Security says no to the cloud. Where does your business AI go?

#The three deployment types differ in where servers and models live

When a company evaluates business AI, the discussion tends to drag once someone asks, "Can we actually put this in the cloud?" The cloud is the fast way to start, but the data makes people want to keep everything in-house. And if your company runs a network with no internet access at all, the first question is whether there is any option.

The answer depends on your company's policy, and every answer leads somewhere. If internal policy allows business data to be processed in an external cloud, you start right away with cloud (SaaS). If policy or your security team won't allow it, on-premises installs OwlAI on your own servers, and if internet access itself is blocked, there is the air-gapped network deployment. All three give you the same screens and the same core features. What changes is where the servers sit and where the AI models run.

#Operating burden and time to start vary by type

With cloud (SaaS), you sign up for the SaaS that OwlAI runs, so there are no servers to prepare. With on-premises, you install OwlAI on company servers and run it yourself. An air-gapped network deployment is on-premises installed on a network cut off from the internet; it runs only on models on internal GPU servers, with no external AI services.

The servers and models a question passes through on its way to an answer also differ by type.

#Core features are the same; only features that reach outside differ

Chat, database queries, document search, meeting minutes and playbooks work the same way in all three.

Web search and web page reading are available wherever there is internet access. On an air-gapped network, research covers internal databases and documents instead of the web. Connectors to external services split the same way: cloud (SaaS) can connect only services reachable from the internet, on-premises can connect internal systems as well, and an air-gapped network can connect only systems inside the internal network.

Model configuration differs too. Connecting self-hosted GPU models is an option only on-premises and on air-gapped networks. An air-gapped deployment runs chat, database queries, document writing and playbooks entirely on open models loaded on internal GPUs. Whether those models fit your company's work is something to check by putting real business questions to them during a pilot.

On-premises and air-gapped deployments should have a proper HTTPS certificate for the server address. Microphone voice input and web notifications work only over a secure connection, and with a self-signed certificate every browser and desktop app has to be set up to trust it. The mobile app connects only to servers at an https:// address.

#Your policy and the systems you need to connect decide the type

If you want to start fast and see the value first, and policy has no objection to processing business data in an external cloud, cloud (SaaS) is the fit. With no servers to prepare, one administrator can try it right away on free trial credits. You subscribe through AWS Marketplace and are billed monthly per user seat, and each seat comes with that month's AI usage as credits.

If you need deep connections to internal databases and systems, or the company must control where its data stays, on-premises is the fit. Its biggest advantage is that you can mix external and internal models to suit each kind of work.

If regulation or security classification rules out internet connectivity on your network, the answer is an air-gapped deployment. It means procuring GPU servers and going through an import procedure, but not a single character of your questions or data leaves the company. On-premises and air-gapped costs vary widely with the number of users, the model configuration and whether new GPU servers have to be procured.

Before choosing a type, settle three questions with your IT department and security team:

• Does internal policy allow business data to be processed in an external cloud?
• Are there internal databases and systems the AI needs to connect to?
• Is your network barred from internet access, and if so, can you procure GPU servers?