Network automation is becoming a must-have for enterprises as networks get more complex and AI projects multiply. Traditional scripting, machine learning and process automation are being augmented with generative AI, making automation more comprehensive and easier to deploy.
One of the largest U.S. insurers is using AI agents to analyze, simulate, and enforce rule changes to security policies in real time. In the past, rule creation and change reviews took hours and required manual validation by multiple analysts, says a senior technology officer at the insurance company. “Now, agentic workflows automatically interpret business requests,” he says.
The agents — which are policy automation agents from AI-powered cybersecurity startup Airrived — also assess the impacts across multi-vendor firewalls, validate compliance against internal policies, and generate approved rule sets. And it just takes minutes. “The impact is 90% faster policy changes,” the insurance company’s executive says. “Three-fold reduction in configuration errors.” And rule propagation across hybrid environments takes place in real time, he says.
This is already translating into business benefits. “We’re seeing efficiency gains that are allowing us to do expense avoidance and flatten our expense curve some, which is positive from an ROI perspective,” he says.
Efficiency gains don’t mean that AI is replacing IT people in droves, however. For example, having the automated firewall rule changes in place frees up half of one full-time employee at
the insurance company. “But, like most companies, that doesn’t go directly to the bottom line because you’re absorbing them into other work,” the executive says. “You’re either getting more done with less, or you’re deferring expenses and not having to take on additional headcount.”
Looking ahead, the insurance company plans to increase its AI spending next year. As the technology evolves, more vendors are expected to step up — not just startups like Airrived, but also existing platform vendors, the executive says. “Pretty much every SaaS platform out there is talking AI and agents,” he says.
Networking is changing. AI is a big part of it.
Networks are becoming more complex, and AI projects are adding to the challenges. Network automation is imperative, and enterprises are looking to augment their traditional automation technologies, such as scripting, machine learning, and robotic process automation, with generative AI-driven capabilities that can make automation easier to deploy, more capable, and more adaptable.
According to an Enterprise Strategy Group survey of 400 networking professionals released in late November, 93% feel that network automation will be essential for keeping up with change — and 89% say that networking is becoming increasingly important as a result of AI. In particular, respondents are looking for proactive outage prevention, predictive maintenance, automated discovery of vulnerabilities and policy violations, reduced human error, and better visibility across hybrid networks.
But only about one-third of companies have fully automated monitoring and visibility, policy enforcement, or diagnostic and troubleshooting workflows. Slightly more — 45% —have fully automated updates and patches. And the rest? They are on their way, with the majority already partially automated.
In terms of tools, three-quarters are using the automation tools provided by network equipment vendors, two-thirds are adopting tools from third-party vendors, 61% use open source, and 50% write their own automation scripts.
All this is possible with traditional technologies —but 99% of respondents say that gen AI will boost the benefits.
The biggest benefit that networking professionals expect to see from gen AI, at 56%, is improved security policy compliance. Accelerated troubleshooting and accelerated testing and auditing tied for second place, at 51% each, followed by improved network policy compliance at 50%.
“We’re using agents to make decisions, but not at all levels. We use it, but at a scale that you can trust.” —Ahmed Abdelaziz, vice president of automation and transformation, Rakuten Symphony
Building confidence in decisions led by AI agents
Solid early use cases for generative AI in network management include root cause analysis, troubleshooting, and similar research-driven tasks.
Ahmed Abdelaziz has been with Rakuten Mobile since its inception six years ago, as the vice president of operations. Rakuten Mobile is a subsidiary of Rakuten, a Japanese conglomerate that offers e-commerce, internet connectivity, travel booking, entertainment streaming, ebooks, credit cards, bank accounts, stock trading, insurance, and mobile phone services.
“We launched in 2020, during the pandemic,” Abdelaziz says. “And we couldn’t get resources to run our network. So why don’t we try to build something?”
Rakuten invested in automation from the start, he says, and, after three years, spun off this technology as Rakuten Symphony, where Abdelaziz is now the vice president of automation and transformation. Its network automation technology dates back to pre-generative AI, back when it was just scripts and machine learning. It’s all merging, Abdelaziz says, and each technology has its place.
“We’re using agents to make decisions,” Abdelaziz says, “but not at all levels. We use it, but at a scale that you can trust.” He points to recent global outages as examples of the kinds of consequences that he’s trying to avoid.
“We have an agent that looks at the previous histories of similar issues, and says, this is a decision we can take with a confidence level of X — and if it’s under 90% confidence, then we send it to an engineer,” Abdelaziz says.
The way the full process works is that the raw data feed comes in, and machine learning is used to identify an anomaly that could be a possible incident. That’s where the generative AI agents step up. In addition to the history of similar issues, the agents also look for other relevant context information, such as other incidents on the network, research possible diagnoses, do root cause analysis, plan a remediation, calculate the confidence level of its recommendation, and explain the basis for that confidence number. And in this process, it’s not just one agent, but multiple agents checking each other’s work.
If the confidence level is high, the agent triggers an action. “We’ve done automation for a long time, and we have a library of actions,” says Abdelaziz.
If the confidence isn’t high enough, and if the action can have a big impact, it goes to a human being, where the generative AI has enriched the ticket fields. If the engineer agrees with the diagnosis and approves the recommendation, that decision is then fed back into the system for future learning.
Right now, this agentic system is only used for limited use cases, not the entire network. For extra security, the automated actions are scheduled during maintenance windows, so there’s no impact on customers.
“We’re doing it gradually,” Abdelaziz says. Over the past year, the agentic system has processed around 6,000 incidents. At the beginning, its success rate was around 88%, he says, and it’s now at more than 95%.
Next, the company is working on reducing its energy footprint by using agents to make energy decisions without jeopardizing network quality. And this is just the beginning. “I believe the best uses are yet to be discovered,” says Abdelaziz.
The current state of network automation
To date, most companies have only partially automated their network management, according to a survey of nearly 700 networking professionals presented in October to the North American Network Operators’ Group. The survey respondents were mostly from companies with 10,000 employees or less, and with networks with less than 5,000 switches.
At the extremes, 9% are still fully manual, and fewer than 1% are fully automated. Roughly 46% are 10% to 30% automated, and 28% are 40% to 60% automated. Only 16% are 70% automated or higher.
More specifically, 31% use automation in configuration deployment, 17% for monitoring, 8% for configuration management, and 5% for semi-autonomous remediation.
When network management is automated, it’s typically with traditional tools. According to the survey, only 3% of companies are in production with AI-powered network automation, 14% are in development or testing, and 34% are thinking about it. Meanwhile, 45% have no current plans or initiatives for using AI.
Larger companies are further ahead.
According to a Dimensional Research survey of more than 1,300 professionals working in networking, operations, cloud, and architecture at medium to enterprise-sized companies, conducted on behalf of Broadcom, 22% of companies are beginning to use AI for automating network operations, and 5% have deployed fully autonomous AI-powered automation. The vast majority, 70%, aren’t yet using AI but are using some form of traditional scripted automation and playbooks. Just 3% are fully manual.
But interest is extremely high. According to Broadcom survey, 98% of companies are interested in deploying AI-enabled network observability technology, and 23% already have at least one such solution in production. Another 49% are in the evaluation or development phases.
And, of course, companies where networking is their primary business, like Rakuten Symphony, are furthest ahead. Another company that falls into this category is Zscaler, a cloud security company.
Zscaler operates a global network with 15 million users, and it has about 100 people on its network management team. The company has been using generative AI to bring down the mean time of outage detection and response.
A human can take minutes to carry out an action that AI can do in seconds, says Dhawal Sharma, Zscaler’s executive vice president of products. “For example, before, humans had to do most of the work of root cause analysis. Now AI is doing most of the work and the human just approves it.”
Zscaler started boosting its network visibility investments in 2024 and followed that with automated remediation capabilities in 2025, Sharma says. The result is that the time to detect and respond has improved by a factor of four to seven, Sharma says, but only because Zscaler already had a lot of automation in place before. For the average company deploying similar technologies, the time savings could be 15 to 20 times better, he says.
For low-risk situations, the remediation process can be automatic, he says. “For mission-critical systems, you always want humans to validate and verify.”
Another company that isn’t waiting to adopt AI for network automation is cybersecurity company N-able. “We always want to be innovators versus imitators,” says Will Ledesma, the company’s senior director of managed detection and response cybersecurity operations.
“We monitor networks and utilize network telemetry,” Ledesma says. “We were taking 60, 90, 120 minutes to find root cause. Using AI brings us down to seconds.”
N-able uses LLMs and agentic AI to power its systems, and it uses both commercial, off-the-shelf tools and ones built internally. The company has already added AI to 70% of its incident and threat remediation capabilities, Ledesma says.
Challenges to adoption
AI accuracy and reliability is just one of the obstacles keeping companies from deploying AI more broadly for network management. But it’s a big one.
According to the Dimensional Research survey, 71% of respondents have limited trust in AI-based functionality for network operations. As a result, the majority of tasks they’re considering for automation have to do with data collection, information sharing, and other simple, low-risk activities.
Another obstacle is the variety of tools that are being used.
For example, enterprises currently use multiple vendors and tools for network performance management and monitoring. According to a Riverbed survey of 1,200 IT leaders and technical specialists released in September, the average company uses 3.9 tools for this function alone, from 2.7 vendors. And then there are observability tools for digital experience monitoring, cloud management and monitoring, application performance management and monitoring —the list goes on. In total, companies have an average of 13 observability tools from nine different vendors.
When it comes to smaller companies, the challenges are even more dramatic.
According to the NANOG survey, 43% of organizations have zero dedicated automation staff, and 27% cite skills challenges as their primary barrier to adoption —followed by organizational challenges at 20% and cultural challenges at 14%. Technical challenges are in fourth place, at just 10%.
As more vendors start to add AI features to their platform, these challenges may become less significant.
And demand for network automation will continue to increase, Gartner predicts, as companies place higher demands on their networks. “Enterprises are increasingly using agile, DevOps and infrastructure as code approaches to meet infrastructure delivery requirements, which demand increased network automation,” the analyst firm says.
