Network Separation and Generative AI
Network Separation (Air Gap) & Generative AI
What are network separation and generative AI?
- A security policy that keeps the business network separate from the internet
- Why commercial AI gets blocked, and what drives demand for on-premises
- Data classification is where adoption talks start
In organizations bound by network separation, the order of AI adoption flips. A typical company starts by asking “which model performs best,” but in finance and the public sector, where the model can be placed is decided first, and the usable models are narrowed down within that. Who owns the project changes too. Business units don’t pick a tool and bring it in; the information security department and legal must approve the deployment plan before work can start, and a large share of the budget goes to infrastructure and controls rather than model usage fees.
Network separation isn’t one monolithic thing either. There is the physical approach, with two sets of PCs and lines, and the logical approach, which only displays a screen through VDI. Locking down the business network and moving only internet access to a separate PC also differ in how tightly they control access. Where AI can run follows this structure. On-premises deployment brings the purchase of GPU servers and operating staff, so teams often start with a lightweight model with fewer parameters, while a private cloud requires a dedicated line, dedicated instances, and outsourcing-related procedures. The boundary people most often confuse here is that network separation and leak control are different layers. Even with the networks split apart, a single copied-and-pasted line can carry out confidential information, and that is the domain of DLP and prompt filters.
The most common disconnect in the field is closing the review with one line: “We have network separation, so it’s not possible.” What is actually blocked is the external API call path, not all AI use, but when the discussion stops there, employees take photos of their screens on personal phones and ask an external chatbot. This shadow AI is the real hole in network separation, with data leaving outside any control. There is a misconception in the other direction too: assuming that deploying an open-source model internally settles everything. If internal documents aren’t organized, search quality collapses and you end up with a system no one opens after two months. Regulations are also revised often, so don’t declare “not allowed” based on last year’s material; check the guidelines of the authority that governs you each time.
Seolyeong ParkGenerative AI and AX training expert with 17 years of hands-on experience and 200+ sessions, making AI usable right away even for non-specialists
Seungtae ParkPractical AX instructor who teaches by building AI business development and marketing automation hands-onKorea began tightening its network separation policy in earnest after the 2011 hacking incidents at major financial and public institutions. The Regulation on Supervision of Electronic Financial Transactions and national cybersecurity guidelines spread it across the financial and public sectors. As demand for cloud and AI use has grown since, discussions on conditional relaxation have continued.
Consider a hypothetical case. Company A, a property and casualty insurer with 1,200 employees, runs every PC in a physically separated network environment, so summarizing a single claims review file took 40 minutes on average. Executives ordered the use of AI, but the information security team replied 'not possible because of network separation' for six months, and a review task force with a 3-month deadline was finally formed.
- When the head of information security sorted 3,200 internal documents into 4 grades (public, internal, confidential, customer information), only 18% actually contained customer information. The rest had no clear criteria and were all lumped in as confidential by habit.
- The task force split out claims review summaries and policy-terms Q&A. When the head of the review team asked, 'The terms are all published on our website, so why can’t we use them?' the plan became a 3-month pilot on 5 internet-network PCs, starting with public-grade tasks.
- For the review summaries that mix in customer information, they loaded a lightweight open-source model onto 2 internal GPU servers and ran it with RAG. First-test accuracy fell short of expectations, and they spent another 4 weeks just cleaning up the policy terms and manuals.
- The legal team and the information security department wrote up the justification for each deployment option in an 11-page A4 audit file, and the compliance officer gave conditional approval, saying, “We’re fine as long as we can explain it later.”
- Six months later, the time to summarize a claims file fell from 40 minutes to 12 minutes, and reports of employees using external chatbots on personal phones dropped from 9 a month to 1.