Why ChatGPT is the new Excel — and what it means for software

One of Xerox PARC’s early PCs, Xerox Daybreak.¹

In the 1970s, Xerox PARC’s pioneering inventions- object-oriented programming, GUIs, word processors, image processing, and even advancements in NLP to power spell-checking — laid the foundation for the PC revolution. Today, we’re on the brink of another seismic shift, powered by large language models (LLMs), that’s poised to redefine software, just as PARC’s innovations did for personal computing.

The first killer app for personal computers was Excel, a software program so essential at the time that people would buy a computer just to use it. Excel introduced a new paradigm of computer programming, teaching fundamental concepts like data types, formulas, functions, and more. Today, Excel is arguably the most impactful software of the last 50 years, used by billions of people globally.

Excel’s universality is a testament to its simplicity and utility. First-time Excel users quickly grasp basic features like data entry, cell formatting, and simple calculations, all with minimal training. However, with more proficiency, users discover Excel’s true potential. Power users can model complex systems, automate workflows with macros, integrate external data, and even build full applications using Visual Basic. Excel is generally useful to most, but it’s true power is only unlocked with mastery.

This paradigm is analogous to today’s LLMs: generally useful, incredibly powerful, yet requires mastery. OpenAI’s ChatGPT or Anthropic’s Claude² is easy for a first-time user — arguably easier than Excel. If a user knows how to read and write, they can simply type in their questions or statements using natural language and ChatGPT provides relevant, coherent responses. Its conversational interface is intuitive, mimicking everyday activities like texting. As a user explores ChatGPT further, they discover its depth of capabilities and potential for more advanced applications.

Mind the (Prompting) Gap: The Gap is Too Large

But mastering ChatGPT is also no small feat. Like Excel, it requires deep understanding of inherent capabilities and limitations. Users must learn the most effective ways to interact with the LLM to achieve their desired outcomes: crafting precise prompts, supplying the right data, fact-checking responses, iterating to improve response quality, staying up-to-date on the latest model releases, and more. Mastery is an eternal treadmill.

LLMs are theoretically incredibly powerful — but why is it so hard to get them to do what you want? The bar for deriving value from LLMs is way too high.

Every company has its 4% of employees that are experimenting with ChatGPT to boost their productivity. While it’s inspiring to see this 4% of employees’ herculean effort to work for their use case, it’s a PITA. The real opportunity lies in making this technology valuable to the other 96%. How do we make this happen?

Context is King: The Push for AI-First Domain-Specific Applications

One of today’s burning questions is whether horizontal platforms, like OpenAI’s ChatGPT, Anthropic’s Claude, or Perplexity, will gatekeep industry-specific or use case-specific applications built on top of their platforms or whether specialized applications circumvent these gates. The problem with the first argument is that it orients around the technology — not the customer. Let’s be real: most customers don’t care about technology paradigms for their own sake. They just want their problems solved, and they want it done yesterday.

Products like ChatGPT and Claude have shown millions of users the power of LLMs. Similarly, Perplexity showcases the magic that happens when you marry LLMs with vast amounts of public information available on the internet. But here’s the kicker: while a general-purpose platform like ChatGPT or Perplexity may solve 20% of a user’s problems, a specialized, AI-first application can solve 50% right out of the gate. Over time, this number could rise to 60%, 70% — maybe even 100% — as the application continues to learn from users and underlying capabilities advance³.

From a customer’s perspective, it’s a no brainer: if a specialized, AI-first application can deliver more value in less time and require less effort to integrate into existing workflows, why would they ever consider a general-purpose platform? People are inherently lazy and, when it comes to adopting new tech, they’ll opt for the path of least resistance.

Just as countless software companies have been built on the premise of “replacing Excel”, a new generation of software companies is emerging on the premise of “replacing ChatGPT.” We can expect a new generation of specialized AI-first applications, each one focused on specific industries or use cases and tailored to the unique needs of their target users, built on the foundation of LLMs.

David vs. Goliath: Why Incumbents Will Fumble to Win the AI-First Future

So why won’t incumbents capture all of the value from new AI-first applications?

First, let’s define what we mean by “AI-first.” When you’re building AI-first, you’re not just slapping AI-powered features onto existing products. Instead, you’re centering AI in the design process, crafting entirely new, AI-centric product experiences that redefine how users interact with software and do their work.

Incumbents will capture — and have already started to capture– value by adding AI-powered features to existing products. Just listen to the earnings calls of companies like Salesforce and ServiceNow, that hype up foreseen revenue growth from their new AI-powered products. As history has shown, the power of incumbent’s deep pockets and distribution networks should not be underestimated — cue the epic showdown of Slack vs. Microsoft Teams.

However, incumbents face the classic innovator’s dilemma: while they see the potential of AI, they’re also weighed down by legacy systems, entrenched business models, and the need to keep existing customers — wedded to existing products — happy. This dynamic is particularly pronounced in a market where technology is evolving rapidly and promising AI-first product opportunities are not yet obvious.

So while incumbents may be quick to add AI-features to their existing products, they’ll likely struggle to create truly groundbreaking AI-first products. Where exactly do I think they’ll miss the mark? AI-first products that unlock entirely new workflows.

What do I mean by new workflows? There’s a body of work that’s currently done by a mashup of human cognition and a smattering of tools like Google searches, Word docs, LinkedIn searches, private database searches, and more. LLMs have the potential to bring these workflows online into purpose-built software that can complete a task 10x faster or synthesize information 10x more intelligently than a human can. These new workflows, powered by the non-deterministic logic and reasoning capabilities of LLMs, are not just innovative — they literally weren’t possible to build two years ago, because the underlying technology didn’t exist.

Let that sink in: you couldn’t build them two years ago. It’s incredibly exciting! Software is no longer about incremental improvements to the last Excel workflow that hasn’t been converted into purpose-built software. Software is no longer just a dumb database with a lightweight CRUD interaction on top of it. With LLMs, software can now understand context, make judgments, and synthesize information in ways that were once exclusively the domain of human cognition.

The true power of LLMs lies in their ability to bring new workflows online, expanding the surface area of work that software can power. Software markets just got a lot bigger.

Revolutionary — Not Evolutionary: Why AI-First Products are a Perfect Fit for Startups

The question still stands: why are startups best-suited to build AI-first applications? Three key factors make startups uniquely equipped to create winning AI-first products:

1. Deep Understanding of Users’ Work:

AI-first products don’t just assist — they replace — chunks of human cognition, fundamentally renegotiating the division of labor between humans and machines. To strike the right balance, product builders must mind-meld with users to understand where AI can deliver the most value and how to craft an intuitive product that makes “artificial” intelligence feel natural.

This requires deep domain expertise and ability to empathize with users’ needs and workflows. Founders with firsthand industry experience have a leg up here. For example, take the former hedge fund portfolio manager who built a finance LLM on gardening leave, then launched AI-first software for hedge fund analysts. This founder brings unique and nuanced insights into user pain points and opportunities for AI to move the needle, resulting in an intuitive product. While direct domain experience isn’t a must for founders, cracking AI-first products requires a new level of mastery of solving user problems.

2. Willingness to experiment and take risk

For AI-first applications, there’s a lot to figure out in the UI/UX, not just the underlying technology. How does a human oversee an AI-application? What are they looking at? How do they nudge the AI in new directions or course-correct when things go awry? It’s clear that there’s a future world where AI agents  — not humans —  perform actions in software, but it’s unclear what UI/UX will facilitate these interaction. As human oversight ascends to higher levels of abstraction, the UI/UXs we know and love today will transform.

By starting with AI at a product’s core, startups can create new user experiences that redefine our interactions with software. For example, the Cognition team’s AI-software engineer Devin provides one lens into this future. Startups have the freedom to take bold risks, pivot quickly, and relentlessly iterate, embracing the fact that, while most product experiments will fail, the few that succeed could revolutionize their industry.

3. Focus and depth

Users have a low tolerance for inaccuracies and expect AI-first products to “just work” with minimal effort. When building AI-first products, focus is particularly important as it allows start ups to build trust with users by delivering a core product that solves an specific problem exceptionally well before expanding to adjacent use cases.

AI-first products often appear simple on the surface, driven by a handful of deep, AI-centric features. However, beneath the surface lies a complex system of agents with different skills, powered by different models, working together. The real magic happens when an AI-first application masters a single use case and then captures adjacent workflows. As the AI-first application continues to learn and improve based on user interactions and feedback, this self-reinforcing system becomes a powerful moat.

You might be thinking — how are these factors different from the recipe of pre-LLM startups? The truth is that the fundamentals haven’t changed — but AI amplifies these principles and takes them to the next level. The recipe remains the same: build a kickass product that solves a burning user problem and find a cost-effective way to get it into the hands of users — but there’s a new superpower available for building products and a new fog of war to navigate.

But let’s be real — in a world where technology is moving at breakneck speed, everything we’ve just discussed could be irrelevant tomorrow. #AGI

1. This photo of my dad will never get old. He worked at Xerox PARC in his 20s and helped build a number of key technologies for the PC era.

2. Spark Capital led the Series C of Anthropic.

3. I suspect that these advancements in the underlying models across text, video, image, robotics, and will happen much more rapidly than we expect.