.3

That is meaningful. But the limit described in the previous section applies here directly. Excel's frontier has moved upward (more convenient) more than rightward (more capable). The AI governance constraint remains.

Python for finance: the biggest shift

Python's frontier has moved the furthest, and in the most important direction. AI augmentation has done two distinct things to Python's position:

First, it has made Python dramatically more accessible to finance professionals. The friction of learning Python (syntax, debugging, environment setup) has been substantially reduced by AI coding tools. A financial analyst who could not previously access Python's capabilities can now, with AI assistance, write and run Python programs they could not have written alone. This is not speculation. 25% of Y Combinator's Winter 2025 batch had codebases that were 95% AI-generated, and those companies were growing 10% per week in aggregate.5

Second, Python is the primary interface through which AI models are accessed programmatically. The OpenAI API, the Anthropic API, and every major AI model's programmatic interface is a Python-native experience. This extends Python's scope well beyond what the original map showed. It now includes the full capability of frontier AI models, applied with user-defined constraints, at scale. For FP&A teams, risk analysts, and anyone building AI automation in finance, this is the access layer that matters.

The key point: AI has not diminished Python's advantage. It has amplified it. Python became simultaneously more accessible (easier to use) and more valuable (it now provides a bridge to introduce AI into your work in a governed, controlled manner). No other tool on the map can make both claims.

Python vs Excel in the AI Era: The Uneven Shift

The frontier has not shifted uniformly. In our assessment, different tools have moved different amounts, in different directions.

Graphical tools (Tableau, Power BI) have gained scope through AI, but their fundamental architecture still constrains them. They have moved modestly.

Excel has gained meaningful scope through Copilot and external AI tools, but is constrained by the governance ceiling Microsoft has acknowledged. It has moved more than Tableau, but significantly less than Python.

Python has moved the most. It gained the most from AI augmentation because it was already a general purpose tool benefiting from strong and growing network effects. When you add AI to a general-purpose tool, the result is near-unlimited scope with meaningfully reduced friction. The distance Python has travelled is, in my assessment, greater than the distance any other tool has moved.

The counterintuitive implication: AI has increased the relative advantage of Python over its alternatives, not decreased it. The common argument, "if AI can generate code, why learn to code at all?", misses this. It assumes AI has moved Python and its alternatives equally. It has not. The gap between what a Python user with AI can accomplish and what an Excel user with Copilot can accomplish has widened, not narrowed. This arguably makes knowing Python significantly more valuable today, compared to a year ago.

Why Python Became the De Facto Language of AI

The shift described above is not a coincidence. Python became the de facto language of AI because of where it already sat on the tool map.

Consider the diagonal again. C++ sits at the far end: maximum power, maximum complexity. VBA sits near Excel: tightly scoped, high-level, locked to a single application. Python occupied a specific middle ground: general-purpose enough to do nearly anything, but with a syntax accessible enough that non-specialists could learn it. It was more approachable than C++ and more capable than VBA. That combination is precisely what AI development required.

When researchers at Google, OpenAI, and Meta needed a language to build and distribute machine learning frameworks, they chose Python. TensorFlow, PyTorch, scikit-learn, and the entire modern AI stack are Python-native. Python now powers 80% of AI agent implementations and commands a 22.6% share of the TIOBE Index as of January 2026, the highest of any language.14 Usage jumped 7 percentage points in a single year on the Stack Overflow Developer Survey, driven largely by AI development.6

This creates a compounding effect. Because Python is where AI lives, every advance in AI makes Python more valuable for finance professionals. And because AI coding tools make Python more accessible, more people can reach that value. The loop reinforces itself.

Vibe coding: what it means for finance professionals

Vibe coding, coined by OpenAI co-founder Andrej Karpathy in February 2025, describes using AI to write code from natural language descriptions.1 In practice, it means someone who has never written a line of Python can describe what they want and get working code back.

The results have been striking. Growth marketers, lawyers, and operations managers have used AI coding tools to ship working products without formal programming training.15 Founders launched MVPs in hours that would previously have taken weeks. By early 2026, 92% of US-based developers reported adopting some form of AI-assisted coding in their workflows.16

But vibe coding also revealed a sharp limit. Cursor's own CEO warned publicly that vibe coding builds "shaky foundations."8 A December 2025 analysis found that code co-authored by generative AI contained approximately 1.7 times more major issues compared to human-written code, with security vulnerability rates 2.74 times higher.17 Across 5,600 vibe-coded applications, researchers found over 2,000 vulnerabilities, 400 exposed secrets, and 175 instances of exposed personal data.18

For finance, this distinction matters enormously. Vibe coding is powerful for prototyping, for building internal tools, for accelerating time to a first working version. It is not, in its current form, a substitute for governed production code in a regulated environment. The accessibility gain is real. The security and quality risks it carries are also real. This tension between convenience and capability will be revisited later in the series.

The Scale of AI Adoption in Finance

Figure 5: The Scale of AI Adoption
Figure 5: The Scale of AI Adoption

GitHub reports that 46% of all code written by active Copilot users is now AI-generated, up from 27% at launch in 2022. 90% of Fortune 100 companies have adopted Copilot.7 Cursor, the AI-native code editor, hit $2 billion in annual recurring revenue in March 2026, doubling its revenue in three months.8

These are not finance-specific numbers. But the finance function is not immune. A 2025 McKinsey study found that 44% of CFOs were using AI in five or more use cases, up from 7% the prior year.9 A joint MIT/Stanford study found that finance teams using AI cut an average of 7.5 days off their monthly close, with a 12% improvement in report detail.10 Gartner's 2026 CFO survey found that nearly 60% of CFOs plan to increase AI investment in the finance function by 10% or more, even as expected headcount growth in finance collapsed from 6% to 2%.11

The implication is direct: finance teams are not growing. AI investment is. The finance professionals who understand how to govern AI, not just use it, are positioned differently from those who do not.

Conclusion: Why Python for Finance Professionals Is More Valuable Than Ever

PyFi's updated tool map shows that Python has become more valuable, not less. Should any finance professional learn Python? Probably. Depending on your exact role. It is more and more likely that it will be useful to you in the immediate and professional sense.

We know that a year ago about 10% of jobs surveyed preferred or required their candidates to know Python. And that was already a lagging indicator.

AI has pulled every tool's frontier toward the upper-right corner. But it has pulled Python's frontier the farthest. The tools that interact with AI most deeply, those that can govern it, constrain it, run it at scale, and audit its outputs, gain the most from AI augmentation. Python is that tool.

The question for finance professionals is no longer whether Python is worth learning relative to Excel. It is whether Python is worth learning as the foundation of a governed AI capability that no other tool on the map can provide.

The upper-right corner of the map, Aladdin's Lamp, is no longer theoretical. It is increasingly reachable. The path to it runs through Python.

References

1 Karpathy, Andrej. "There's a new kind of coding I call 'vibe coding'..." X (formerly Twitter), February 2, 2025.

2 Taft, Darryl K. "Vibe coding is passe. Karpathy has a new name for the future of software." The New Stack, February 10, 2026.

3 Microsoft. "Finance scenario library." Microsoft Adoption, 2025.

4 Sherr, Ian. "Microsoft launches Copilot AI function in Excel, but warns not to use it in 'any task requiring accuracy or reproducibility'." PC Gamer, August 23, 2025.

5 Mehta, Ivan. "A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated." TechCrunch, March 6, 2025.

6 Stack Overflow. "2025 Developer Survey." Accessed March 2026.

7 "AI Is Writing 46% of All Code: GitHub Copilot's Real Impact on 15 Million Developers." Medium, 2025.

8 "Cursor CEO warns vibe coding builds 'shaky foundations'." Fortune, 2026.

9 McKinsey & Company. "How finance teams are putting AI to work today." 2025.

10 "AI cuts monthly financial close time 7.5 days, MIT/Stanford study finds." CFO Dive, 2025.

11 Gartner. "Gartner Research Reveals CFOs' Budget Plans Prioritize Growth Functions, Technology and AI in 2026." February 10, 2026.

12 "AI Hallucination Report 2026: Which AI Hallucinates the Most?" All About AI, 2026.

13 "LLM Hallucinations: What Are the Implications for Financial Institutions?" BizTech Magazine, August 2025.

14 "Python's Reign Continues: Navigating the Top Programming Languages in 2026." Oreate AI, 2026.

15 "Vibe coding examples: Real projects from non-developers." Zapier, 2025.

16 "Vibe Coding in 2026: What It Is, Who It's For, and What People Get Wrong." Abhishek Gautam, 2026.

17 "The Reality of Vibe Coding: AI Agents and the Security Debt Crisis." Towards Data Science, 2025.

18 "Vibe Coding Security Risks: Why 53% of AI Code Has Security Holes." Autonoma, 2026.

Back to blog