
For a long time, enterprise security architecture could be explained using three reasonably tidy boxes:
- Firewalls protecting networks
- Endpoint security protecting devices
- Secure connectivity linking branches, data centres and cloud infrastructure
It was never quite that simple, of course. Nothing in enterprise IT remains simple once there are 2,000 users, 40 branches, three cloud platforms, six operating systems and one business-critical Windows Server that nobody is permitted to restart.
Still, the broad model worked.
Then applications moved to the cloud, employees began working from everywhere, and artificial intelligence introduced an entirely new question: what exactly are users sending outside the enterprise?
Start With the Control Point
Every security control has a vantage point. It can only inspect and control what passes through that point.
A firewall at the branch can see traffic entering and leaving that branch. Endpoint security can see activity on a managed device. A SASE platform can inspect traffic from remote users if their traffic is routed through it. A browser extension can see what happens inside the browser.
This sounds obvious, but many security designs become complicated because we start with a product category rather than the control point.
Before buying anything, the better questions are:
- What are we trying to observe?
- Where does that activity occur?
- Does the proposed control actually see it?
- Can users or applications bypass that control?
- What action should be taken when something goes wrong?
If the answer to the third question is “sometimes,” the answer to the fourth is usually “probably.”
What the Firewall Can Control
The network firewall remains essential. It protects the enterprise perimeter, segments internal networks, controls applications, blocks threats and enforces access policies.
It is particularly effective when users and workloads are behind a predictable network boundary. Branch offices, campuses, factories, data centres and cloud networks still need strong network-level security.
The firewall can decide which systems may communicate, which services can be accessed and what traffic should be blocked. With HTTPS inspection, it can also inspect permitted encrypted traffic, subject to policy, privacy, performance and application compatibility considerations.
But the firewall can only control traffic that reaches it.
If employees are working remotely and connecting directly to cloud applications, the office firewall is sitting in the office doing its job admirably while much of the workforce is somewhere else.
It is not the firewall’s fault. It was not informed that the perimeter had moved.
Why Remote Work Makes SASE Important
When users work from homes, hotels, airports and customer locations, security enforcement must follow them.
This is where SASE becomes useful. A SASE client can route endpoint traffic through a cloud security service, allowing the organisation to apply web filtering, access policies, threat protection and data controls regardless of the user’s location.
When HTTPS inspection is properly enabled, a SASE platform has much better visibility into conventional web traffic. It can inspect uploads, downloads, web requests and application activity that would otherwise appear as encrypted connections.
This creates a more consistent security layer across branches and remote endpoints.
However, HTTPS inspection is not a magical “show me everything” button. It depends on certificate deployment, browser and application behaviour, endpoint control and correctly designed exclusions. Some non-browser applications use their own trust mechanisms or certificate pinning, which can interfere with TLS inspection. Apple itself advises that pinning is not required for most applications, but the mechanism remains available and is used in some environments. Apple Developer
SASE provides an important control point. It does not automatically provide every possible control.
Every control sees from its own vantage point
AI Traffic Changes the Question
Traditional web security asks:
Which website is the user visiting?
AI security must ask:
What did the user type, paste, upload or ask the model to process?
Those are very different questions.
Knowing that an employee accessed an AI service may be useful. But it does not tell you whether the employee asked for help writing an email or uploaded a confidential customer database.
Network-level visibility into AI traffic is becoming more difficult. Encrypted Client Hello, or ECH, can hide information in the TLS handshake that network controls traditionally used for domain identification. Certificate pinning and application-specific trust behaviour can prevent interception. AI services also operate across dynamic cloud and content-delivery infrastructure.
ECH does not make all traffic invisible, nor do all AI services and applications behave identically. But the overall direction is clear: network metadata alone is becoming a weaker basis for understanding user intent.
Even where HTTPS inspection works, the security platform still needs to understand the application protocol, distinguish prompts from responses, detect file uploads and apply context-aware policies. Merely decrypting traffic is not the same as understanding it.
The Browser May Be the Best AI Vantage Point
For browser-based AI usage, the browser is often the most useful place to observe behaviour.
A browser-level control can potentially see:
- The AI service being used
- The prompt entered by the user
- Text pasted into the conversation
- Files selected for upload
- The user’s identity and device context
- The response returned by the model
This allows policies to be applied before sensitive information leaves the device.
For example, the organisation may allow employees to use AI for general research but block the submission of source code, financial information, credentials, personal data or confidential documents.
The important distinction is that the browser control sees the user action. The network control sees a connection.
Unfortunately, users are highly creative participants in security architecture. If the browser is controlled, somebody will eventually install a desktop application. If the desktop application is blocked, somebody will discover a command-line client. If all of that is blocked, somebody will use a personal phone and describe the confidential spreadsheet manually.
Technology cannot eliminate every form of poor judgement, but it can remove the easy paths.
Desktop AI Applications Create an Observability Gap
Native applications for Claude, Gemini and other AI services may not provide the same visibility as a managed browser.
They may use different certificate stores, networking libraries or trust models. Some can work through enterprise proxies, while others may require additional configuration or fail under TLS inspection. Even applications from the same provider may behave differently depending on their runtime and operating system.
This creates a considerable loss of observability. The organisation may know that an AI service was contacted, but not what was submitted.
An endpoint agent can help by controlling application execution, monitoring files and enforcing device policies. However, reconstructing the full meaning of an AI conversation from endpoint or network events is difficult.
If the security objective is to understand and govern prompts, the control should ideally sit where the prompt is created or where it is deliberately routed.
Perhaps Enterprises Need Their Own AI Gateway
A more manageable approach may be to restrict direct access to public AI services and provide an approved enterprise AI portal.
Users submit their requests through the organisation’s portal. Behind it sits an AI gateway or AI firewall that can:
- Authenticate users
- Inspect prompts and uploaded files
- Detect sensitive information
- Apply departmental policies
- Maintain appropriate audit records
- Select approved models
- Remove or mask sensitive data
- Enforce cost and usage limits
- Route requests across different AI providers
- Inspect model responses
- Block unsafe or non-compliant requests
This AI firewall is not simply another network firewall with the word “AI” printed on the box.
Architecturally, it is closer to a secure proxy combined with a policy engine, data-loss prevention controls and a multi-model routing layer. Platforms such as OpenRouter demonstrate the routing concept by providing access to different models and providers through a common API. OpenRouter documentation
The enterprise version must add identity, governance, privacy, security policies, auditability and control over where data is processed.
Direct access to public AI services can then be blocked or limited using the firewall, SASE platform, endpoint controls and browser policies. Employees still receive access to useful AI models, but through a route that the organisation can govern.
And Now We Have Accidentally Created a Vendor Meeting
Let us count what the IT manager may need:
- A firewall for branches and data centres
- Secure SD-WAN for interconnecting locations
- Endpoint security for managed devices
- SASE for remote users
- Browser-level AI controls
- An AI gateway for approved model access
- Centralised monitoring and policy management
Depending on how the market packages these capabilities, this can quickly become four or five vendors.
Each vendor brings a demonstration, a proof of concept, a licensing structure, an agent, a dashboard and a strong opinion about why its dashboard should be the first one opened every morning.
The IT team must then integrate identity, policies, logs, endpoint agents, network rules and incident workflows across all of them. Security architecture slowly becomes integration architecture.
This is precisely the problem NetSense wants to solve.
The objective is not to claim that one appliance can perform every security function. Different controls genuinely belong at different vantage points. The objective is to make those controls part of one coordinated architecture.
The firewall should understand the endpoint. The SASE platform should share policies with the branch. The browser control should work with the AI gateway. The AI gateway should use the organisation’s identity, data and access policies. Administrators should be able to manage all of this without maintaining a handwritten map of which vendor is responsible for which packet.
Security architecture should not become integration architecture
The Future Is Not One Control Point
Enterprise security is no longer located only at the perimeter.
It exists at the network, endpoint, cloud edge, browser and application gateway. AI adds another layer because the organisation must understand not only where traffic is going, but what the user is asking an external system to do with enterprise information.
The right question is therefore not:
Do we need a firewall, SASE or an AI security product?
The better question is:
Where must each control operate, and how do we make all those controls work as one system?
That is the big-picture security problem facing IT managers, IT directors and CTOs today.
And if answering it requires four vendor demonstrations, twelve dashboards and three different endpoint agents, perhaps the architecture itself needs a rethink.
See it in your environment
Talk to a NetSense security architect about applying this to your stack.