APRICOT 2026 Indonesia’s Universitas Islam conducted experiments that found using generative AI vastly reduces the cognitive load on network pros during IPv4 to IPv6 migrations, but that organizations may not be ready for both AI and the new network protocol.
The university’s CIO Mukhammad Andri Setiawan, who also teaches at its Department of Informatics, discussed the experiments on Tuesday at the Asia Pacific Regional Internet Conference on Operational Technologies (APRICOT) in Jakarta, Indonesia. Setiawan explained that the university created a tool called “Net AI Copilot” that converts existing IPv4 implementations to dual-stack IPv4/IPv6 configurations and also generates ready-to-run Ansible playbooks to make the move a reality. The tool includes configuration validation checks and automatic rollback triggers to help prevent errors.
The CIO and his team sought to answer three questions:
Does generative AI reduce the cognitive workload experienced by network engineers during the complex IPv6 renumbering process?
How does AI-assisted configuration compare to traditional manual methods in terms of error rates and time-on-task efficiency?
Do technical improvements from AI tools directly translate into organizational readiness for broader IPv6 adoption?
Setiawan told The Register investigating cognitive load was important because some network engineers find 128-bit IPv6 addresses harder to work with than the dotted quads of IPv4. Some are therefore reluctant to work on migrations and find them tiring. The CIO felt testing for organizational readiness was a worthy line of inquiry to see if it is a factor in the slow pace of IPv6 adoption.
In his conference presentation, Setiawan said the university asked a group of seven experienced network engineers who should have been capable of a network migration to test a configuration manually, then with Net AI Copilot. In that test, subjects reported using AI reduced cognitive load by 65 percent, a measurement derived by asking participants to rate their frustration, effort levels, and mental and physical exertion. Participants reported that using AI reduced manual work and shrank the time required to complete the task from 170 seconds to five. The test subjects finished all assigned tasks, improving on a 71 percent completion rate using traditional working methods. Accuracy improved from 65 to 100 percent.
The second test, which involved an interactive simulation of a migration, saw cognitive load fall by 72 percent, completion time fall from 9.4 hours to 96 seconds, and firewall configuration errors disappear entirely.
The researcher then pondered why, if AI can improve network migrations, everyone isn’t using it already.
The CIO said researchers found organizations aren’t ready to use AI because they fear being unable to do so well or safely. Ironically, those are the same reasons that prevent migration from IPv4 to IPv6, leading Setiawan to suggest that developing operational readiness is the key to success with adopting both AI and IPv6.
Network admins need to read novels
In conversation with The Register, Setiawan said Universitas Islam uses AI to summarize its documentation and configuration data, and that doing so has improved the efficiency of his tech team to the point at which he no longer feels the need to immediately hire replacements for departing staff.
The CIO fully understands the possible implications of that choice but thinks his network engineering team can protect their jobs by becoming better users of AI.
To illustrate why, he used the example of AI alerting staff to network problems and potential security issues.
Some network engineers, he said, won’t get past using AI as an automated alert system – but he wants them to go deeper with AI.
“We need to have people that can understand a business process, not just do simple configurations.” He therefore thinks network admins need to learn how to use AI so it guides them towards actions that meet an organization’s strategic goals.
To get there, he thinks technology workers could do worse than to read more novels to help them understand how to build stories and spark ideas about how to make AI their co-authors.
Sovereignty and tokenomics
While IT workers worry about what AI might do to their careers, Setiawan is pondering how it will affect the university.
He currently finds the cheapest way to consume AI is with a simple subscription but worries that AI companies will lower the number of tokens they process for a set fee. He fears that by the time those price changes kick in, some of his team will have built automations and integrations the university cannot do without, meaning it will need to find more money.
Running AI on-prem is one alternative, but the CIO says quotes for setups that compete with SaaS-based AI can top $1 million.
He therefore thinks that organizations may need to contemplate shared and/or distributed infrastructure to insulate themselves from the high cost of AI, and the potential for lock-in and enshittification.
Setiawan jokingly suggested another approach to make AI pay: buy his own AI infrastructure, and quickly resell the memory and solid-state disks it contains.
“SSD is the new gold,” he said. “Last November and December we bought some, and now it is almost double the price.” ®