An AI agent on top of your ERP: what it does and doesn't do
An AI agent working on top of your ERP reads PDFs and emails, matches them to records, writes the result into the ERP and prepares reports. It does not decide, does not change the ERP and does not move data outside the company. The agent produces the draft and an authorized person gives the approval; your ERP's screens, permissions and data stay exactly as they are.
What does an agent on top of the ERP mean?
An ERP keeps records: customer/supplier accounts, stock, invoices, orders. But the information that goes into a record often comes from outside the ERP. A PDF invoice from a supplier, an order by email from a customer, a bank statement in Excel, a WhatsApp message from the sales team. The person who reads these documents and enters them into the ERP is someone on your team.
The agent is a software layer that takes over the first steps of this job. It does not replace the ERP; it sits next to it. It uses your ERP's screens, permissions and data structure as they are. So your ERP does not change; what changes is the path a document takes until it reaches the ERP.
How does the agent connect to the ERP?
The connection method varies by ERP. If the ERP has an API, the agent talks to it. Read-only access to the database may be enough. Some older programs require file transfer or working through the screen. The method is set at setup, and the integrations page lists which systems we work with.
Three rules hold for every method. The agent logs in as a separate user, so every record shows who wrote it. Its permissions cover only its job: an agent that enters invoices cannot open personnel records. You can switch off its permission at any moment and the ERP works as before.
What exactly does the agent do in daily work?
Four verbs are enough: it reads, matches, writes, reports.
- Reads: PDFs, scanned documents, email bodies, Excel tabs, WhatsApp messages. Even if every supplier's invoice looks different, it extracts the amount, date, tax number and line items.
- Matches: links an invoice line to an order, a statement row to a customer/supplier account movement, a product name in an email to a stock card. It also writes the reason for the match.
- Writes: enters the matched record into the ERP as a draft, or prepares it for entry. The record becomes final after approval.
- Reports: every morning it summarizes open orders, late payments and stock differences on a single page.
Example: if someone opens 40 PDF invoices every day and enters them one by one, the agent reads those 40 documents and prepares the drafts. The person looks only at the five or six records the agent is unsure about. A job that took two hours a day becomes checking and resolving exceptions.
What does the agent look like in a working day?
At 07:00 in the morning the agent reads yesterday's open orders, late collections and critical stock, and emails a one-page report. During the day it reads incoming invoices in turn, matches them to orders and puts the drafts in the approval queue. The accountant opens the queue, approves the matches and corrects those that do not match.
In the evening the agent writes the day's summary: how many documents were read, how many approved, how many put to a person. This summary is the easiest way to measure whether the agent is working.
In this flow the agent finalizes nothing on its own. The report is read, the queue is approved; behind every line that enters the ERP there is an approval. You set the rule: which document type, which amount, whose approval.
What does the agent not do?
The boundary has to be drawn at the start. What the agent does not do matters as much as what it can do, because trust is built by limits.
It does not decide
The agent does not decide whether an invoice will be paid, whether a discount will be given, or whether a customer's payment term will be extended. The agent prepares the case: it shows the document, the match, the difference. An authorized person makes the decision.
It does not change the ERP
The agent does not change the ERP's code, modules or settings. It writes with the methods the ERP offers, under the permissions of a user you define. A screen that user cannot see is a screen the agent cannot enter either. You grant the permission and you take it back.
It does not move data
The agent does not pile a copy of your data into another system. The data stays on your server or in your cloud account; the agent reads from there and writes there. Data does not leave the company. We cover this topic in detail in a separate article.
What happens with a record it is unsure about?
A good agent does not try to match everything. If a customer/supplier name fits two records, if an amount does not match the order, if a document cannot be read, it sets the record aside and asks a person.
That is why the approval point is the heart of the design. The agent prepares the draft and puts it in the queue with its question; an authorized person approves, corrects or rejects. No record that is not approved becomes final in the ERP.
Every step is written to the log: what the agent read, what it matched it with, who approved, when. In an audit or when investigating an inconsistency, this log rests on a document, not on memory.
What happens to the team?
The team does not shrink; its work changes. The person who opened documents and retyped them becomes the person who checks the draft the agent prepared. They are left with less data entry and more exception handling.
The difference looks small, but it changes the daily load. At month-end, instead of entering all invoices from scratch, you look only at the three invoices that do not match, flagged by the agent. The time is left for what the person who knows the job should really be doing: calling the supplier, finding the cause of the difference.
It gets easier for a newcomer too. They do not need to know by heart which account to post an invoice to; they read the agent's reasoning and approve if it is right. An experienced employee corrects the rule, and the agent does not make the same mistake on the next document. Knowledge leaves a person's head and becomes a written rule.
For which jobs an agent, for which another solution?
An agent makes sense where the input varies and the job involves reading and matching. If the input always has the same format and the rule can be fully written, simpler tools are enough. You do not have to put an agent on every job.
- Agent fits: PDF invoices in different layouts, order emails with free text, matching bank statements to customer/supplier accounts, the morning report.
- A simple rule is enough: importing into the ERP an Excel file that arrives every day in the same format with fixed columns.
- A person should stay: price negotiation, payment-term decisions, disputed invoices, working with a new supplier for the first time.
The difference between an agent and a classic software robot is a separate topic; the AI agent vs RPA comparison puts the two approaches side by side and also states plainly where RPA is enough.
What is needed to try it?
First a single job is chosen; for example incoming PDF invoices. A few samples are taken from your real documents and the agent's reading and matching are shown. Panda gives a demonstration within 3 business days; if you like it, you start with a 30-day pilot on one process, with approval points.
The measure of the pilot is simple: how many hours a day did the person spend before, how many now? You calculate the number with your own team's time. If the agent is not working, you see it in the first weeks, with the documents in hand.