AI Network Tools: The Case Studies Are Vendor-Published, the Surveys Are Not

How CodoraTech is funded: CodoraTech is supported by advertising and, in some articles, by affiliate links. Where an article contains affiliate links, we say so at the top of that article.

Ask what AI has done for enterprise networking and two incompatible bodies of evidence come back. One is a set of large percentage improvements published by Juniper, Cisco and Nile about their own products and, in Cisco’s case, about its own IT department. The other is survey work from Enterprise Management Associates and IDC, which finds adoption climbing while measured success falls. Both are worth reading. They are not the same kind of evidence, and keeping them apart matters more than any single number in either.

Every headline percentage here was published by the company selling the product

The most quoted named-customer results in AI networking come from vendor marketing documents. That does not make them false, but nobody outside the vendor and its customer has checked them, and no independently audited AIOps customer outcome could be located at all.

Juniper’s ServiceNow study, the first entry below, also carries a customer line from Venkat Lakshminarayanan, vice president of infrastructure and operations at ServiceNow: “When we first implemented Juniper, we saw a 90% drop in issues related to wireless.” The quote and the percentage sit on the same vendor page.

One frequently recycled example does not belong in this category at all. Cisco’s Kamstrup case study of 6 August 2025 describes SD-WAN, ThousandEyes, SD-Access, Firepower and Umbrella. That is network automation and assurance, not an AI product, and the quote from network administrator Palle Lyng Raun makes the point himself: “After installing Cisco ThousandEyes, we were able to prove that there was a peering problem between two ISPs.” Good telemetry found a fault. No model was involved. The three deployments that do carry AI claims and published numbers, with their publishers, are these:

• ServiceNow, per a Juniper Networks case study dated October 2023: 60 percent cost avoidance in network capital and operating expenditure over three years, a 90 percent reduction in wireless-related employee issues, 50 percent faster network deployment and a 60 percent reduction in switch port density, after deploying Juniper Mist cloud AI across wired, wireless and WAN with the Marvis Virtual Network Assistant.
• Cisco’s own IT organisation, per a Cisco on Cisco page dated 11 June 2026: Catalyst Center with Cisco AI Network Analytics and Splunk Enterprise across more than 200,000 devices, a 97 percent reduction in code vulnerabilities, 59 percent shorter upgrade times and a mean time to detect of 41 seconds. Cisco is both the vendor and the customer in that account.
• Nile, per its own March 2025 announcement: a campus network-as-a-service with AI-driven automation and zero-trust architecture deployed at LEAP 2025 in Malham, Saudi Arabia, covering 2 million square feet across seven halls for more than 200,000 attendees, installed in three days with four on-site installers. Nile published no uptime, fault or user-experience figures afterwards.

Thirty-seven percent of network alerts indicate a real problem

For an independent counterweight, the useful dataset is EMA’s Network Management Megatrends 2026, a survey of 352 IT professionals in North America and Europe reported by Denise Dubie in Network World on 8 June 2026. Its central finding is that only 37 percent of network alerts indicate a real problem. Nearly two thirds of what reaches an operations team is noise, and that is the environment any AI layer works inside.

The surrounding numbers describe teams under pressure rather than teams being relieved of work. EMA found 58 percent of network problems detected proactively before users are affected, 28 percent of problems caused by manual administrative errors, staff spending 29 percent of the day troubleshooting, and 52 percent finding it difficult to hire network experts, up from 26 percent in 2022.

Shamus McGillicuddy, EMA’s vice president of research, summarised it as a support problem: “Network operators clearly know they need to do better, but they aren’t getting the support they need.” A monitoring architect at a Fortune 500 entertainment company, quoted anonymously in the same coverage, gave the reason: “What used to be done by a 25-person team, management now wants us to do with a ten-person team.”

Complete success fell to 31 percent from 42 percent

The most awkward figure in the same EMA survey is the trend line. Only 31 percent of respondents reported a completely successful network operations strategy, down from 42 percent two years earlier, so self-reported success went backwards by eleven points over exactly the period in which AI tooling spread through the category.

EMA’s dedicated AI-Driven NetOps research, published on 20 January 2026 and based on 458 IT professionals, breaks that down. Only 35 percent report complete success with AI-driven network management initiatives, only 39 percent are completely confident evaluating AI-driven solutions, and only 44 percent have full confidence in their own network data quality. At the same time 59 percent already use AI features from their network management vendors and 52 percent train AI models on their own IT and security data. EMA calls the distance between those two sets of numbers a significant execution gap.

McGillicuddy names the mechanism: “Network data quality is the AI killer…Today’s network operations struggle with a variety of data issues, including data collection errors, poor documentation, and proprietary data formats.” That reconciles the two halves of this article. The vendor case studies describe results on estates that are unusually well instrumented, run by the vendor itself or by an engineering-heavy customer. The survey data describes everybody else, feeding a model documentation that is out of date and telemetry in formats that do not agree with each other.

Adoption moved, maturity did not

IDC’s 2026 AI in Networking Special Report, reported by Network World on 6 April 2026, found that AI adoption maturity had not shifted in 18 months. Research director Mark Leary put it flatly: “The people who were at select use were still at select use. The people who were at substantial use were still at substantial use.” The same report found 81 percent of organisations increasing spending on managed service providers to support AI initiatives, buying in capability they do not have, in what IDC calls a widening gap between intent and execution. IDC’s sample size was not disclosed in that coverage.

Gartner reached a comparable position earlier. In a press release of 18 September 2024 it predicted that 30 percent of enterprises would automate more than half of their network activities by 2026, while simultaneously placing intelligent automation for infrastructure and operations in the Trough of Disillusionment on its hype cycle. Frances Karamouzis, distinguished vice president analyst, added the measurement problem: “less than 20% of organizations have mastered the measurement of hyperautomation initiatives.”

None of this argues against buying the tooling. It argues for one question at the point of purchase, asked of every number in a sales deck: who measured this, and were they selling something at the time?

The failure modes are not new either. CIO.com catalogued them on 21 December 2021, quoting analysts at Gartner, Forrester, Omdia and McKinsey. That article is older context, but its list still maps onto what EMA measured four years later, and it remains a serviceable checklist:

1. Adopting AIOps with no strategy set before purchase.
2. Poor or incomplete data feeding the system.
3. Inadequate coverage, so the model sees part of the estate.
4. Paying twice for toolsets that duplicate existing monitoring.
5. Optimising components while missing the big picture.
6. Culture change and distrust of AI: 22 percent cite fear or distrust, and a typical deployment runs about 16 months.

Sources: Network World · PR Newswire · Juniper Networks · Cisco · Network World