News
Local AI found fake invoices and duplicates in a real accounting database
Asked a simple question about IT spending, the CLARYEL agent found fake invoices and duplicates in the ERP on its own and realised the total was wrong. It carried out the clean-up only after a person approved it, then re-checked the database. The main AI runs locally, on the Box.
Real work, real business data and a real result, running entirely locally. Scenario: ask, analyze, detect, act, verify.
A simple question was asked: “How much did we spend on IT components in the last year?” The system analyzed the ERP data, independently detected fake and placeholder records and duplicates, recognized that the initial total was wrong, carried out the cleanup after human approval and verified the database again.
What happened
- Ask: a business question in natural language.
- Analyze: Hermes queried ERPNext and analyzed the purchasing data.
- Detect: the system identified 64 fake invoices and duplicate records on its own, without being told which records were wrong.
- Act: it analyzed dependencies, ran into accounting constraints, corrected its approach and proceeded after human approval.
- Verify: it performed a final database check instead of simply reporting “done”.
The cleanup removed the placeholder invoices and related records, placeholder payments, duplicates, fake suppliers and customer, sales and warehouse placeholder data, plus about 240 orphan accounting records.
Result
Final database state: 188 valid purchase invoices totaling €38,627.70, zero fake records and zero orphan records. Corrected real IT spending for the last 12 months: €36,993.58.
The conversation was conducted in Italian, by voice, through Telegram. The same local environment is available in 100+ languages through Telegram, WhatsApp, a local browser, a computer, a phone, Home Assistant or a connected smart speaker. Business data stays local: it does not leave the office and is not sent to cloud AI services.
Technical stack
- AI agent: Hermes
- Inference engine: Strata
- Model: Qwen3.8-Flash-Next IQ3_S
- Business system: ERPNext
- Interface: Telegram voice; also WhatsApp, local browser, computer, phone and Home Assistant
- Data processing: fully local; no cloud transfer of business data
The next morning the same approach, a local agent with local AI, was applied to technical work on computer configuration. The goal is a local, private AI environment built from proven, ready-to-use solutions, with orchestration configured and everything kept as private and local as possible, on an encrypted drive.
What this means for a Box customer
- Questions about the business can be asked in plain language, by voice, in your own language.
- The agent checks the data before answering and changes nothing without approval.
- Accounting data stays in the office.