Content author: Uvid ERP team — certified Odoo partner (ISO 9001 / ISO 27001, Bureau Veritas)
A word of advice to directors and owners: however tempting it is, do not vibe-code your next ERP. As a small-company director, I understand the attraction. This article explains where AI accelerates work, why ERP core is different and how to combine them without costly consequences.
Finally, a colleague who listens
An AI colleague has arrived who actually listens. Describe a need and it gets done, without meetings, waiting or “out of scope”. It works so fast that you become the bottleneck.
The initial excitement is almost addictive. You finally have help for tasks your team lacks capacity for: your own digital development agents.
This is not an argument against AI. That would be like arguing against Office ten or twenty years ago. We sell agents ourselves, narrowly specialised in knowledge relevant to each user.
But the jump from excitement to “let us build the entire ERP” is dangerously short.
What is vibe coding?
Vibe coding means describing software in natural language while AI generates and iteratively changes most of the code. Development follows visible results through prompts and rapid corrections, without detailed review of every line.
Three levels of AI-assisted development are often confused:
|
Approach |
How you work |
Risk |
|
Vibe coding |
Describe the goal and largely accept output if it works |
High if you do not understand or review the code |
|
AI-assisted coding |
AI drafts code, tests and repetitive parts; you direct design and review |
Reasonably manageable |
|
Agentic engineering |
An AI agent performs clearly bounded tasks within repository, testing and permission constraints |
Best for teamwork with safeguards |
Vibe coding is excellent for speed: a working prototype in hours instead of weeks or months. Trouble starts when the prototype is declared a product.
Where vibe coding makes sense
It suits low-cost mistakes, limited scope and results you can quickly verify manually:
- Prototypes and proofs of concept: small internal apps, client demos and clickable interfaces
- Low-risk internal tools: CSV parsers, report generators and data-cleaning utilities
- One-off automation: file transformations and scripts first run on data copies
- Extensions to existing systems: portals, dashboards and specific workflows
The common factor is that mistakes are inexpensive to rerun or correct. AI is excellent for these extensions.
ERP core fails that test because it is responsible for something unique in the business: the truth.
Why you should not prompt-build ERP core
It is not that tool — yet
AI cannot currently replace experienced teams building complex ERP on proven systems. Accounting entries, taxes, inventory, permissions, audit trails and corrections embody decades of expertise and edge cases already paid for by thousands of companies.
AI can build a car that looks good and drives around a car park. But the question is whether it brakes in rain at 130 km/h and who services it in five years. ERP is similar: visible bodywork matters less than the equivalent of brakes, ABS and certification.
The demo works. Then the real work starts
Claude or Cursor may produce an ERP-looking application over a weekend. Then it needs fiscalisation, e-invoicing and consistent accounting transactions for which you, as director, are responsible.
Can you trust the stock figures and values? Mature ERP has been tested for years. Is valuation correct and provable? Are figures traceable to financial statements? Do processes follow your preferences or actual business requirements? There is a major difference.
Then come cases a demo never sees: locked periods, mismatched accounts, duplicate bank transactions, or an e-invoice stuck after posting. Most importantly, who tested all this?
A vibe-coded system addresses each problem after it happens. Mature ERP such as Odoo models them coherently before they become incidents, ensuring events remain linked, controlled, auditable and financially correct over time.
A demo proves you can build a screen, not close a month accurately.
AI code looks right. Accounting asks whether entries are right
The biggest risk is not an obvious crash. It is plausible code with hidden flaws: weak validation, excessive permissions or incorrect tax logic discovered at month-end.
In Croatia, that includes fiscalisation, e-invoices, JOPPD, reverse-charge VAT, exchange differences and statutory reports. These change and need years of maintenance. With your own ERP, every regulation, security patch and user request remains yours. There is no community, documentation or upgrade path — just a chat history.
Code is not ERP’s most expensive part
Fast AI coding does not make ERP cheap. Forms, lists and dashboards have long ceased to be the main implementation cost.
The real work is defining process owners and rules, cleaning and migrating data, covering exceptions and corrections, designing permissions and segregation of duties, testing with users and maintaining legal compliance.
ERP projects fail through poor data, unclear rules and insufficient testing, not too few screens. AI removes none of those costs; it only makes them less visible initially.
Where AI belongs: around a proven core
At Uvid, we build agents daily and believe AI changes business software profoundly. We build around a proven core rather than replacing it.
Layer |
Examples |
Approach |
|
Core (system of record) |
Accounting, tax, inventory, permissions, audit trails and corrections |
Proven ERP, such as Odoo, with structured implementation |
|
Edge (surrounding layer) |
Portals, dashboards, import validators, workflows and AI assistants |
AI accelerates development; errors are reversible |
ERP is the proven system of record, backed by decades of expertise. Specialised AI agents provide intelligence around it, retrieving answers within minutes.
An agent is only as reliable as its structured data. An unreliable core produces unreliable data and AI. The sequence is clear: first a sound ERP implementation with a structured working methodology, then an AI layer delivering value. Our agents developed with FERIT follow that order: retrieve, propose and flag; people decide.
AI generates code that looks correct. Accounting asks whether it was posted correctly.
A final recommendation
Vibe-code extensions: portals, dashboards, validators and prototypes. AI brings speed, reversible errors and help where your team lacks capacity. We use and recommend it for those tasks.
Buy a proven business system of record, then build around it.
An ERP that cannot close the month accurately is not a saving. It transfers cost into incidents, manual corrections and lost trust in your numbers.
Frequently asked questions
Can AI write a complete ERP system?
It can generate ERP-like code, but cannot take responsibility for accounting accuracy, compliance, permissions and years of maintenance. A demo with test data does not prove accurate month-end closing.
What is vibe coding useful for in business?
Prototypes, low-risk internal tools, one-off automation and extensions around ERP: portals, dashboards and validators. If mistakes are easy to identify and reverse, AI can safely accelerate development.
How can AI work with Odoo without compromising data?
AI reads ERP data, prepares proposals and flags anomalies for human approval. It does not independently post entries, change stock or alter permissions. ERP remains the system of record.
Is building your own ERP worthwhile?
No. Processes, rules, migration and testing cost money either way. Custom ERP adds lifelong responsibility for localisation, taxes, audit trails and upgrades that proven systems already provide. Custom development makes sense for specialist modules around the core.
About the author
Marko Carević founded and leads Uvid d.o.o., a certified Odoo partner (ISO 9001, ISO 27001) with over 250 implementations in manufacturing, construction and food industries. Uvid develops ERP AI agents with FERIT Osijek. If you are considering digitalisation or business AI agents, get in touch for a no-obligation conversation.