Self-Driving ERP: An Autonomous Enterprise System Integrating Sixty AI Capabilities, from LLMs and RAG to AI Agents

This research presents the Self-Driving ERP, an autonomous enterprise system in which 60 AI capabilities, from large language models (LLMs) and retrieval-augmented generation (RAG) to AI agents, are built into ten ERP modules: Sales & CRM, Purchase, Inventory, Accounting, HR & Payroll, Projects, Manufacturing, Helpdesk, Reporting & BI, and a Developer Studio. Modelled on the components and six levels of automation of self-driving cars, the system senses business events, understands documents and voice, reasons with a shared pool of models, acts through the ERP's own transactions, stops at a risk gate when stakes are high, and hands control to a person with an explanation. The paper specifies a layered architecture (ERP core, perception, knowledge and memory, cognitive core, agent runtime, interaction layer, plus governance and operations planes) and provides an open-source reference stack on the Odoo framework with 24 reference implementations, covering hybrid RAG, cost-aware model routing, LoRA fine-tuning, demand forecasting, anomaly detection, invoice extraction, MCP tool servers, a risk-gated agent loop, PII redaction, a tamper-evident audit ledger, evaluation harnesses and deployment. Nine reproducible experiments on synthetic ERP and software-operations data test the core mechanisms: forecast error fell from 17.3% to 15.0% WAPE; fraud-style anomaly detection rose from 0.814 to 0.997 AUROC; a risk gate automated 67% of actions while leaving only 12.2% of error cost unreviewed, against 48.9% for a value-only limit; production incidents were detected 23 times out of 24 with five times fewer false alarms; and all 1,000 tampering attempts on the audit ledger were caught. A worked example estimates payback in under five months for a 40-person company. A comparison with SAP, Oracle, Microsoft Dynamics 365, Odoo, Zoho and ERPNext shows where the Self-Driving ERP is stronger (affordability, privacy, bilingual Arabic-English use and measured safety) and where commercial suites remain ahead. The paper also covers the history of AI and ERP, AI for software engineering, governance aligned with ISO/IEC 42001 and the NIST AI RMF, an implementation roadmap, the future of AI skills and a research agenda. It extends the author's earlier study of AI app builder platforms (https://doi.org/10.5281/zenodo.19280643). This work is a preprint and has not been peer-reviewed.