Podcast Atlas
source Episode summary Updated 2026-08-08 Tags: Podcast, Enterprise-Ai, Enterprise-Software, Erp, Saas

174: AI冲击企业软件巨头？与SAP原欣聊大模型to B的颠覆与边界

Summary

This LateTalk episode with [[YuanXin|原欣]] uses SAP and [[EnterpriseResourcePlanning|ERP]] to test the claim that large models, AI coding, and agents will disrupt traditional enterprise software. The discussion argues that AI can reshape interaction, implementation, pricing, and workflow automation, but it does not erase process rules, data governance, compliance, auditability, or industry know-how. Its main contribution to the wiki is the ERP Trust Moat frame: enterprise software is defended less by code volume than by trusted business-process substrate, structured data, global localization, and accountable human review.

Key Claims

  • [[EnterpriseResourcePlanning|ERP]] is presented as the backend system for people, finance, materials, procurement, suppliers, orders, revenue, and payment flows, not merely a visible software interface.
  • Feishu / 飞书 and DingTalk are framed as office-collaboration surfaces, while ERP, HCM, finance, procurement, and manufacturing systems carry deeper operational state.
  • AI has already worked in bounded enterprise use cases such as invoice recognition or OCR, but agentifying an entire enterprise system is harder because identity, security, integration, payroll, tax, e-invoicing, and global compliance all matter.
  • The episode qualifies AI Native SaaS Threat: Vibe Coding and AI coding can make temporary internal tools feasible for small companies, but listed, global, or audited companies need permissions, audit trails, trusted data, and external-consensus platforms.
  • SaaS Trust Moat becomes sharper in ERP: recreating code is not the same as recreating standard business processes, localization, industry rules, and operating credibility.
  • AI pressures seat-based SaaS pricing because agents may replace some human users; the source discusses consumption-based pricing and Result As A Service as alternatives.
  • [[SAPJoule|SAP Joule / Joule Work]] is described as the natural-language front end for intent recognition, agent dispatch, and self-built agents through a studio-like layer.
  • Autonomous Enterprise in this source remains human-in-the-loop: financial close, exchange-rate differences, bad-debt judgment, and data-error handling may be mostly automated but still need human review.
  • SAP’s model strategy is not to build a general foundation model first; the source says SAP works with Anthropic, Google, Microsoft, and Chinese partners such as [[AlibabaCloud|Alibaba Cloud]] and Qwen, while using SAP-specific structured-data models where appropriate.
  • The gap between model and enterprise application is reliability: 99% accuracy may be unusable in finance or compliance-critical workflows without reflection, correction, structured data, and responsibility boundaries.
  • Enterprise Operational Memory is treated as a prerequisite for enterprise agents: companies need business objects, ontology, history, standard workflows, unstructured records, and offline decision context before agents can act reliably.
  • [[ForwardDeployedEngineer|FDE]] work is framed as a hybrid of product engineering and business consulting rather than pure coding; it has to translate customer needs into automatable scenarios.
  • The source argues that OpenAI, Anthropic, Microsoft, and SAP are all exploring FDE-like enterprise deployment, but their strengths differ: model companies have coding/model capability, while SAP claims business-process knowledge.
  • For [[China|Chinese]] enterprises, China Enterprise AI System Debt is a recurring constraint: many firms optimized top-line growth before building strong information systems, data foundations, and software-payment cultures.
  • SAP’s China route includes [[AlibabaCloud|Alibaba Cloud]] infrastructure, Qwen model access, and joint exploration around FDE, post-training, and customer scenarios; the clothing-order example connects SAP ERP, DingTalk data, and an AI agent for order fulfillment and scheduling.
  • The episode ends by placing AI inside a longer technology cycle: natural-language interaction accelerates adoption, but industrial, organizational, and human diffusion still takes time.

Key Quotes

“模型即产品” — the source’s rejected shortcut for core enterprise workflows.

“人在回路中” — the source’s boundary for complex enterprise automation.

“从记录型系统变成可执行系统” — SAP’s described direction for agent-era enterprise software.

Connections

Contradictions

  • No direct contradiction found.
  • Productive tension with AI Native SaaS Threat: prior sources emphasize AI-native entrants, coding agents, and result-priced AI labor as threats to SaaS; this source agrees on pricing and interface pressure but argues that ERP-like systems retain strong defenses where data trust, auditability, compliance, standard processes, and global localization are core to the product.
  • Productive tension with stronger Agent Native Software claims: the source supports agent layers and thinner applications, but it does not expect agents to make enterprise systems disappear; instead it frames agents as a control and execution layer over governed systems of record.