The short answer
It goes wrong in the foundations. An assistant can explain a chart of accounts, how ERPNext handles stock or where permissions live. It cannot know how you price, ship, invoice and approve, or see your data. Early choices feel like simple settings, but once transactions sit on them they are slow and risky to change. Use AI to learn; leave design, migration and testing to accountable people.
What usually goes wrong
Patterns we see when an ERP is configured from general advice (forums, videos or an AI assistant):
| Area | Typical mistake | Why it hurts later |
|---|---|---|
| Company and chart of accounts | A generic or foreign template | Statements unlike your accountant's reporting; painful restructuring |
| Canadian sales tax | GST, HST, PST and QST as one rate, or rates hard-coded on items | Wrong tax across provinces; messy returns |
| Items and warehouses | Too many variants, or none; unrealistic warehouses | Inaccurate stock, confusing reports, item rework |
| Stock valuation | Method chosen without the accountant | Cost of goods will not reconcile; hard to change once stock moves |
| Naming series | Defaults clash with existing invoice numbers | Gaps and duplicates auditors query |
| Permissions | Broad roles for everyone "to get going" | Staff seeing payroll or margins; approvals bypassed |
| Data import | Straight into the live system | Duplicate customers and items, tedious to merge |
| Customization | Editing core code, not supported customization | Upgrades break; nobody knows what changed |
ERPNext handles this when configured deliberately: roles control access (role-based permissions) and data import inserts or updates records from CSV or Excel (data import). The tools are sound; the risk is how they are applied.
Why an AI assistant struggles here
It cannot know your processes
Invoicing on shipment or order, partial deliveries and approval limits change the setup.
Version drift
Screens, fields and defaults change between major versions; answers may describe another.
It cannot see your data
Duplicates, inconsistent units and missing tax codes show only in the real spreadsheets.
Privacy
Customer lists, employee records and pricing are what people paste into prompts.
The Privacy Commissioner of Canada recommends anonymized or de-identified information in prompts instead of personal information where possible (OPC).
A hypothetical example
A distributor selling in Ontario and British Columbia sets up ERPNext over a few weekends with an AI assistant. One sales tax template is applied to all customers, and stock is imported straight into the live company, with one product under three slightly different names. At the first quarter's tax filing, the bookkeeper finds BC customers were charged the wrong tax and stock valuation does not match the physical count. Fixing it means correcting months of invoices and merging duplicate items that already have transactions.
Rather talk it through? If you are partway through an ERP setup, or about to start one, we can review your configuration and data before transactions make changes expensive. Talk to a Promatics specialist
A safer way to use AI during an ERP project
AI helps when it supports people rather than replacing design work.
- Learn the vocabularyBefore workshops, so meetings are about your business.
- Draft processes"How we handle a return", for the team to turn into configuration.
- Clean spreadsheetsFormulas and cleanup, using de-identified sample data.
- Prepare test scenariosReal transactions your ERP must handle before go-live.
- Accountable decisionsAccountant: financial structure. Partner: configuration, migration, testing.
Setup readiness checklist
- Chart of accounts agreed with our accountant, not copied from a template
- Sales tax set for every province we sell into, tested on invoices
- Stock valuation method chosen with the accountant, before any stock moves
- Item master cleaned: one record per product, consistent units of measure
- Roles designed around jobs, with sensitive data restricted
- Data imported to a test copy first, reconciled with the old system
- Customizations made through supported methods and documented
- End-to-end tests passed: quote to cash, purchase to pay, month-end close
- A named person or partner responsible for support after go-live
When to bring in help
DIY can work for a very small, single-location business with simple tax and no inventory, ideally reviewed by an accountant. Bring in a professional if you hold stock, sell across provinces, have more than a handful of users, need approvals or integrations, or are migrating years of history. A structured ERPNext implementation costs less than a rebuild, with one team owning configuration, migration and support. ERPNext customization keeps extra workflows and reports upgrade-safe.
With the foundations right, AI has a real place: reading supplier invoices, preparing orders from emails, answering questions from ERP data. See AI in ERP.
Sources and further reading
Product capabilities and guidance change. These are the primary sources this article relies on, checked on the review date above.
- ERPNext documentation, Frappe
- Role based permissions, Frappe (ERPNext documentation)
- Data import, Frappe (ERPNext documentation)
- Principles for responsible, trustworthy and privacy-protective generative AI technologies, Office of the Privacy Commissioner of Canada
This article is general information, not legal, accounting or security advice for your specific situation. Examples are hypothetical unless stated otherwise.