Do SaaS Founders Dream of System Prompts?
Talking about vibe coding, MCP servers, and ephemeral UI
For years, SaaS companies have been built as separate, self-contained products—each with its own database, rules, and interface. But now, AI-native startups challenging with a new stack are growing fast, reaching tens or even hundreds of millions in ARR in no time. In addition, with more tools at hand like Browser Use, MCP, and AI-powered IDEs and app builders, SaaS founders are asking: What’s next for them?
My hot take? SaaS as we know might collapse into middleware, reduced to a system prompt.
Compound Startup with AI
Software used to be fragmented, with each tool managing its own data silo. John Luttig’s analysis of Rippling showed how bundling multiple products around a central data source—like the employee record—let it scale fast by solving problems holistically. Deel did the same, reaching $800M in revenue in no time. The compound startup was born.
AI is taking this further. Rippling, Deel, or Personio built around a single function—employee records. AI isn’t tied to one function. It can coordinate workflows across an entire company, dynamically deciding what to do instead of following fixed rules. AI middleware acts like a quarterback—instantly deciding whether to fetch data, trigger services, or generate a temporary UI, all guided dynamically by a system prompt as its playbook. This fits with Software 2.0, a term Andrej Karpathy coined in 2017 to describe software that writes itself instead of being hand-coded.
Klarna is an early glimpse of this shift. They reportedly shut down Salesforce and over 1,200 SaaS tools, consolidating company data into a Neo4j database. Employees now access it through tools like Cursor. They reportedly cut data-query response times by half. Klarna’s AI story may be exaggerated ahead of its IPO, but the trend is real: AI-driven middleware is replacing rigid SaaS workflows with flexible, autonomous systems.
Interestingly, Palantir has been setting up companies for this paradigm shift for years with its Foundry platform, connecting data sources across enterprises long before AI became relevant. Without realizing it, Palantir positioned itself perfectly to benefit from this shift.
The application and business logic that once defined SaaS products and forced users to adapt their behavior is shifting to AI middleware. SaaS could shrink into a system prompt. It’s the next step in what Rippling popularized: bundling everything around a single source of truth, but now fully AI-driven.
The New SaaS Stack: System of Record + AI Middleware + Dynamic UI/Agents
Custom UIs tailored to user needs aren't new. Salesforce, with its army of consultants, built an entire industry around UI customization, and Retool made a business out of exactly this. According to Dylan Patel, there is no large SaaS platform in China, because its cheap but capable labor could produce custom software cheaply removing the reliance on larger platform players.
Now, AI is only an amplification of this. Instead of hiring expensive teams to configure CRMs, AI middleware generates interfaces on-demand at almost no cost, removing friction by creating exactly what's needed and when it’s needed—and nothing more.
This is especially disruptive for CRUD-based SaaS. Most SaaS apps today are frontends for basic CREATE, READ, UPDATE, DELETE operations. AI middleware makes most static UI flows largely redundant. For most jobs-to-be-done, rather than clicking through endless menus, users simply describe what they need, and AI dynamically fetches, modifies, and stores data. Sometimes AI even anticipates user actions, making changes proactively without ever showing a traditional UI.
The paradox of traditional SaaS is that it traditionally forces users to conform to the software’s rigid, opinionated, predefined workflows. It should be the other way around: software adapting to user intent, in real-time. (imv, opinionated software is great, but (1) can live inside a system prompt and (2) should get the job done without confusing or slowing down the user)
But it's not just the UI collapsing; application and business logic are dissolving too. Previously hardcoded workflows and conditional logic can now be expressed within a system prompt. AI middleware dynamically interprets rules and policies as data flows in, generating workflows on the fly. Of course, today’s AI middleware isn’t flawless—concerns around security, compliance, and predictability remain—but these issues are rapidly improving. Outside of cost and latency (which is trending to 0), there is no argument why most of business and application logic cannot be described, maintained, and updated as a system prompt.
Take Applicant Tracking Systems (ATS). Traditionally, companies like Greenhouse or Lever require hiring teams to conform to rigid software workflows. AI middleware flips this: hiring managers simply instruct, ‘Auto-screen candidates for Python skills, schedule two-stage interviews, and reject applicants without portfolios,’ and the AI adjusts workflows dynamically in real-time. The AI executes dynamically, adjusting workflows based on real-time feedback and given possibilities (e.g., tool use) and constraints (e.g., data security or access rights). The business process itself becomes a living system prompt.
This fundamentally rewrites the software stack. The old structure—database → hard-coded logic → frontend UI—now shifts towards a headless model:
data → AI middleware → multiple dynamic interfaces
Just as the 'headless' approach separated frontends from backends years ago, allowing multiple different UIs to be built on a single backend via APIs, AI middleware allows various 'heads'—it can spin up ephemeral UIs for transient interactions, maintain/update persistent, fixed interfaces where consistency matters (like analytics or reporting dashboards), or eliminate UI entirely by automating tasks invisibly within familiar tools like email, spreadsheets, and documents. It’s federated—pulling dynamically from existing tools, internal databases, or third-party services using emerging protocols like Anthropic’s MCP. Systems of record dissolve into AI middleware, which in turn abstracts away the underlying databases. And if you trust the AI quarterback in the middleware, why would you care where your data is stored, updated, or retrieved—as long as the job gets done?
Interestingly, companies that thrived in the original headless era—like Vercel on the frontend or Supabase on the backend—might again be perfectly positioned to ride this new AI middleware wave. (Curious readers should explore Supabase’s growing influence in the vibe-coding scene or check out Vercel’s AI launches like v0, AI SDK, and Generative UI.)
The future isn’t vibe-coding replicas of old-world SaaS (same goes for no-code tools that never delivered on this very promise)—even though that’s what trends on X/Twitter today. It’s about adaptive UIs that appear exactly when needed and disappear once the task is done—or no UI at all, with AI quietly augmenting workflows inside the tools
The playbook of AI-native challengers
Looking at these new AI-native challengers, they follow a clear playbook. They (1) wedge into existing workflows, (2) quietly bring AI middleware under the hood, and, once (3) they’ve built trust, become the system of record, the holy grail of all SaaS companies.
1. Start with the application layer—make adoption a no-brainer.
AI doesn’t have a technology problem—it has a UX and discovery problem. The best AI products solve this by not asking users to change their behavior. They embed into existing workflows, making the first interaction feel effortless.
Cursor initially forked VSCode, providing a familiar interface for developers while adding a simple AI chat on top. Granola entered a crowded space of meeting transcribers but stood out by deciding not to join calls as a bot, focusing on high-quality transcription, and letting users take notes alongside AI-generated ones—augmenting meetings instead of disrupting them. Langdock provides familiar AI interfaces like chat and an assistant/workflow builder while giving users flexibility on what model to use, avoiding vendor lock-in from the start.
2. AI-native products aren’t just UI layers—they operate as deep middleware from the start.
Cursor doesn’t just autocomplete generic code; it allows developers to RAG across entire codebases and external documentation. It can also spin up proactive and eager agents that generate entirely new files in parallel, fixing linting errors by itself until the user is ready to hit “Accept All.” Langdock connects with GDrive and Slack, not just pulling data but autonomously triggering workflows based on context. Granola doesn’t just transcribe—it lets users chat across past AI summaries, linking discussions into a structured knowledge graph that retains context over time.
3. Earn trust, then expand into the system of record.
Once users rely on these tools, they start feeding them more data. Cursor won’t just assist developers—it will become the go-to interface for writing, querying, and navigating codebases, integrating company-wide repositories as context if they aren’t already written entirely inside Cursor. Langdock begins by syncing with external data for individual use cases but could evolve into the unified abstraction layer for entire enterprise knowledge. And Granola? It won’t stop at transcriptions—it’s positioned to become an AI-native CRM or even your personal AI assistant for managing and recalling company-wide conversations.
What This Means for Traditional SaaS Founders
For founders building traditional SaaS, this shift raises new questions. If software becomes a system prompt, what happens to the way companies buy and implement software? Instead of adopting rigid SaaS applications, businesses may soon trade system prompts—where they can choose how they run HR, finance, or procurement simply by selecting the best system prompt. A company might implement "Stripe’s HR function as a system prompt" rather than buying yet another standalone HR SaaS tool. As a result, critical system prompts will be secured like API keys and passwords today.
This doesn’t mean existing SaaS players are obsolete—far from it. Verticalized CRUD apps, in particular, have an opportunity to adapt. By opening up their system of records (e.g., ERP, CRM, etc.) and exposing their services as APIs or MCP servers, they can still participate in the AI-native age. Instead of forcing customers into outdated UIs, they let AI-native companies build on top of their infrastructure, leveraging existing distribution and customer relationships. But realistically, most SaaS companies won't readily surrender control over their user interface and experience. We'll likely see a twofold strategy: opening core systems via APIs and MCP, while simultaneously offering proprietary ephemeral and fixed UIs, along with internal AI agents, to retain direct user engagement.
Do SaaS founders dream of system prompts? Maybe they should.
Many thanks to Christoph Deckert, Christian Eggert, and Judith Dada who patiently listened to my initial incoherent ideas found in this post and who gave me valuable feedback, and many thanks to John Luttig for the original thoughts around Rippling.
Love this! Feels like a great read for founders how to ride the next AI wave instead of getting left behind