LLMS (LARGE LANGUAGE MODELS) IN FLIP
A visual diagram of Flip interacting with a Large Language Model and how it’s being leveraged to identify the intention of the caller.
AI is all the rage – and as you may have guessed, we’re into it. What ChatGPT—and Alexa before it—have shown is just how good this technology has gotten. And that’s important, given all the hype cycles it’s gone through.
But it really is a pretty transformative flip. Overnight, the general public has gone from hating bots to loving them – or at least having a bot they love. But I mean, wow, what are the IVRs out there thinking when they see how popular ChatGPT is? People try every trick in the book to avoid talking to an IVR, and yet ChatGPT is the fastest consumer app of all time. Bonkers.
The thing that GPT, and LLMs more broadly, have technically solved is the Dialogue Management Challenge. In other words, based on everything the bot knows—the goal it is trying to accomplish—and what the customer actually said last in the conversation…what should the bot say next?
The last generation of bots relied on pre-scripted decision trees – you know, like your classic chatbot flow builder. And the instant someone veered off the happy path, well, they were lost forever. As we’ve all experienced, there are no such rigid tracks in ChatGPT.
So now, once and for all, may the flow builder in your customer service bots die. Instead, say hello to the era of Generative AI.
We’re officially on our way at Flip – LLMs and Generative AI now underpin our Intent Classification Model. This is the point in the call where we’re trying to identify the reason the customer is calling (so we can resolve it for them). Flip has the 200+ most common commerce intents in our classifier.
Sometimes when we ask the reason the customer is calling, we get an answer that matches up exactly to one of those 200+. But as the number of intents has grown, so too have the number of situations where what the customer says could apply to >1 intent. This is what we call Ambiguous Intent.
We used to have a dozen different ambiguous types (ie ambiguous subscription, ambiguous order, ambiguous returns, etc) that then gave a specific follow up to narrow it down.
Now, we’ve removed the rigid structure and allow our LLM model to determine which intents the customer could mean and what Flip should say in order to disambiguate and get to the right one.
A quarter of all calls are not clear with their intent and with our new LLM-driven model, over half of these people are now making it to their single correct intent. This means we can map less clear intents with a recommendation that matches more often.
We’re officially out of the starting gate. Stay tuned for much much more where this came from.
NEW RESOLUTIONS – SMS, SUBSCRIPTION INTEGRATIONS, TAGGING, TASK-ASSIGNING, & MORE
After Flip is able to identify the customer/caller’s intent, the next step is to resolve it. This can involve serving up information from their account (like latest delivery info), directing them to where they can start a workflow (like a return), or simply answering FAQs.
In the past month, we’ve put some mighty muscle into how we can resolve things.
Live currently for WISMO and Start_Return.
Instead of just sharing the latest shipping info, we now also offer to text the customer with their tracking link.
Instead of just instructing them on where to go to start a return, we offer to text them a direct URL.
2. Subscription Integrations
For those in the subscription biz, you know this is a hot topic on the phone lines.
Now, through native integrations with Ordergroove, Flip can:
- Share subscription account info,
- Pause or Cancel your subscription,
- Skip the next delivery.
So when customers call about these topics, Flip can take care of them. No waiting to speak with a person, no being sent on a wild goose chase to some hidden page in their account. Just a wicked fast and slick resolution.
3. OMS tagging, Helpdesk task creation
Shoutout to one of our new customers for this, Bartesian. When a caller wants to cancel their order, Flip can now edit the fulfillment tag in the OMS to avoid shipment.
Flip now has the ability to create a task in the helpdesk for an agent to take action on when a customer’s intent requires it. Sharing the customer info, what needs to be done, and the full call transcript if they want it too. No more work for the customer – 90% of the work done for the agent.
4. Catalog API
Special shoutout to Plow & Hearth for this one. When a caller wants to subscribe or unsubscribe, Flip can now do so directly for them via API. If you’re one of our catalog friends, we know you’ll love this.
ADDITIONAL PORTAL ANALYTICS
Now, for the pièce de résistance. The analytics in our portal contain all sorts of juicy info (some may say tooooo much*), and now we’ve added even more.
We present to you—by popular demand from Plow & Hearth, Bonafide Health, and others—The Automation Breakdown page.
Our existing analytics focus on revealing at what point in the customer journey people are calling, what they’re calling about, and how Flip is automating their inquiries.
Now we’re getting into the weeds to show the building blocks of how we’re achieving this automation.
Give it a gander, and don’t hesitate to holler if you want a walkthrough!
Until next time ✌️