Coding

The next great software company won't sell software

A new breed of "service-as-a-software" startups—like LayerX—is dismantling the traditional SaaS model by embedding AI agents directly into enterprise workflows, charging per transaction rather than per seat. By abstracting away the software layer entirely, these companies monetize outcomes (e.g., automated invoice processing, fraud detection) while letting clients bypass licensing, integrations, and even UI. The shift threatens legacy SaaS incumbents by turning software from a product into an invisible, pay-per-use utility.

A new business model is emerging in enterprise software, one that bypasses the traditional SaaS subscription entirely. Instead of selling access to a tool, companies like LayerX are selling the completed work itself — charging per transaction or per outcome rather than per seat. The model, which Sequoia has termed the "autopilot," represents a fundamental shift from selling the tool to selling the result.

Overview

The core insight is straightforward: for every dollar spent on software, roughly six dollars are spent on human services — the work that the software couldn't finish. Reading emails, reconciling spreadsheets, drafting contracts, filing claims, triaging tickets. Traditional SaaS provided a dashboard that told you what to do, but a person still had to do it. Service-as-a-Software (SaaS, redefined) aims to eliminate that gap by having AI agents perform the intelligence work directly.

What it does

Instead of buying a CRM, you buy qualified leads delivered to your inbox. Instead of a contract drafting tool, you buy the completed contract. Instead of a support helpdesk, you buy resolved tickets — in multiple languages, with satisfaction scores attached. The product is the outcome, not the interface.

LayerX, an AI studio based in Portugal, describes its own work this way. When a staffing agency hires them, the client doesn't want a WhatsApp tool — they want their inbound candidate flow handled end-to-end: screening, qualification, scheduling. When a hospitality group hires them, they don't want a chatbot — they want guest communication handled 24/7 in five languages, with measurable conversion rates. The software layer becomes invisible.

Why now

The model depends on a distinction between two types of work: intelligence and judgement. Intelligence is rule-following — complex rules, sometimes thousands of them, but rules. Translating a clinical note into a code, drafting an NDA from a template, reconciling a bank statement. Judgement is taste — the gut call from experience, deciding which feature ships next or whether to take a deal.

For decades, software handled neither. It held the data while humans did both. Current AI models can now handle the intelligence layer reliably enough that companies can sell the outcome of that work and actually deliver it. The judgement layer still belongs to humans, but the intelligence layer — the expensive, repetitive work — is now addressable by code.

The wedge: outsourced work

The most practical entry point is work that is already being outsourced. If a company already pays an external accountant to close the books, three things are true: the work can be done externally, the budget is already a line item, and the buyer is already paying for an outcome, not a tool. Replacing an outsourced contract with an AI-native provider is a vendor swap. Replacing a full-time employee is a reorg. One happens in a week; the other takes a year.

The playbook, according to LayerX, is to find the small business already paying a bookkeeper a few hundred euros a month and offer to do it better, faster, and cheaper, with humans-in-the-loop for the messy parts. The wedge is the outsourced job; the long game is the insourced one.

Tradeoffs

Selling the tool puts you in a race against the model. Every six months, the model improves and your wrapper gets thinner. Selling the work inverts that dynamic: every improvement in the model makes your service faster, cheaper, and more defensible. The model getting better becomes your roadmap, not your obituary.

The moat in this model is not the code — anyone can write code, and code gets cheaper every month. The moat is proprietary data on what good work looks like in a specific vertical, accumulated one job at a time. Knowing exactly how a Portuguese staffing agency talks to a Brazilian candidate at 11 PM on a Tuesday, and what makes that candidate actually show up to the interview — that knowledge compounds faster than any competitor can replicate.

Bottom line

The next great software company, the argument goes, won't sell software. It will sell the answer. For founders, the practical advice is: pick a vertical you understand, find a job that is already being outsourced, sell the outcome of that work, charge for the work not for seats, use AI to deliver it, use humans for the parts AI can't do yet, and compound your data with every job completed.

Similar Articles

More articles like this

Coding 1 min

Show HN: Adam – An embeddable cross-platform AI agent library

A new embeddable AI agent library, Adam, has emerged, leveraging SQLite's database expertise to integrate machine learning models with relational data. By utilizing SQLite's query language, Adam enables developers to query and manipulate AI-driven data in a familiar, SQL-like syntax. This novel approach could simplify the integration of AI and data storage, potentially accelerating the development of AI-powered applications.

Coding 3 min

GLM 5.1 offers a low-cost alternative to Claude Opus for developers

Zhipu AI's GLM 5.1 is emerging as a budget alternative to Anthropic's Claude Opus 4.6, priced at $18 per month—three times cheaper than Opus. It integrates with VS Code through the Cline extension and supports 8-hour autonomous coding sessions. Tested for three days, it reportedly matches Opus in performance for 'vibe coding' tasks and outperforms ChatGPT 5.4 and Gemini. Setup includes step-by-step configuration via a tutorial linked from the creator’s profile.

Coding 3 min

Claude Code vs. Codex/GPT-5.5: When to Use Which AI Coding Assistant

Claude Code and Codex/GPT-5.5 excel in different coding scenarios, with Claude handling complex architecture and refactoring while Codex shines in UI, prototyping, and security testing. Token efficiency and task-specific strengths make both tools complementary rather than competitive. Here’s how to integrate them into your workflow for maximum productivity.

Coding 2 min

Optimizing Claude Code with Graphify

Graphify is a plugin that turns raw files into a knowledge graph, mapping connections between files and providing context to Claude Code without burning tokens. This solution was inspired by Andrej Karpathy, co-founder of OpenAI, and has gained 39K stars in a couple of weeks. By using Graphify, users can save 70 times on token per search and get better answers. The plugin creates a knowledge graph and preemptively creates semantic relationships between files, allowing Claude Code to understand the relationships between them.

Coding 2 min

Building a Second Brain with Obsidian and Claude Code

Obsidian and Claude Code can be combined to create a personalized knowledge base that helps with decision-making. This setup allows Claude Code to access a user's notes and data, providing more accurate responses. The process involves installing Obsidian, setting up Claude Code, and connecting the two. With this system, users can store notes, goals, and data, and Claude Code can help find connections between projects and provide answers based on the user's own knowledge base.

Coding 7 min

Claude’s Hidden Modes: Why One Chatbot Is Actually Five Tools in Disguise

Claude isn’t just a chatbot. Behind its familiar text box lie five distinct AI modes, each optimized for a specific workflow—from generating landing pages to writing code in plain Russian. Yet most users remain stuck in the default chat interface, unaware of the productivity gains locked behind these unadvertised tools. The revelation raises a question: Are we underestimating AI’s potential by treating it as a single, general-purpose tool?