Two Paths for Enterprise AI: Retrofit or Rebuild?

The next wave of AI transformation will reshape how businesses build and maintain their core software systems.

The Magic Has Arrived

Developers have already experienced their “magic moment” with AI. Tools like Cursor and Claude Code have fundamentally changed how code gets written—you can have a conversation with your IDE or terminal, describe what you want, and watch as AI iterates, catches its own mistakes, and reasons through problems alongside you.

But developers are just a fraction of knowledge workers. Most people are still clicking, typing, and navigating through software as they did before ChatGPT. The question is: when will the magic come to everyone?

The Enterprise AI Gap

Let’s take a very tangible example. Say you run sales operations at a mid-sized company. Your VP of Sales has told you that your company is redoing territory assignments this quarter.

First, you need to update the lead routing software, which funnels leads to the right sales rep in the CRM. Then you need to update the CRM to have the new territories configured. Finally, you’ll update the commission management software so that reps get paid correctly. Then you need to test it all. You can’t screw up lead routing or the company may lose a deal, and you can’t screw up payroll for your employees.

Most enterprise software needs to be customized like this. And these customizations can take weeks, months, or longer, depending on the complexity. Today, we’re still pointing and clicking our way through these updates. But the promise of AI is that we can “vibe code” our business processes. Updating territories in Salesforce should feel more like Cursor or Claude Code—conversational, fast, intelligent—and less like manual configuration.

Two Divergent Paths

This is the fork in the road for enterprise software: do we retrofit AI into yesterday’s platforms, or build entirely new, AI-native systems?

Path 1: Incumbents Add AI

Existing platforms could develop conversational interfaces where users describe desired changes in natural language and the system proposes modifications to workflows, custom fields, and business logic. Users would get an interactive experience where they could see proposed changes, accept or reject them, and quickly test their modifications.

The challenge of providing this experience on legacy software platforms is architectural. These platforms sit on complex foundations. Training AI to understand a typical Salesforce instance requires parsing custom fields, Apex code, Flow definitions in proprietary XML, permission structures, and countless other abstractions built over decades.

Path 2: AI-Native Business Systems

The alternative is more radical: building business systems from the ground up to be AI-native.

LLMs excel at understanding and modifying code and natural language. They struggle with proprietary configuration formats and domain-specific languages. Instead of teaching AI to understand Salesforce's XML abstractions, what if CRM logic were simply defined in TypeScript or Python? LLMs are good at understanding these, and they’re only getting better.

There are already early signs of companies doing this, letting knowledge workers configure business rules through a chat interface, which generates code, rather than clicking through endless configuration screens. Their products feel like Cursor, but they’re purpose-built for business functions.

The Competitive Battleground

This creates a real opportunity for startups. Incumbent platforms have enormous advantages: integration ecosystems, established communities, and switching costs that border on prohibitive. But they're also burdened by decades of architectural complexity that makes AI-driven customization genuinely difficult.

New entrants face a different equation. They can build workflow engines and business systems designed specifically for AI, but they face a competitive go-to-market landscape against both other startups and incumbents.

The Emergence Enterprise AI Framework

We believe the transformation is inevitable. In five years, we'll be having conversations with our business systems to configure and modify them. These systems will need new architectures that lean into the models’ strengths.

We're particularly interested in companies that:

1. Start with a wedge: Target a specific workflow or process that can be configured with a conversational interface.
2. Capture workflow data: Own the business logic and data as processes run.
3. Build toward systems of record: Use workflow ownership as a path to becoming the primary system, even if you start with just a subset of the incumbent platform.
4. Embrace code-first architectures: Enable customers to configure your software with languages and frameworks that AI models excel at (e.g. TypeScript, Python). Provide opinionated, typed SDKs in your platform that enable customers to string together pre-approved tools and actions.
5. Provide a new UX for code: Customers don’t necessarily need to see the underlying code. Novel UX’s should display workflow logic visually or in natural language to represent what’s being changed, while still allowing inspection of the underlying code.

The best opportunities will look like “Trojan horses.” They’ll start with specific workflows backed by code that gradually capture more of the underlying business logic and data, and eventually become the new system of record. The experience will feel like vibe coding, on rails. The opportunity is massive for those who capture it.