Larry Weidig
AI isn’t coming for your job. It is coming for your workflow.
That distinction matters, especially for ISPs. Most teams I know aren’t short on effort or expertise. They’re short on time. AI is showing up as a lever that gives someof that time back—if you apply it in the right places, with the rightguardrails.In this episode of Bandwidth, we stayed out of the “someday” conversation and focused onwhat operators are actually doing right now: where AI is saving minutes thatbecome hours, where it’s improving reliability, and where it can quietly createnew problems if you treat it like a set-and-forget tool.
Bandwidth is sponsored by Sonar, but the goal of the show is simple: share what’s working in the real world for the folks building and running networks.
The real-world AI wins we’re seeing right now
1) Tier-zero“office work” that frees up real operator time
The lowest-friction wins are the ones that don’t require specialized data or custom integrations:
- Drafting and replying to emails (with a human edit pass)
- Summarizing long threads and extracting action items
- Meeting transcription + automatic task lists
- Research that used to take days (vendors, competitors, markets)
That last one is bigger than it sounds. “Weeks of research” turning into “a day of review” changes how fast you can make decisions—especially when you’re balancing operations, growth, and the usual fires.
The key: treat AI like a fast first draft plus a faster search assistant. You still own the final output.
2) Contact center intelligence that actually helps customers (and your team)
Georgette called out the industry-specific shift we’re seeing: AI going beyond “chatbot on the website” and into the heart of the support operation.
Examples that are landing well:
- Call analysis for the calls already flowing through your PBX
- Sentiment detection and pattern spotting across thousands of interactions
- Surfacing upsell opportunities (when it’s appropriate and not tone-deaf)
- Giving CSRs “fingertip access” to SOPs and account context:
- tickets opened in the last 90 days
- recent truck rolls
- chronic issue patterns tied to the account
The big unlock isn’t that AI “replaces a CSR.” It’s that it shortens the time to context. Support improves when the person helping the customer doesn’t have to hunt through five systems and three browser tabs to figure out what’s going on.
3) Field operations: routing, scheduling, and wasted miles
Routing optimization is not glamorous, but it hits real costs:
- fuel
- vehicle maintenance
- technician time
- missed appointment windows
Static data(office location, service address) plus changing conditions (traffic, weather,constraints) is a good fit for AI-driven optimization. That’s not science fiction. Operators are already testing it because the ROI is measurable.
“We don’t have time to learn AI” is a warning light, not an excuse
This came up because it’s real. Small operators often feel like there’s no margin anywhere.
But when teams say “we don’t have time,” what I hear is: “We don’t have a plan.”
A plan doesn’t mean an eight-hour training day. It can be:
- 10 minutes a day
- 30 minutes a week
- One use case per month per department
The point is consistency.
If leadership wants AI adoption, leadership has to participate. It doesn’t work when the message is, “You all go figure this out.” That approach creates uneven adoption, uneven risk, and very uneven results.
The hidden blocker: “garbage in, garbage out” is still true
The uncomfortable part: some operators can’t get to the bigger AI wins because their data foundation isn’t ready.
If you’re notlogging tickets, if support is still operating off informal notes, if accounthistory isn’t structured—then AI can’t correlate what it needs to correlate.
That doesn’t mean you wait.
It means you“plant a flag” and start collecting clean data going forward:
- Pick a ticketing process and stick to it
- standardize resolution codes
- define what gets tagged and when
- Make sure account notes are consistent and searchable
Even if your historical data is a mess, your next 90 days don’t have to be. And those next 90 days become the base layer for everything you want to do later.
Where ISPs should aim first: churn prediction
In the episode, we kept circling back to churn for a reason: it’s where AI’s strengths line upwith ISP business reality.
AI is good atpattern analysis—finding signals that are too subtle or too spread out forhumans to reliably catch early. Churn is a pattern problem.
And churn is expensive:
- lost MRR
- Higher CAC pressure
- compounding brand damage when the churned customer leaves a bad review on the way out
Fraud detection matters too, but for most Tier 2/3 ISPs, the immediate business impact of churn is usually bigger than the typical fraud hit. As you scale, fraud risk grows. But churn is the steady leak that flips the boat over slowly.
If you’re choosing one direction to point your energy first, churn is a strong place to start.
AIguardrails: what “responsible use” looks like inside an ISP
This is the part that can’t be skipped.
1) Corporate accounts, not personal accounts
If you’re putting any customer, network, billing, or internal operational details into a model,it needs to be under corporate control.
Even with anonymization, “oops” happens. And “oops” with customer data is not the kind ofstory you want to tell.
Corporate accounts also make it easier to:
- manage access
- apply retention policies
- enforce usage rules
- audit what’s happening
2) An internal AI usage policy
Policy sounds boring until you need it. You need something that covers:
- What data is allowed, and what is not
- approved tools and accounts
- anonymization expectations
- who can do what (roles and permissions)
- consequences when someone ignores the policy
This also helps with training because you’re not asking teams to guess what “safe” looks like.
3) Team sharing and “AI show-and-tell” meetings
One of the best ideas we talked about: regular internal sessions where teams share:
- How they used AI this week
- What worked
- What went wrong
- What risks they spotted
This does two things:
- spreads good practices fast
- surfaces red flags early
Screening resumes with AI: where it helps, where it hurts
We hit this in the “Love it or hate it” segment, and the nuance matters.
Screening: AI can be useful to remove obvious non-starters (wrong location, missing required certifications, etc.). It can shrink the pile.
Vetting: AI is still likely to miss great “fringe” candidates—people who don’t look perfect on paper but are strong problem-solvers, curious, and great in the role.
My concern is simple: if you lean too hard on AI early, you’ll optimize for “good resumewriters” and filter out the people you’d be happiest you hired six months from now.
Use AI to reduce noise. Don’t use it to decide who has grit.
The KPI trap: faster ticket resolution with angrier customers
This hypothetical made all of us tense for a reason.
If a chatbot is closing tickets faster but frustrating customers, you’re watching a bad trade happen in real time. That frustration doesn’t stay inside the support queue. It shows up as:
- repeat contacts
- escalations
- negative social posts
- churn signals
The fix is not“turn it off forever.” The fix is operational discipline:
Step 1: Measure CSAT on chatbot interactions
If you don’t measure satisfaction the same way you measure your human team, you’re guessing.
Step 2: Find the friction source
Is it:
- tone
- phrasing
- too technical / not technical enough
- forcing the wrong flow
- failing to recognize when to hand off to a human
Step 3: Put a human in the loop during rollout
One of the smartest rollout approaches we discussed: enable the chatbot only during office hours for an initial period (90 days was the example), so staff can monitor outcomes, catch failures, and adjust before it becomes 24/7.
That’s the right instinct: you keep control while the system learns and while your customers learn how to use it.
What to do next if you’re an operator
If you want something practical to take back to your team, here’s a straightforward sequence:
- Pick one low-risk use case (email drafts, meeting summaries, research) and require a human review.
- Write a one-page AI usage policy and require corporate accounts for anything work-related.
- Start cleaning data now (tickets, tags, account notes). Don’t wait for perfect history.
- Choose one high-impact operational target (churn is a great first bet) and define how you’ll measure success.
- Build a feedback loop (weekly or biweekly). AI outcomes need tuning.
AI moves fast. Your plan will change. That’s normal.
What you can’t do is sit idle while the gap widens between operators who are learning in public and operators who are still waiting for the “right time.”
Episodelink
The episode is available in the blog post with links out to Spotify, Apple Podcasts, and the Bandwidth YouTube channel.