AI without data leaving your company: on your own server or cloud account

· Panda Yazılım

An AI agent does not require your data to go to another company. The agent runs on your server or in your own cloud account, and reads and writes your ERP's data there. Before deciding, look at three things: where the data sits, what the agent can access, and where every step is logged.

What does "data does not leave the company" mean?

This sentence describes three separate things, and each must be asked about separately. Where is the data stored? Where does the agent run? To which third party does the agent send data while doing its job?

Many products say "secure" but keep a copy of the data on their own server. The test here is simple: does a copy of the invoice, customer/supplier account and stock information from your ERP stay permanently outside the area you control? If it does, the data has left.

Where can the agent run?

There are two ways. In both, the agent is installed on your infrastructure and no copy of your data accumulates at the agent provider.

  • On your own server: a server inside the company or one you rent. If your ERP also runs there, data never leaves your network.
  • In your own cloud account: a cloud account whose subscription is opened in your name. You are the account owner; you open and close access.

Which one suits depends on your ERP. If you use a server-based ERP, the agent mostly runs on the same network. If you use a cloud ERP, the agent connects to that cloud's API with a key you give it. The difference is explained in detail in the cloud ERP vs server ERP comparison.

The question to ask when choosing is: where does your ERP run today, and do you plan to change that? If you do not, the agent is installed where you are. Opening new infrastructure is not required just for the agent. Managing one place is easier than managing two.

Setup time varies by environment; access permissions need your IT owner's approval. Getting that approval up front saves time on demo day.

How does a document reach the agent?

An incoming PDF invoice first lands in your mailbox or your folder. The agent picks the document up from there, reads it and writes the result into the ERP as a draft. Both the original document and the extracted fields stay on the same infrastructure. There is no intermediate file copied to a third place; if there has to be one, you must know and approve that in advance.

The point to watch here is temporary files. Some systems upload the document to another server for a while to process it and then delete it. Even if deleted, the data has left for a moment. Ask the provider plainly whether such temporary steps exist.

How is the access scope set?

As much as where the data sits, who accesses what matters. The agent is defined as a separate user. That way every record shows who wrote it and you manage permissions from one place.

  • Only what its job needs: an agent that reads invoices has no need to see personnel records.
  • Read and write are granted separately: a reporting agent only reads, an agent preparing records can write drafts.
  • Finalizing stays with people: a draft becomes final after approval.
  • Reversible: the moment you switch off the permission, the agent stops working and your ERP is used as before.

Keep the permission list in writing. Which table, which field, for which operation is accessed should be on paper too. If asked one day, the answer is ready.

Why is an activity log needed?

The agent must write down what it did. The log holds: which document was read, which value was written to which field, which rule it matched, who approved, when. This log stays on your infrastructure.

The log does two jobs. The first is debugging: if an invoice was posted to the wrong customer/supplier account, you see why and when. The second is accountability: when an auditor or manager asks "who entered this record", the answer rests on a document.

Put access records next to the log too. It should be visible at what time the agent connected to the ERP, what it did and how many records it read. If there is reading at an unexpected hour or in an unexpected volume, you should be the first to notice.

Backups are thought of with the same logic. The log and the agent's settings go into your backup routine together with your ERP backups. Opening a separate backup location means opening one more place where data is copied.

Which KVKK responsibilities stay with you?

This section is not a legal opinion but a general framework. For documents containing personal data (employee payrolls, customer contact details, ID numbers), the data controller is generally your company. You are also the one who decides why you process the data, who accesses it and how long you keep it. KVKK is Turkey's data protection law.

The agent does not change this responsibility; it adds a new processing step. So before putting the agent into service, it is good to check whether your privacy notices and your inventory are up to date. On matters such as whether data is transferred abroad, retention periods and deletion requests, your company's legal counsel or KVKK officer should make the assessment.

In practice the most useful thing is to list up front the types of data the agent will touch. Invoices and waybills are mostly commercial information. Documents containing payroll, personnel files or customer contact details are more sensitive. If possible, running the first pilot on a process with no personal data narrows the risk framework.

Which questions should be asked before buying?

When talking to a provider, ask for written answers to these questions:

  • Does the copy of my data sit on my server or on your infrastructure?
  • To which external services does the agent send my documents while doing its job?
  • Are my documents used to train the agent?
  • What happens when I switch off the permission, and who keeps the log?
  • When the contract ends, what traces remain on my infrastructure?

If every answer says "you", the data is not leaving the company. If any one is unclear, that item should be clarified before the pilot.

How does Panda do this?

Panda AI is installed on your server or in your cloud account; no copy of your data is kept at Panda. The agent connects as a separate user and the log stays on your infrastructure. First a demonstration is given with your sample documents within 3 business days, then you start with a 30-day pilot on one process.

Your ERP does not change, data does not leave the company, the team does not shrink; its work changes. These three assurances should be discussed in writing before the pilot begins.

Frequently asked questions

What does "data does not leave the company" mean?

It means that no permanent copy of the invoice, customer/supplier account and stock data in your ERP is kept outside the server or cloud account you control. The agent reads the data there and writes it there.

How do I limit what the agent can access?

The agent is defined as a separate user and given permission only for the tables and fields its job needs. Read and write are granted separately. When you switch off the permission the agent stops and your ERP works as before.

Who keeps KVKK responsibility?

For documents containing personal data, the data controller is generally your company. This is a general framework, not a legal opinion. For transfers, retention periods and deletion requests, consult your legal counsel.