Karpathy Says Most Software Shouldn't Exist. What Does That Mean for Finance Teams?

Tuesday, 2 June 2026  |  By Timothy, CPA — Managing Director, Professional Financelink (PFL)

At Sequoia Capital's AI Ascent conference in April 2026, Andrej Karpathy — an OpenAI founding member, former head of AI at Tesla, and one of the clearest thinkers in the field — made a remark that stuck with me.

He described building a small app called MenuGen: you photograph a restaurant menu, and it renders images of each dish alongside the items. Useful. Clever. And then, within months, completely obsolete — because Gemini could simply take the photo and overlay the images directly. No app needed. "A lot of this code shouldn't exist," he said. "The neural network does most of the work."

That's a statement about software. But when I heard it, I immediately thought about finance teams.

Because finance teams have spent years building elaborate workflows, spreadsheet templates, reporting routines, and reconciliation processes. A lot of that work exists because the tools we had in 2018 couldn't do what we needed. The question worth asking in 2026 is: how much of what we built back then shouldn't exist anymore?

⚠️ AI & Data Privacy Reminder: When using AI tools to review or redesign finance workflows, avoid inputting identifiable client data, employee records, or commercially sensitive figures into public AI models. Use anonymised or generalised data for workflow design discussions.

Software 1.0, 2.0, 3.0 — and Why It Matters for Finance

Karpathy's framework is worth understanding clearly, because it reframes the whole conversation about AI in finance.

Software 1.0 is the traditional world most of us grew up with: humans write explicit instructions in code. Every behaviour is designed, debugged, and maintained by engineers. It's deterministic and predictable — and it's the foundation for every piece of accounting software, payroll system, and ERP that finance teams rely on.

Software 2.0 is where AI first entered the picture: systems trained on data rather than explicitly programmed. Machine learning models that learned to recognise patterns, detect fraud, flag anomalies. Still largely invisible to finance teams, but embedded in many of the tools used every day.

Software 3.0 is what's happening now. The programming surface is no longer code — it's context. You describe what you want, in plain language, and an LLM interprets and executes. Karpathy's framing: "You are no longer only writing deterministic instructions for a computer. You are giving context to an intelligent interpreter that can read, reason, call tools, inspect environments, debug errors, and adapt."

For finance leaders, this matters because it collapses the distance between someone who has the finance knowledge and someone who can build the tool. That gap — which used to require a developer — is shrinking fast.

The MenuGen Question for Finance Teams

Karpathy's MenuGen story is more than a clever anecdote. It's a diagnostic tool. The question it asks is: does this workflow exist because it's genuinely the best way to do this — or does it exist because the tools available when we built it couldn't do better?

Apply that to a typical finance team's monthly routine. The reconciliation process that involves exporting from three systems, combining in Excel, and manually flagging exceptions — does that exist because it's good practice, or because the systems didn't talk to each other in 2020? The board commentary template that takes two days to populate — is that format actually the most useful output, or is it an artefact of the tools we had at the time?

The honest answer, in most organisations, is a mix. Some processes exist for good reasons — they reflect genuine compliance requirements, audit obligations, or governance controls that can't be automated away. But a significant portion of the manual effort in a typical finance team exists because the tooling hasn't kept pace with what the work actually requires.

Software 3.0 doesn't just make existing processes faster. It makes some of them unnecessary. That's a different kind of change — and it requires a different kind of leadership response.

Dec 2025 The agentic inflection point Karpathy describes — when tools shifted from requiring frequent correction to reliably producing large chunks of useful output. "I can't remember the last time I corrected it."

83% Of finance leaders identify AI adoption as a key force reshaping their function, with 72% believing it will have significant impact within three years (Wolters Kluwer, 2026)

What This Looks Like in Practice — A Finance Perspective

I've been testing this thinking on my own work. Not theoretically — actually building things, running workflows through AI tools, and asking the MenuGen question: does this process need to exist in the form it currently does?

Some of what I've found confirms the hype. Tasks that used to take hours — drafting variance commentary, restructuring a reporting template, building a first-pass scenario model — now take a fraction of the time when approached with Software 3.0 tools. The knowledge required to do the work hasn't changed. The execution time has collapsed.

Some of what I've found confirms the limits. Processes that touch real compliance obligations, that require entity-specific context, or that depend on clean integrated data don't respond to a prompt. The constraint isn't the AI tool — it's the data and governance architecture underneath it. Software 3.0 is a powerful interpreter, but it can only work with what it's given.

The practical implication: finance leaders who are best positioned in this environment are those who can distinguish between "this process exists for a real reason" and "this process exists because we didn't have better tools." That distinction requires finance knowledge first, and technology curiosity second. Both matter. But they're in that order.

Note: The scenarios in this post are based on real experiences — mine and those shared by colleagues across the sector. Details have been modified slightly to protect confidentiality, and I've used a first-person perspective throughout for readability.

The Leadership Shift This Requires

Karpathy's analysis of AI job exposure ranked financial analysts among the highest-exposure roles — not because AI is replacing finance professionals, but because so much of what a financial analyst does in 2026 can now be performed, at least in draft form, by an LLM given the right context.

The useful distinction is that exposure is not the same thing as replacement. It means AI can perform significant components of the work — draft analysis, summarisation, modelling support, structured output generation. The human value shifts toward judgement, verification, and deciding what matters — using AI to amplify capability rather than competing with it on raw output speed.

For finance leaders, this translates into a specific kind of leadership shift. The value of a finance function in a Software 3.0 environment isn't in producing outputs — it's in knowing which outputs matter, what questions to ask, and how to verify that the answers are reliable. The judgment layer is where human expertise becomes irreplaceable.

That's a different job description than what finance teams were hired for ten years ago. It requires finance professionals who are comfortable directing AI tools, reviewing and challenging AI outputs, and designing the governance structures that keep automated processes accountable. This is less about coding and more about understanding what good looks like — and holding the standard.

RELATED ON FINANCE INTELLIGENCE

We built a ChatGPT agent for a finance workflow — and the most important lesson wasn't about the technology. Read: We Built a ChatGPT Agent for Finance. Here's What We Learned About Safeguards First →

The Question Finance Leaders Should Be Sitting With

Karpathy's MenuGen example is memorable because it's concrete. An app that made sense to build became unnecessary in months — not because the problem went away, but because a better tool solved it at a lower layer of the stack.

Finance teams should be running that same audit on their own workflows. Not as a cost-cutting exercise, and not as a technology project. As a leadership question: what are we doing because we have to, what are we doing because we chose to, and what are we doing because we haven't yet asked whether we still need to?

The finance leaders who are going to be most effective in the next few years aren't necessarily the ones who know the most about AI. They're the ones who know their workflows well enough to ask the MenuGen question honestly — and are willing to act on the answer.

Rethinking how your finance function is structured?

PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations. If you're asking the right questions about what your finance team should look like in 2026, we should talk.

Talk to PFL →

Timothy, CPA

Managing Director of Professional Financelink (PFL). 20+ years in finance leadership across NFP, NDIS and SME sectors. Interested in the intersection of finance practice and AI — not theoretically, but from actually building and using the tools.

SOURCES

• Sequoia Capital AI Ascent 2026 — Andrej Karpathy Fireside Chat (April 2026)
• Karpathy — 342 Jobs Ranked by AI Exposure Risk (2026)
• Wolters Kluwer — 2026 Future Ready CFO Report (via FutureCFO)
• FutureCFO — Using AI to Reshape the Finance Function in 2026

Tomorrow on Finance Intelligence: Property finance is one of the last holdouts of the spreadsheet era. Here's why — and what's finally starting to change.