Consider the customer whose AI prompt times out repeatedly, moments after a successful speed test. The network operations center reports healthy infrastructure. Latency is within target. There are no outages. Somewhere between those measurements and the failed request, the provider’s account of the service has stopped explaining the customer’s experience.

The customer still needs help.

For broadband operators, the last twenty meters between the gateway and a subscriber’s device are a useful way to frame that problem. Trouble here can turn a working connection into a support call. If it keeps happening, the provider has a retention problem to investigate, even when its network metrics look good.

In a 2025 survey of U.S. telecom customers, 69% cited reliable network connectivity as a top decision factor, narrowly edging out affordability.¹

The industry has spent decades building the infrastructure and monitoring systems broadband depends on. That investment was necessary. Operators need to know when capacity is constrained, or a node is failing. Cloud-managed Wi-Fi has since extended their view into the home, giving them access to device behavior, Wi-Fi conditions, and application performance that were once difficult to observe.

That visibility has already transformed day-to-day operations. Many of the issues that once generated support calls, from Wi-Fi channel selection and device steering to routine security threats, can now be identified and addressed automatically. For a large share of subscriber interactions, the network can increasingly optimize itself.

The remaining issues are different. They are less frequent, but significantly harder to diagnose because they don’t point to a single root cause. A Zoom meeting freezes even though the speed test passes. A Claude AI prompt times out despite healthy Wi-Fi. A household reports “slow internet,” but the underlying issue could involve the application, the device, RF conditions, or something outside the provider’s network altogether.

The challenge is becoming more urgent as AI evolves. Our recent research found that 86% of homes in Plume’s dataset now interact with a large language model at least once a quarter, and two-thirds do so regularly. But the next wave isn’t simply more AI conversations. It’s AI agents that perform increasingly complex tasks on behalf of users, creating longer-running workflows, more application interactions, and more opportunities for something to go wrong.

Together, these harder-to-diagnose issues create what we call the experience gap: the difference between observing that something went wrong and understanding why it happened well enough to act and improve it. The challenge is no longer collecting signals. It is determining which signals matter, connecting them into meaningful context, and doing so while there is still time to improve the subscriber’s experience.

---

Inside the Home

The average household in Plume’s dataset now has more than 22 connected devices. Increasingly, those devices are interacting not only with people, but with AI services acting on their behalf. On any given day, a laptop joins a video call while someone else streams a match in 4K. A PlayStation console competes with a cloud backup. A Nest camera uploads footage in the background. The mix changes as people move between rooms and devices, often without thinking about the connection.

This is where the experience gap becomes visible.

A subscriber trying to finish the yearly objective-setting session with their boss on Teams has little reason or ability to distinguish an access-network problem from a Wi-Fi or device problem. The meeting is being interrupted. A diagnosis that stops at a successful speed test leaves that interruption unexplained.

A speed test alone cannot reconstruct what happened during the call. That requires relating the relevant measurements to the device, the application, and the time of the interruption. Even then, the cause may lie outside the provider’s control. The useful outcome is a diagnosis that narrows the problem enough to guide the next step.

A stalled AI workflow introduces another kind of symptom to investigate. Unlike a single prompt-and-response interaction, an AI agent may complete dozens of coordinated actions across multiple cloud services before returning a result. If one step fails, the operator still has to determine whether the interruption originated in the home, the access network, the application, or somewhere else entirely.

Operators now have access to network, Wi-Fi, application, device, and customer data that would have been difficult to assemble a decade ago. A single interrupted call can leave evidence across several of those systems. The work is in connecting that evidence. Which signals explain the interruption? Has it happened before? Does it persist after the apparent fix? Those answers give a support agent something to work with when the customer calls.

---

Where the Information Stops Being Useful Across the Board

Customer care makes the cost of incomplete context visible. An agent may begin with little more than “unhappy customer on line 2.” The subscriber describes the symptoms, reboots equipment, and repeats a speed test. These steps can be useful. Their value is harder to defend when the provider has already collected information that could have narrowed the investigation.

The test for an operator is whether that information reaches the person handling the call in a usable form. Collecting device and Wi-Fi data somewhere in the business does not mean an agent can connect it to the interruption the customer is describing.

Operations teams face the same difficulty when a healthy access network leaves them with no clear next step. They need to follow the problem beyond the modem and identify which part of the connection warrants attention. A dashboard can show where a measurement crossed a threshold; someone still has to decide what that means for this subscriber.

Commercial decisions depend on that distinction, too. A household struggling with coverage and one that needs more capacity may both report slow internet. Treating those complaints as the same upgrade opportunity risks ISPs recommending a service that leaves the original problem unresolved.

For product teams, the question becomes whether a service changes the experience it was designed to improve. Leadership needs to know whether spending on that service reduces avoidable support work or helps retain subscribers. Answering either question requires connecting what happened in the home, what the provider did, and what happened afterward.

More telemetry alone cannot supply that answer. Nor can an analysis delivered after the customer has called again or canceled.

---

What Happens After an Alert

An alert can help an operator intervene earlier. It can also become another item in a queue. Its value depends on the response it makes possible.

If each alert requires someone to assemble the same evidence before deciding what to do, the provider has made the problem more visible while leaving much of the investigative work intact. As the subscriber base grows, that work grows with it unless the operating process changes.

A useful operating model would bring the relevant evidence together and help teams decide which problems need attention. Where a remedy can be applied safely, the provider should be able to act and check the result. Where the cause remains uncertain, the agent should inherit the investigation already completed rather than ask the subscriber to start again.

Preventing a support call is a demanding standard. The operator must identify a developing problem, choose an appropriate response, and confirm it helped. A prediction without a workable intervention adds another alert. An intervention without a check leaves the provider guessing about its effect.

---

A New Operating Model for Broadband

An operations leader should be able to ask why customer satisfaction dropped across a region and get help investigating. Today, that question may send people searching through dashboards, reports, and logs before they can even decide where to look.

Consider how that investigation could work. The system identifies the affected households and finds that most share a gateway model. Reboots increased during a heatwave, with failures concentrated on a particular firmware build. The operations team can examine the evidence and test a specific explanation. Engineering receives an investigation it can act on, including a proposed remedy to evaluate. What begins as an NPS question ends with a concrete engineering action.

Commercial teams need that depth of understanding when deciding which households should receive an upgrade offer. Video calls that deteriorate during working hours might suggest demand for a work-from-home service. But if weak Wi-Fi coverage is causing the interruptions, the offer needs to address coverage. More bandwidth alone would leave the customer paying more for the same problem.

An investigation could establish which households have a coverage issue and whether the proposed service would resolve it. Commercial teams could then make an offer with a clear rationale. Households needing a straightforward fix could be routed to care, with the evidence already attached. What begins as a support symptom ends with a targeted revenue opportunity.

These decisions require information to travel with the customer’s problem. A finding about a gateway fault should inform the next support conversation. An unresolved coverage issue should be visible to the team preparing an offer. After an intervention, the provider needs to know whether the interruptions stopped.

That kind of operating model changes when and how decisions get made. As AI evolves from answering questions to completing work, that shift becomes even more important. Operators won’t simply be supporting more connected devices. They’ll be supporting more autonomous activity across those devices.

Operators should be able to put that understanding to work across their business. Soon, they’ll have a new way to do it.