Why internal government documents should not go to public AI platforms

It happens more often than people think. An officer needs to summarise a forty-page report before an afternoon meeting. They open a browser, paste the whole thing into a free AI tool, and have a summary in thirty seconds. The job gets done, nobody complains, and the habit sticks.

The problem is that the report has just left the organisation’s control, and nobody inside the organisation knows it happened.

What actually happens when you paste text into an AI service

The content you paste travels over the Internet to the provider’s servers, usually in another country. There it is processed, stored in the account’s conversation history, and — depending on the terms of service — may be retained for a period for operational or improvement purposes.

Three consequences follow:

You no longer control the copy. Once data leaves your network, where it is stored, for how long, and who can access it are all outside your authority.

There is no log on your side. If you later need to answer “has this document ever been sent outside”, your systems have nothing to check. The action happened in a personal browser and passed through no control layer at all.

Personal account, organisational liability. The officer used their own account, but the data belongs to the organisation. When something goes wrong, the line of responsibility is very hard to draw.

This is not hypothetical

Several large organisations worldwide have issued internal bans after discovering staff pasting source code or internal documents into public AI tools. In the public sector the constraints are tighter still: operational data is usually tied to citizens’ personal information, and transferring personal data abroad falls under Vietnam’s Decree 13/2023/ND-CP on personal data protection.

The important point: almost none of these cases involve bad intent. Officers are simply trying to get their work done faster. That is precisely why a ban on its own does not work.

Why banning does not solve it

Issuing a directive against public AI tools is a necessary step, but on its own it usually produces one of two outcomes.

Either officers comply and go back to doing the work manually. The organisation loses the efficiency the technology could have provided, while the workload stays the same.

Or — more commonly — they keep using it, just more discreetly. A ban does not remove the need; it pushes the behaviour somewhere nobody can observe it.

The need to summarise a long document, look up a regulation quickly, or draft a report is real. It does not disappear because of a memo.

The workable answer: bring the tool inside

The sustainable approach is to give officers a legitimate AI tool to use, running inside your own infrastructure. When the model runs on servers at your premises:

• Documents never leave the internal network, including dedicated networks with no Internet access.
• Every query passes through your permission system — officers can only work with data they are cleared to see.
• Every action is logged, so “who viewed which document, and when” always has an answer.
• Officers no longer have a reason to reach for an outside tool, because the internal one is good enough.

That last point matters most and is the one most often missed: a technical solution only works if it is convenient enough that people choose it naturally. A slow, awkward internal system gets ignored, and everyone returns to their old habits.

Questions worth asking before you decide

If your organisation is considering this direction, these are worth putting to any vendor:

• Where does the model run? If the answer contains the word “API”, ask where that API is hosted.
• Are there any functions that require an Internet connection to work at all?
• What does the system log, and where are those logs stored?
• How many permission dimensions are there, and can access be controlled by document sensitivity?
• If the vendor ceases operations, can your organisation keep running the system?

That last question is frequently skipped, but for a system expected to run for years it is the important one.

---

Insightech builds AI software installed on the customer’s own servers, capable of running on a LAN or a dedicated network. If your organisation is considering this route, we are happy to discuss your specific situation before proposing anything.