Published
8 MIN READ
ERP AI Integration: What to Prepare First
ERP AI integration means connecting artificial intelligence to the system that holds orders, inventory, receivables and invoices, so it can look things up, prepare drafts and, within a limit, write records. The work that decides the outcome happens earlier, inside the ERP itself: the right version, a technical user just for the AI and the tax layer in order.
In 2026, that preparation got more demanding in Brazil. The tax reform is changing invoices throughout the year, and an AI that suggests an entry or a draft invoice depends on fields whose rules changed in August. Anyone who connects AI without looking at this is automating on top of a foundation that is still being replaced.
At Wama, this kind of project starts with the AI assessment for companies, the stage that comes before any connection. The overview of how spreadsheets, CRM and ERP talk to AI is already on the blog; the focus here is what to prepare inside the ERP.
What is ERP AI integration?
It's giving artificial intelligence controlled access to the ERP, through the interfaces the system itself offers, so it can work with data that already exists. It isn't switching ERPs or asking the model to learn the company's accounting. The ERP remains the system of record: it applies the inventory, billing and tax rules, and the AI works around it.
In practice, there are three levels. In lookups, the AI answers how much is in stock or which receivable is overdue, and nothing changes in the system. In drafts, it puts together an order, a draft invoice or an expense classification, and a person confirms. In limited writing, it records on its own only what falls below a ceiling set by the process owner.
The protocol that connects the model to tools is a separate topic, covered in MCP for companies. With an ERP, the question that comes before the protocol is simpler: what does this system, in this version, let the AI read and write?
What should you check in your ERP version?
Check the interface your installation actually exposes, not the one in the vendor's sales material. The same ERP can have a modern API in the cloud and a much more limited path in the version installed on the company's own server. Saying the system has an API is an incomplete answer: what matters is whether it reads and writes the objects your flow uses.
Bring four questions to whoever manages the system:
Which objects does the API cover in our version: order, customer, product, receivable, invoice?
Does it allow writing those objects or only reading them?
Is there a test environment separate from production, with data similar to the real data?
What call limits, maintenance windows and license rules apply to an integration user?
When the API doesn't cover what the flow needs, three paths are left: file exports, an intermediate layer that exposes only the authorized operations, or custom software alongside the ERP. When to build is described in custom software with AI.
Technical user: the AI inherits what it can do
The AI sees and does in the ERP exactly what the integration user can see and do. That's why the most important preparation is creating that user from scratch, with minimum permissions, instead of reusing the login of someone in finance or, worse, the administrator's.
TOTVS's Protheus documentation on TDN, updated on January 29, 2025, shows how this works in an ERP widely used in Brazil: the REST API issues an access token whose context depends on the user who authenticated, valid for one hour. The token doesn't create new permissions. It carries those of whoever logged in.
In Wama projects, every AI output gets a human read before it reaches the client. The same standard applies near the ERP: the technical user starts with read-only access, gets draft permission once the reading is right and only then receives writing limited to specific operations.
Why does the tax reform matter here?
Because the tax side of the ERP is changing in 2026, and an AI that touches invoices or entries inherits that instability. Joint Act RFB/CGIBS No. 4, released by Brazil's Federal Revenue Service on July 31, 2026, set August 3, 2026 as the start of mandatory reform tax documents for NF-e and NFC-e (Brazil's electronic product and consumer invoices), and October 1, 2026 for NFS-e (service invoices) in general.
On August 3, Brazil's Ministry of Finance announced it was suspending the requirement to fill in CBS and IBS information (the new consumption taxes) on documents such as NF-e, NFC-e and CT-e: they would not be rejected if those fields were missing. The Federal Revenue Service's schedule page, updated on October 28, 2026, warns that the layout publication dates are estimates and may be adjusted.
In Wama's view, an AI flow that touches tax fields should only write after the ERP vendor confirms the version with the new layouts and the accountant validates the company's rules. Until then, AI helps more by reading, checking and flagging discrepancies. It's the same logic as back-office automation with AI: start with what can be undone.
A five-step preparation order
The order below works for most ERPs and doesn't depend on the vendor. It goes from process to technology, never the other way around.
Pick a flow with volume and recoverable errors, such as order lookup or receivables reconciliation, and measure what it costs today.
Find out what your version's API covers for that flow, testing in the test environment.
Create the technical user with read-only access and log every call: what was requested, what came back and who approved it.
Allow drafts with human approval and set, with the process owner, the value and quantity ceiling for writing without approval.
Review the flow whenever the ERP gets a tax or version update.
At Wama, the AI Assessment produces a custom Implementation Plan and the baseline for measuring before and after, and that's where this first flow is chosen. The AI implementation plan turns the five steps into stages with deadlines and owners.
Common mistakes
The most expensive mistake is using the administrator account to test quickly. The test becomes production without anyone noticing, and the AI ends up with a power no one in operations has alone.
The second is testing directly on the live database. Without a test environment, the first wrong draft is already a real order, with inventory reserved and the customer notified.
The third is treating the ERP as if it were a CRM. In a CRM, a wrong field gets in the way of a sale; in an ERP, it becomes an accounting entry and an invoice. The permission logic of CRM integration with AI helps, but the write ceiling in the ERP needs to be lower.
The fourth is not keeping an audit trail. If nobody can reconstruct what the AI looked up, what it proposed and who approved it, an error turns into an argument from memory, and the customer data that went through the integration has no one accountable for it.
Frequently asked questions
Do I need to switch ERPs to integrate with AI?
In most cases, no. The starting point is the interface your version already offers and a small flow. Switching ERPs mixes AI adoption with a system migration, which carries its own risk. It only makes sense to discuss a switch when the installed version can't provide the data the flow needs.
Can AI issue invoices on its own?
It can prepare the invoice, but issuing on its own calls for special care in 2026. Tax documents in Brazil now follow the reform schedule, with CBS and IBS, and the validation rules were adjusted throughout the year. Until the ERP vendor and the accountant confirm the setup, issuing stays with a person approving.
How much does ERP AI integration cost?
It depends on what your ERP version offers, the chosen flow and how much needs to be built around it. In general, using an existing API takes less building than setting up an intermediate layer. The factors that affect the budget are gathered in how much it costs to implement AI, with the model cost separated from the project cost.
Where should ERP AI integration start?
Start with a lookup flow that has volume and recoverable errors, such as order status or open receivables. Create a read-only technical user, test in the test environment and log every call. Writing comes later, with a ceiling written by the process owner and human approval for anything above it.
Who should take part in the preparation?
At least the process owner, whoever administers the ERP and the accountant, when tax fields are involved. Wama doesn't sell boxed solutions: each company gets its own plan, and that's where it's written down who approves what. Without those three people, the AI ends up deciding on its own things nobody authorized.
If your ERP already holds what the company needs, but nobody can say what the AI could read or write in it, the first step is to describe one flow clearly. Talk to Wama to start with the assessment and get a custom plan for the system you already use.
