Extreme's fix for AI networking starts with reconciling four fields all named client ID

Extreme sat network administrators down in front of an open chat box and watched them do almost nothing with it ^[9]. The question they eventually landed on was how many devices they had. Bukhari called that "a completely useless question, because it's right there on the dashboard" ^[10]. Then, when a real incident hit, the engineers forgot the AI existed and went back to old habits ^[11].

The data problem under that behaviour is specific and checkable. Inside Extreme's own portfolio, the fabric, switching, Wi-Fi and SD-WAN databases each carry a field called client ID, and each one means something different ^[6]. Feed all four to a model unnormalized and it has no idea what it is looking at. "Unless you absolutely fix your data and have a custom-built context layer, it's never going to be good enough for a network admin to actually use it," Bukhari said ^[7]. Extreme's answer is a purpose-built context layer plus a knowledge graph mapping relationships across users, devices, applications, services and network conditions ^[8].

The figure Extreme has put behind the release measures something else. Bukhari said that "in the first eight days, we have seen about a 900% increase in interaction" ^[15], which is roughly ten times the prior level ^[16]. What changed in those eight days is the delivery. The agent is always running and reaches out instead of waiting to be asked, through a Nudge bar that blinks once it has completed a pre-investigation ^[14]. An agent that interrupts you gets used more than a box that sits there. Whether the incident closed faster is a different measurement, and on that the story rests on unnamed beta customers who told SiliconANGLE that resolution time falls from hours to minutes and that onboarding time for new engineers is cut by as much as half ^[17].

The part a pilot can actually falsify is the refusal. Extreme built what it calls an awareness scale, so the agent decides quickly whether it can help, says plainly when it cannot, and then explains what it can do instead, often by packaging the evidence it collected and opening a support ticket for the user ^[12]. In an analyst pre-briefing, Michael Jones, Extreme's vice president of AI, asked whether Agent ONE could remediate wired devices; it said no, then listed what it could do ^[13].

This is for shops already paying for Platform ONE, because it arrives inside the existing subscription with no new line item ^[1]. Nobody has to sign a purchase order, which also means nobody formally owns the rollout.

Two tests, then, before a pilot closes. Pick two field names that appear in more than one of the vendor's modules and ask what each one means and where the reconciliation lives; Extreme has 18 months of its own production data behind Platform ONE's first-generation generative AI to draw an answer from, and a vendor without that history will show it here ^[5]. Second, insist on two numbers next to the interaction count: time from alert to closed ticket, and the share of nudges the engineer dismissed. If most nudges get dismissed, the interaction count is measuring interruptions.