On a Trust-Building Trajectory: AI in Network Automation

Three and a half years ago, ChatGPT was being used to write silly poems. Today, it has the ability to write applications with co-pilot code editors for mobile applications in minutes. We are in the midst of a similar trajectory with AI in network automation.

AI has not yet earned the trust required to independently make operational decisions in production networks, and it does not replace the foundational work of deterministic automation. But it is far more than a passing trend. AI is already transforming how engineers troubleshoot, reason about complex systems, and interact with automation platforms. The question is not if AI belongs in network automation, but how it should be introduced safely.

Just as “cloud-first” initiatives were largely driven top-down, artificial intelligence adoption is often initiated at the executive level. But its impact is felt across the entire network engineering organization. That means success depends on understanding where AI fits, where it does not, and how trust in the technology is built over time. Consider this as a guide for networking professionals to navigate the differences between deterministic automation and cognitive automation (GenAI), when to use one versus the other, and the considerations for companies to assess as they plan their AI journey in 2026.

This is fundamentally a conversation about trust.

See also: Are You Getting the Best Results from RPA?

When to Use What – GenAI vs. RPA

When differentiating the application of cognitive automation (Gen AI) and deterministic automation (RPA), Gen AI is effective for “passive” functions that primarily assist the human rather than making direct changes to the network. The human would then take action based on the AI’s assistance. Examples of this are root cause analysis, troubleshooting, parsing large datasets, finding “needle in a haystack problems,” and debugging by analyzing logs and output between devices (e.g., identifying errors in OSPF adjacency timers). It is recommended for GenAI implementation to be done as a stepped approach for organizations to mature Gen AI capabilities by tying it to existing automation execution engines.

While organizations might lean on Gen AI to gather information from a source of truth, build reports, and understand capacity (e.g., number of ports available, circuit information) by tying together disparate pieces of information, which Gen AI is pretty good at, checks and balances with these types of functions are required.

When it comes to choosing when to use deterministic automation (RPA), it is recommendedfor any job or execution engine with a defined trigger (input) and output. This means that no matter how many times the same operation is performed, it will always create a consistent outcome.

New users should fundamentally understand that they will pass information in, expecting the system to perform an action or return information as an output. The two main mechanisms for making changes are screen scraping/control or API integration. When using RPA, organizations must understand the data that drives the change. Inputs must come from a consistent place, and organizations need systems to ensure the data is “good,” as automation can execute a bad change just as easily as a good one if the data is faulty.

The core issue with either technology comes down to trust. RPA can fail if the format of data scraped or parsed from a device changes, while AI can hallucinate or guess incorrectly, especially with mathematics. While improvements are being made to these technologies all the time, guardrails and checks and balances are key. 

See also: From Automation to Autonomy: Building the Architecture for Agentic AI

When AI Use Can Be Dangerous – The Need for Guardrails and Checks and Balances

While many organizations publicly emphasize their focus on artificial intelligence, most stop short of allowing AI to directly make network changes. In practice, they continue to adhere to established operational frameworks such as ITIL-based change control, and for good reason.

There are very few organizations today using AI in a fully autonomous, end-to-end manner from “cradle to grave.” Due to its non-deterministic nature and unresolved trust concerns, caution is warranted before allowing GenAI to make changes on production devices. Trust must reach a defined threshold before broader adoption is considered.

The critical decision point arises when organizations consider whether AI should automatically execute a fix once it has identified a solution. This is where human intervention becomes essential. Allowing GenAI to make functional changes on network devices without guardrails is risky at the current stage of artificial intelligence development and should be avoided.

One effective mitigation strategy is layered validation. An AI agent may gather and present information, but instead of acting on the first response, a human can review it, or a second AI agent can independently validate the output. This approach provides checks and balances for areas prone to error, such as calculations, percentages, or inferred relationships.

Just as self-driving cars still include steering wheels and manual overrides, AI systems require clear control mechanisms. Autonomy without oversight is not a feature; it is a liability.

See also: RPA vs. AI Automation: Is Robotic Process Automation Being Replaced?

Building Confidence in AI with a Stepped Approach

As GenAI matures in network environments, the safest starting point is to tightly couple AI to existing automation systems, where AI is not the execution engine but instead directs another system, such as a script or orchestrator, to perform the change.

The first step involves defining the automation job and its scope, establishing guardrails, and exposing that limited, human-created job to the AI. Initially, Artificial intelligence should be granted access only to narrowly scoped, low-risk operations. As reliability is demonstrated over time, additional complexity can be introduced.

This approach requires execution engines and orchestrators to provide a tooling layer that AI can interact with. Technologies such as MCP (Model Context Protocol) can serve as this abstraction layer, exposing well-defined tools to AI agents while enforcing constraints around how and when they are used.

Limiting early use cases to “safe changes,” such as updating interface descriptions based on traffic type, allows organizations to build confidence incrementally. As outcomes become more predictable and deterministic, the scope of tools available to AI can expand.

GenAI can also play a valuable role in strengthening deterministic automation itself. It can assist in building automation jobs, generating documentation, and writing tests that validate expected outcomes. For example, an AI co-pilot can help create an automation job along with its test coverage, ensuring that when an AI agent later triggers that job, the resulting change remains deterministic.

With guardrails, checks and balances, and a stepped approach in place, organizations can embrace AI rather than fear it. AI’s ability to accelerate troubleshooting and reduce time to root cause can significantly improve operational efficiency. When analyzing massive datasets, such as a 50,000-line debug output or a packet capture, AI can surface insights quickly and without operational risk, even if the answer is not perfectly precise.

The most effective mental model is to think of AI as an autopilot rather than a self-driving. Autopilot excels during long, stable stretches but still relies on human oversight for critical decisions. By placing the appropriate level of trust in the technology for the appropriate tasks, organizations can leverage AI where it excels while allowing it to mature safely over time.

Trust Comes Before Autonomy

AI will continue to improve, and its role in network automation will expand. But autonomy without trust is not progress. The most successful organizations will treat GenAI as an accelerant for human decision-making, not a replacement for deterministic execution.

By clearly separating reasoning from execution, applying guardrails, and adopting a stepped approach, teams can safely integrate AI into their automation workflows. Trust is not granted all at once; it is earned through predictable outcomes over time.

Those who approach AI with discipline rather than hype will be best positioned to benefit as the technology matures.