Natural-language analytics looks simple in a demo. Ask a question, generate SQL, return a chart. In a real ERP environment, the hard part is rarely writing syntactically valid SQL. The hard part is choosing the right tables, joins, filters, date definitions, business meaning, and data scope before SQL is written.
This is why generic text-to-SQL tools often disappoint operators. They can produce polished answers against the wrong context. The answer looks confident, the chart looks useful, and the business decision is still wrong.
The 10,000-table problem
ERP and source-system environments are full of legacy tables, technical names, custom fields, regional variations, archived records, and business rules that live in people's heads. A sales question may require finance definitions. A receivables question may require customer hierarchy. A production question may require plant, material, batch, and posting-date nuance.
If the retrieval layer picks the wrong context, the model can still write beautiful SQL. That is the dangerous part. The system must first narrow the world to approved tables, known joins, business definitions, saved queries, schema documentation, and user permissions.
What governed ERP answers require
- Approved data scopes for each team, so users cannot wander into sensitive or irrelevant data.
- Business-language expansion, so a question like overdue invoices maps to the right ERP vocabulary.
- Visible SQL, so data owners can inspect what actually ran.
- Validation question sets, so the pilot is judged against known answers before rollout.
- Audit logs that show the user question, retrieved context, generated SQL, result, and response.
The value is not replacing analysts
The best ERP answer layer does not remove the data team. It removes repetitive interruptions. Analysts still own definitions, review complex questions, improve context, and govern access. Business teams get faster answers to recurring operational questions that should not require a queue every time.
Where SemeLabs fits
SemeLabs is Dotnitron's governed ERP operational answer layer. It is built around context retrieval, visible SQL, approved data scopes, and validation evidence. The goal is not a magic chatbot over a database. The goal is a trusted answer workflow for one team, one scope, and one recurring question set before expansion.
Research notes and sources
- IBM Institute for Business Value frames agentic AI as a process automation shift for intelligent business operations: https://www.ibm.com/thought-leadership/institute-business-value/en-us/report/agentic-process-automation
- McKinsey's 2025 AI survey emphasizes workflow redesign and enterprise scaling challenges: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- Gartner predicts growth in task-specific agents inside enterprise applications, reinforcing the need for controlled workflow orchestration: https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025