AI agents are starting to eat SaaS

Original Article ↗ Hacker News Discussion ↗

Role of AI agents vs SaaS: build vs buy

  • Many argue writing code is now “easy” with agents, but lifecycle is hard: upgrades, bugs, onboarding, security, and changing requirements still dominate cost.
  • Several commenters stress corporations buy SaaS primarily to mitigate risk and get accountability, SLAs, compliance, support, and a legal entity to sue—none of which agents provide.
  • Internal “vibe-coded” tools are compared to spreadsheets: fast and personal, but fragile, undocumented, and hated by everyone except their author.

Concrete uses of AI to replace / extend tools

  • One detailed story: using an AI assistant to discover a diff algorithm, wire up an open-source library, and build a custom HTML diff viewer with watch mode in an evening; contrasted with failing to get existing diff tools to behave as desired.
  • Some report canceling small, narrow SaaS (e.g. retrospectives, internal dashboards, Retool-like tools) after quickly rebuilding minimal equivalents with LLM help.
  • Others use AI to extend or customize open-source SaaS-alikes rather than adopt new commercial products.

Skepticism: economics and scale

  • Recurrent theme: AI-generated code still needs engineering, ops, security review, monitoring, and on-call. For most orgs, it remains cheaper to pay per-seat SaaS than to own 100% of maintenance.
  • Economies of scale: with SaaS you pay 1/N of maintenance; with in-house you pay N/N.
  • Several anecdotes of companies abandoning in-house systems for Jira/SaaS even when the internal code was “free,” because maintenance and feature demands overwhelmed small teams.

Where SaaS is likely resilient

  • Systems of record, high-uptime / high-volume systems, products with strong network effects, and offerings based on proprietary datasets or heavy regulation are widely seen as safe for now.
  • Vertical/“boutique” SaaS built on deep domain expertise and tight customer feedback is seen as hard to replicate by an internal dev + agent in a weekend.
  • Some expect AI to increase demand for SaaS-like integration, middleware, and niche vertical tools, not reduce it.

Data usage and trust in AI providers

  • Long sub-thread debates whether Copilot/Gemini/Claude train on enterprise or consumer data; some cite ToS and enterprise contracts as safeguards, others cite lawsuits, opt‑out policies, and “paraphrased data” as loopholes.
  • Consensus: enterprises must carefully read contracts and assume vendors will follow the letter, not the spirit, of data promises.

Long-term outlook

  • Optimists predict agents will eventually clone most software cheaply, commoditizing many generic SaaS features.
  • Skeptics note current agents are brittle, can’t reliably handle complex infra or business logic, and are more like very good IDEs than autonomous systems.
  • Many expect a split: large orgs and non-technical industries will keep buying SaaS; technical teams and indie builders will increasingly assemble bespoke tools with agents, raising the bar for flimsy, single-feature SaaS.