What a company perimeter is, and when data never leaves it

Why this matters

• Understanding exactly what protection a vendor is charging extra for, and when that premium is justified
• Answering security and legal before a project starts, rather than after the first incident
• Splitting a process so the expensive part runs inside and the cheap part runs outside

What a company perimeter is, in plain language

A warehouse has a gate. Inside are the goods, the racks, the company's own storekeepers, cameras and a log of who came in and what they took out. Outside are contractors, couriers, buyers. Goods can leave, but only through the gate and only against a delivery note.

A company perimeter works the same way, except the goods are data. Inside the boundary sit servers, databases, mail, accounting systems. Every employee has an account; every action leaves a trace in a log.

When a language model is added to a process, the question comes down to one thing: does the model work inside the gate, or are the documents carted out to it? The answer determines cost, time to launch, and what will have to be explained to auditors.

How it works

1. Draw the boundary. List what counts as internal: servers, networks, accounting systems, mail, file storage. Anything not on the list is automatically external.
2. Label the data step by step. Sorting incoming mail, reconciling invoices, searching the knowledge base — each step is marked with what it handles. Personal data and trade secrets go on one list, public information on another.
3. Assign a mode to each step. Sensitive steps are covered by a model with open weights on company hardware. The rest go to an external model API, where the price is lower and the quality is higher.
4. Put a gateway on the boundary. Every request to the external model passes through it. The gateway strips names, contract numbers and amounts from the text, writes a log, and throttles volume once a limit is exceeded.
5. Put the arrangement in writing. The contract with the vendor fixes where logs are stored, how long copies of data live, and who has access to the server running the model.

The boundary is not drawn by hardware alone. Data leaks through exports to spreadsheets, through employees' personal accounts, and through intermediate integration services. Those three channels are checked separately.

A case from practice

A manufacturing company processes roughly 4,000 incoming documents a month: invoices, delivery notes, acceptance certificates. The documents carry counterparty details, contract amounts and the names of signatories. Security will not approve sending them out.

The process was broken into steps, and for each step the number of documents containing sensitive fields was counted. It turned out that classifying the document type and extracting the number manage without them: the first few lines and the page layout are enough. The full text with amounts is needed only at reconciliation.

Then two bills are calculated. The first is the per-operation price at an external vendor, multiplied by monthly volume. The second is the lease on a GPU server, engineering time and updates, divided by the same volume. The figures for a specific process come from the in-house perimeter versus API calculator; there is no general answer here, because the break-even point depends entirely on volume.

The figures above are an illustrative model for a company of this profile, not a report on a completed rollout.

What the business gets out of it

• The conversation with security turns into a conversation about lists. Instead of arguing whether AI is allowed at all, the discussion is about specific fields in specific documents. That argument is settled in a single meeting.
• The bill becomes manageable. The expensive mode covers only a minority of steps, while high-volume routine goes to an external model at a per-operation price.
• The project survives a change of vendor. Once the boundary is documented, replacing the model inside the perimeter touches one section rather than the whole system.
• An audit is answered with documents. The network diagram, the access log and the contract terms are prepared in advance rather than assembled in a rush when the regulator asks.

When this can be skipped

If a process only ever handles public text — product descriptions, articles for the website, newsletter drafts — the boundary protects nothing and adds cost. Here an external API is the answer, with the bill calculated by volume.

The same goes for a pilot on de-identified data. Testing a hypothesis on a hundred prepared examples requires neither in-house hardware nor a gateway. The perimeter goes up at the step where live traffic begins, and that is a separate budget line to plan for in advance.

Low volume also changes the answer. At a few hundred operations a month, the fixed cost of owned hardware is divided across too small a number, and the cost per operation ends up several times the external rate.

What to check carefully

First: de-identification does not always work. A surname turns up in the body of an email, a contract number in a file name, an amount in a table caption. The leak rate is measured on a sample of 200–300 documents and reviewed by eye, not estimated by feel.

Second: a model on company hardware is weaker than the best external ones. The quality gap is measured on the same sample of the company's own documents, by the same method. It often turns out that on a narrow task the gap is smaller than general benchmarks suggest.

Third: ask the vendor for a data flow diagram — which fields cross the boundary, where logs are written, how long temporary copies live. Without that diagram, there is no way to verify the promise that "everything stays with you".