AI text support platform for internet providers
The problem
I spent 20 years in the ISP business, from installs to executive vice president, through three acquisitions. Most support is the same handful of conversations on repeat:
Can I get service here?
What do I owe?
When’s my tech coming?
Is there an outage?
Those conversations take up a lot of a support team’s day, and customers with harder problems wait behind them. So I’m building a platform any ISP can plug in to handle them.
What I’m building
Customers text their ISP the way they’d text a person. Calling is next. AI agents (software that reads each text and takes the next step) work out what the customer needs and get it done:
- Signups
- Billing
- Plan changes
- Appointments
- Retention, when someone’s ready to cancel
Answers come back in seconds, any hour of the day. It also tries to work out why someone is reaching out before they say it. If a customer’s install is today, it opens with something like “Hi Hannah, are you texting about your install between 12 and 4?”
- Customer textsthe way they’d text a person
- Twiliosends and receives the texts
- Eve agentswork out what the customer needs
- AI modellocal or cloud
- Billing and business rulescode handles the money and the rules
- Replyin seconds, any hour
How it’s kept safe
The AI runs the conversation. Code handles the money and the rules. It can’t make up a price, skip a confirmation or promise something the system didn’t approve.
Every conversation gets graded against the ISP’s own policies and procedures, the way you’d grade a customer service rep. A real person reviews anything that gets flagged.
How it’s built
- Vercel’s Eve framework, which runs the agents.
- Twilio, the service that sends and receives the texts (phone calls next).
- Any AI model the ISP wants.
My prototype runs open models such as Qwen and Gemma on local hardware. Run it that way and no customer data goes to an AI company.
Built with
- Vercel Eve
- Twilio
- Qwen
- Gemma
- Ollama
- ChatGPT
- Claude
- Grok
What it costs to run
Per customer text and reply:
- Twilio: roughly 2 to 4 cents.
- A local model: pennies a day in power.
- Paid AI models from the big AI companies: 0.1 to 4 cents, depending on the model.
The same grading scorecard tells me the cheapest model that still does the job right.
The hard part
Tech support is next, and it’s a challenge. The network side can be automated: check the radio, the ONT (the box where fiber enters the home), the network path and the router before anyone asks for a reboot. Past that, it gets messy. Routers from a dozen vendors, every phone and laptop, and customers who all describe the problem differently.
Where it stands
This is a working prototype I’m still building. It isn’t running with an ISP customer yet. I’m sharing it because the same pattern fits a lot of local businesses: a few common conversations, handled quickly and within firm rules, with a person stepping in when it matters. For a local business, an after-hours answering service or a FAQ and booking chatbot does the same job on a smaller scale. A bigger job along these lines is a custom AI integration.
Good service is making things right, and that’s what I’m building toward.