How do we stick it to the rest of the industry and actually use LLMs to responsibly help with customer support?
Most people these days would probably say "You can't." I get the instinct. Everyone is using AI wrong. EVERYONE. I cursed endlessly at the Disney Plus AI just the other day, worthless junk. Everyone is trying to use AI to replace humans on support, and force customers to talk to an AI. AI can't handle customer support, it's simply too complex.
For me though, I've been waiting for AI my whole life. This isn't something I'm trying to use to save money. I'm not trying to weight the loss of revenue vs the savings from firing the support team. So while everyone else is misusing AI to completely bomb customer support, I have a question to answer: How do I make everyone else look bad and use it the right way?
I actually have an answer. Most of you that received a support ticket response in the last two weeks unknowingly saw the answer. It's actually pretty simple. That's why it makes everyone else look stupid, and I like that. I needed a confidence boost.
The answer is that it doesn't replace support at all. It's a tool that helps support run faster, smoother, and with more accuracy. Here's how it works:
Our obsessively secured, completely private and locally hosted LLM, read every support ticket that we've answered this year. It can access logs and basic account information by internal API, and no it has absolutely no way of exfiltrating that data. Nor does it have a way of accessing the functions that limit it's access. Simply, it guesses what we were most likely to say in reply to a support ticket, and it writes that in a draft. The human walks in, checks the ticket and draft, and makes a decision. Is this what I would have sent? If yes, send it. If no, reject it (and send feedback to the LLM). If almost, edit and then send. If the reply requires action from us, it determines what we were most likely to say and leaves us a note with the required actions to send the draft (if we agree, of course).
The customer never speaks to an AI. The AI simply helps the human reply faster. Less browser tabs, less log greps, less "I forgot how that feature works, let me check before I reply."
Just because everyone else is using it wrong doesn't mean we're not going to use it. We're just going to use it as a middle finger to all of them.
Our support ticket backlog is down to the smallest levels in years. We're almost completely caught up. A human has been involved in every single ticket. And best of all: We didn't have to increase prices to do it.