According to António Pombeiro, Deputy Secretary-General of the Internal Administration, who spoke to journalists on 20 June in Porto, “if the pilot project goes well, we are prepared to start using the system to answer calls as of 2025.”
Currently, we are facing “a very recent technology”, and there is the “need to do many tests”, admitting that for now we are “very much in the unknown”, so the operation of the pilot project will be key.
“In certain situations, we have waiting periods due to the great amount of calls. This happens when there are incidents that involve a lot of publicity, a lot of people watching what is happening and everyone has the initiative to call 112”, said António Pombeiro, giving the example of urban fires.
I’m actually working on an LLM “AI” augmented call center application right now, so I’ve got a bit of experience in this.
Before anyone starts doomsaying, keep in mind that when you narrow the goal and focus of the machine learning model, it gets exponentially better at the job. Way better at the job than people.
ChatGPT on its own is a massive scope, and that flexibility means it’s going to do the bad things we know it too do. That’s why chatgpt sucks.
But build a LLM focused to managing a call center that handles just one topic. That’s what’s going on, virtually everywhere right now. This article gets that “based on chat gpt” in for clicks and fear mongering.
What should I change the title to? “Based on LLM” or “Based on Machine Learning”?
Your title is true to the article. Its the reporter who should have done it right.
No need, it’s the journalist and editor who need to clarify things.
Are these systems able to recognize when they don’t know something, or when to contact a human? All the focused learning base in the world won’t help if the machine still confabulates answers, and so far afaik that’s a flaw with all the models.
Absolutely, 100%. We aren’t just plugging in an LLM and letting it handle calls willy nilly. We’re telling it like a robot exactly what to do, and the LLM only comes into play when it’s trying to interpret the intent of the person on the phone within the given conversation they’re having.
So for instance as we develop this for our end users, we’re building out functionality in pieces. For each piece where we know we can’t do that (yet), we “escalate” the call to the real person at the call center for them to handle. As we develop more these escalations get fewer, however there are many instances that will always escalate. For instance if the user says “let me speak to a person” or something to that effect, we’ll escalate right away.
For things the LLM can actually do against that users data, those are hard coded actions we control, it didn’t come up with them. It didn’t decide to do them, we do. It isn’t skynet and it isn’t close either.
The LLM’s actual functional use is quite limited to just understanding the intent of the user’s speech, that’s all. That’s how it’s being used all over (to great results).
I’m not calling customer service unless I need a human, so the automated assistants are a huge waste of my time.
The biggest problems I have with these systems is when the company using them force you to use them, especially on the phone. They can have big accessibility barriers and it’s really frustrating when they don’t have a “let me talk to a human” function - more and more companies are using these things and not offering that and it’s a genuinely horrible experience for me, every time.
That’s great, you’re not everyone though, and you’re not fielding everyone’s calls either.
I’m in Healthcare. A massive chunk of our calls are simply “you have an order expected on (date), and shipping to (your address), is this information correct? Yes? Awesome, kthxbye”.
That’s it. By utilizing automatic dialers for that kind of thing we’re freeing up a ton of time for the real people to do more difficult hands on customer service.
I’m gonna say it, you’re the same person my great grandfather was, complaining about ATMs because they were over complicated.
Concerning support calls, if my errand can be handled by bots I can most often just get it done on the website. When i actually pick up the phone, unfailingly my issue is enough of an edge case that the automated bots cant handle it, and just end up wasting my time.
Look at the date, 2025, those issues will be very probably mostly resolved by then. Not saying that those things are not real problems now, just pointing out that we are talking about pretty damn far future in terms of ai.
Yeah, I do understand the frustration the same as with ATM when they were new if you don’t have the background knowledge then navigating menus and pressing buttons feels strange - that tech soon became ubiquitous though and for most people today ATM are basically obsolete because phone apps are far more convenient.
I think we’ll see the same happen, everyone will get used to talking to AI and systems will be refined - it won’t be long before you don’t even need to call up to get them to change your account because you’ll just tell your AI to sort it out and it’ll handle all the communication.
I’m not opposed to technology or progress. If the task I’m calling about can be easily automated then it should be available to do online, and if it’s online, I’ll do it every time on there. But companies are more and more using forcing phone calls and horrible automated systems to discourage people from doing things they don’t want them to do (requesting a refund, cancelling an account, making a complaint) in the hopes that the experience will be so discouraging that they’ll give up. It’s literally a dark pattern.
Besides which, if the problems are that straight forward and easy to automate, machine learning is entirely unnecessary for dealing with them.
I think you’re being quite naive about how developments in this area could make our lives significantly worse.