Key Takeaways

• AI in ISP customer support can take on several jobs at once: deflecting contacts, assisting the rep, automating tickets, and detecting network degradation. Where each fits depends on whether the subscriber needs a person and whether the tool has the data.

• An AI tool that can’t access the subscriber’s account and service data has limited ability to address their specific need; it can mainly route the contact and capture details for a person. On technical calls, that data has to include the network.

• At one fixed wireless operator, about 25 of 38 daily support contacts weren’t technical. Deflection can reach contacts like those; the technical calls need a diagnosis.

• When a subscriber has a problem, AI is most useful working alongside the rep, gathering what they need to answer so the rep can spend the call on the subscriber. A person also checks its answer before the subscriber hears it.

• Support contacts ran from about 0.1 to 0.4 per subscriber per month across three regional operators, so judge an AI tool against your own contact volume and call mix rather than anyone’s average.

What Does AI in ISP Customer Support Actually Do?

AI in ISP customer support can take on several jobs: deflecting contacts before a person picks up, assisting the support rep during the call, automating ticket work around it, and detecting network degradation early. You don’t have to choose between them; an operator can use several at once, each on the contacts it suits. Where each one fits comes down to two questions: does the subscriber need a person for this, and does the tool have the data to do the job?

Some contacts don’t need a person. A subscriber asking how much they owe and when it’s due wants a number, and an AI tool with access to the billing system can provide it at any hour, making it a good job for deflection. 

A subscriber with a problem is in a different position: their service isn’t working properly, they’re often frustrated, and they usually want human interaction. There, AI is most useful beside the rep, gathering everything the rep needs to answer, so the rep can spend the call on the subscriber: listening, explaining, and making sure it gets fixed.

The data decides how far either job can go. Billing and account answers come from the OSS and BSS systems; technical answers come from the network, in how the subscriber’s connection and the equipment around it have been performing. Some calls need both, since a “my internet is down” call can come from an account suspended for non-payment. An assistant that can’t see the network leaves the rep to troubleshoot using manuals and knowledge-base articles, which is slow for everyone on the call. That’s why the place to start is the assistant beside the rep, connected to the network data, rather than a chatbot in front of the subscriber trying to troubleshoot.

How much each job pays off comes down to the mix of calls your desk actually takes. In Preseem’s operator data for one fixed wireless operator, about 25 of 38 daily support contacts weren’t technical. The rest were the calls that ran long, were escalated, and sometimes ended in a truck roll.

For the rest of the support operation, see the complete guide to improving ISP support performance.

What Are The Types of AI in ISP Customer Support?

“AI in support” covers four categories. For each, the useful questions are what job it does, what it has to see to do that job, and where it falls short without network data.

Degradation detection is really a network operations tool that customer support benefits from, and it can’t run without network data. The other three can be deployed without it, and on technical calls they then fall back on generic scripts: questions, knowledge-base articles, reboots, faults classified from the subscriber’s words. That looks like troubleshooting, but it’s done without the evidence that would locate the fault, so much of what the tool costs to buy and run goes into calls it can’t resolve. Deflection on account and billing contacts is the exception, since the data it needs sits in OSS and BSS.

Which Calls Can Each AI Tool Actually Answer?

Deflection can reach the non-technical contacts, which were the majority at the fixed wireless operator mentioned above. A well-built virtual agent with OSS and BSS access can handle those at any hour.

A technical call is a different problem. “My internet is slow in the evenings” can mean congestion on a shared sector or PON, a degrading RF link or optical signal, bufferbloat under load, or a router behind the television. Each needs a different fix, even though the subscriber describes all four the same way. Answering it means looking at what that connection has been doing and how its neighbours compare, and deciding whether it’s the Wi-Fi or the network is often the first fork. Without that view, ticket automation sorts these calls by the words the subscriber used, and the most useful thing a deflection tool can do is get the call to someone who can quickly see the connection.

The technical calls are where escalations and dispatches come from, even when they’re the smaller share of the count. They’re also the reactive load that keeps ISP customer support teams stuck in firefighting mode. A tool that handles the non-technical contacts well can lower your contact count while leaving that work exactly where it was.

What Happens When A Chatbot Troubleshoots Without The Connection?

It asks the subscriber questions such as whether the lights on the router or ONT are green, whether it’s every device or just one, whether the problem occurs at a particular time of day, and whether they’ve tried restarting. Each answer adds to a picture the chatbot can’t finish, because what would settle the question is on the network side, in how the subscriber’s link has been performing and whether neighbours on the same sector or PON are seeing the same thing. The conversation goes into understanding a problem the tool doesn’t have the information to solve, and the subscriber’s time pays for it.

That’s a frustrating experience for the subscriber. Someone whose video calls have frozen every evening for a week has probably restarted everything already and got in touch for an actual diagnosis. A string of questions that ends with a reboot reads to them as a sign that nobody can see the problem, which is accurate. If the conversation then ends in a ticket, they tell the whole story again to a rep, so the minutes spent with the chatbot didn’t move the call forward.

None of this is a criticism of any vendor’s chatbot; it describes a configuration, and there are two ways to fix it. Connected to the subscriber’s connection history and the state of their sector or PON, the same tool could tell the subscriber that their part of the network is congested most evenings and what’s being done about it. Scoped down instead, it handles the account and billing contacts it can resolve and passes a contact to a person as soon as it turns technical, with what the subscriber has already said attached.

Why Start With The Assistant Beside The Rep?

The assistant works on calls that are already being handled by someone who can judge what it produces. With the subscriber’s recent experience, the state of the AP or PON, and the history of their radio or ONT in front of them, the rep can be told where the fault is and what to do next while the subscriber is still on the line. If the assistant’s view is incomplete on a particular call, the rep can catch that before the subscriber hears it. A tool placed in front of the subscriber has nobody in that position.

It also asks less of the subscriber. Someone who phones about a technical problem has chosen to reach a person. With the assistant, they reach a rep who already knows what’s wrong and doesn’t have to check scattered systems mid-call, so the rep’s attention stays on them. A deflection tool has to win the subscriber’s trust first, on exactly the calls where it’s least likely to see enough to deserve it.

The assistant’s value also scales with call volume, which varies widely between operators. In Preseem’s operator data, support contacts ran from about 0.1 to 0.4 per subscriber per month across three regional operators, so two operators of the same size can take very different numbers of calls each week. Measure your own contact rate and split it into technical and non-technical contacts. That tells you which category you’re actually considering.

Why Do ISPs Buy AI Support Tools Before Deciding What They’re For?

In Ian Khan’s AIRS AI Readiness Score, administered to 29 operators at the Proactive ISP Summit: AI in Action, experimentation was the highest-scoring readiness dimension at 73 and governance the lowest at 36. That 37-point spread is the gap between the appetite for trying AI and the discipline for managing it.

The AIRS sample was self-selected from AI-interested attendees at an AI-themed event, so it describes the leading edge rather than the sector middle. Scores are self-reported perception rather than audited capability, from 29 operators. The instrument and analysis are Ian Khan’s.

In support, that gap tends to show up as a tool bought for the category rather than for a specific call. Before anything is bought, name the decision it’s meant to improve: which calls it will take, what it has to see to take them, and who acts on its output. If a tool can’t be tied to one of your calls in those terms, there’s no way to tell afterwards whether it changed anything.

Frequently Asked Questions about AI in ISP Customer Support

Can AI customer service handle broadband outage calls?

Yes, and outages suit deflection better than most technical contacts, because the cause is already known and shared. If the tool knows which subscribers are on the affected AP, OLT, or sector, it can tell each of them what’s wrong and when service is expected to be restored, without their calls joining the queue. Without that topology, it can only post a general status message and hope the right people read it.

Will an AI chatbot reduce call volume at an ISP?

It can handle non-technical contacts, which, for some operators, make up most of the day’s calls. Technical contacts deflected without an answer can come back as repeat contacts, sometimes from a subscriber with less patience than the first time. Measure repeat contacts alongside deflection rate; otherwise a deflected call and a resolved call look identical in the report.

Do AI support tools for ISPs work across multi-vendor fiber and fixed wireless networks?

They work as far as their data reaches. Many regional operators run several radio and OLT vendors, often with equipment inherited through acquisitions, and a tool that reads one vendor’s portal can only see part of the network. Ask a vendor which access equipment, CPE and billing systems the tool reads today, not which ones are on its roadmap.

Who is responsible when an AI support tool gets a diagnosis wrong?

With an assistant beside the rep, the rep is responsible, and they can catch the error before acting on it. With a subscriber-facing tool, an error usually surfaces when the subscriber calls back or cancels, unless someone is reviewing the tool’s conversations. That’s a good reason to put AI in ISP customer support where a person reviews its output first, at least until you’ve seen how often it’s right on your own network.

What An AI Tool That Can’t Answer The Call Costs You

The budget spent on AI that can’t answer the calls it’s given is easy to miss, because the tool runs and the dashboards fill up while the technical calls take as long as they did. Meanwhile the cost of serving each subscriber stays where it was. The complete guide to ISP support performance covers how to measure what a change actually does to your desk.