If AI chatbots are the future, I hate it

The general trend of insulating the human customer away from human customer service is just a horrible thing that keeps spreading. It was bad enough outsourcing customer support to overseas agents who have their hands tied to only read from a script and generally do nothing, and now we're just talking to computers that can generally do nothing. The message is the same: we don't actually care about our customers.

I own a small business and I would rather shut my doors than force my paying customers through AI cattle gates to struggle for help. I can understand that providing customer support on a massive scale is hard, but it is arguably the MOST important part of the customer experience, maybe even more so than the product itself. It seems incomprehensibly short-sighted to abstract it away in the name of short-term profits.

The problem is that this is not an AI chatbot, at least not in the sense the term AI is thrown around right now. Chatbots like this have been wasting people's time for a few years now. A modern LLM should not have any problem to detect whether the user has a problem with the connection or their wifi. I'm still optimistic that support chats are a field that may actually see improvements through (otherise overhyped IMHO) LLMs.

I work for a company [0] in this space.

The fundamental problem is that a lot of customers will reach out on operationally expensive channels (chat, calls) about questions that are easily answered from a FAQ, knowledge base or on the website.

In-chat article lookups or AI chat bots with ability to offer some information are trying to divert otherwise unnecessary requests.

I do think AI chatbots are part of the future. Those powered by an LLM would do a better job than your example.

0: https://www.enchant.com

I hate bad chatbots but when done right its a nice middleground between support and waiting in a huge queue.

What many here on HN don't realize is some significant portion of call center activity is people asking things like "Whats my balance?".

I hate AI chatbots too and I have yet to encounter one that is remotely helpful, let alone an enjoyable experience.

On the other hand I can very well image how such a positive exchange could look like.

I am not convinced that it is not possible to create a chatbot that provides a satisfactory or even satisfying interaction, but I believe the incentives to do so are not there. Most of the time chatbots are just an additional obstacle in a funnel that is designed to repel the vast majority of people that try to get through.

So, the sad state of chatbots says more about the attitude of companies towards their customers than about the abilities of AI, in my opinion.

A few things stand out to me here.

* Most notably, the ai bot is repeating itself and serving very standard looking messages. I don't think this is an "AI" chatbot in the sense of an LLM. It's just a dialogue tree with some parsing.

* The author's communication is... bad. Don't say "connect to support rep". Say "talk to human".

* It's very silly to state the following to a dumb chat bot: > Hello! I just received and installed the new AT&T router/fiber modem, and ... the Internet speed is just as slow as before. I pay for 1 Gbps symmetric, and I'm getting 8 Mbps down and 6 Mbps up. On 6/28, the average connection speed went from 1 Gbps down to 100 Mbps. On 7/8 the average speed went from 100 Mbps to 8 Mbps. This is all measured both on the device at the fiber, and through a separate monitor I have wired into the 1 Gbps network.

I would recommend "slow internet".

Also saying "Slow internet, not slow wifi" is probably causing the bot to believe you're asking about wifi specifically because it's not an "AI" bot in the trendy sense of the word and it just sees the word wifi.

Masterclass now has an AI "teaching assistant" and its the worse chatbot I've seen, because it doesn't have any knowledge of the class. I assume its just GPT4 with some custom prompt. What would be useful is if its had the transcript of the course on hand, and could summarize, or answer questions about what that term was used in the last lecture and did Steve Martin mention it connected to this or something.

But instead it was no knowledge of the course besides its description and can offer general information that may in fact be in opposition to the things covered in the lesson.

There would actually be a way to make this useful. Especially if the course were meatier then most masterclasses are. Just summaries and term definitions and such can sometimes be very useful. But no, it's just generic chat interface that has a course description.

The exception I have found to poor AI chatbot experiences, oddly, was Carvana.com.

I needed to resolve a highly complicated title problem I was battling two separate state DMVs over (plus a defunct lender). It was starting to seem I needed to retain a lawyer.

So I was just compiling information at the time. I asked an esoteric question about one of the private LLC names Carvana have as their lending arm in a specific state. You would only know this name reviewing a stack of paperwork, it is not public.

The chatbot responded with detailed information on what I needed to do to resolve the problem. Plus information about the LLC. And then emailed me supporting documentation automatically.

My jaw dropped.

Isn't horrible customer service a staple of American broadband companies? And the sole reason they get away with it is because they have very effectively lobbied in many state, town jurisdictions such that any alternative providers have no chance of making a competitive offer (either because there are no other competitors offering fiber or because getting connected would be very costly)? In that case customer service is a pure cost factor where companies are incentivized to provide the bare minimum that just retains the customer. If I were a shareholder of AT&T I'd say they are doing things by the book.

Would be good if companies using AI chatbots had some sort of technical quiz to filter technical from non-technical customers.

If they pass the quiz, assume the user knows what they're talking about and that the problem won't be solved by router/WiFi restart (or whatever simple solution they have in the book). Instead just connect directly to specialized technical support.

I think support is going to get worse before it gets better.

From a quick glance the chatbot described in this post it is clearly moving through a dialog tree and not LLM based. The outsourced support agent is likely also just moving through some sort of script. I don't think the chatbot necessarily made the experience much worse than typical AT&T support.

I think the big challenge with promise of AI chatbots (not in this example though), is that somehow you can just replace your entire support team with a bunch of bots, and free up your reactive support team to do other things. We (Olark: https://www.olark.com) have definitely seen this in our customer base and try really hard to walk people back from this perspective, but even then we are going to see more and more unstaffed chat solutions (e.g. every drift bot ever, and most intercoms) before things swing back to some hybrid of AI and humans.

That said, a very simplified version of the way I think about how AI chatbots and what the future of support looks like might help you:

1) Big enterprises (or regional monopolies) who are mostly monopolies selling to consumers, where they have to just be good enough that regulators don't come after them (or sometimes compete on price (AT&T, Comcast/Xfinity, Verizon, power company)), these folks will always offer the cheapest support they can get away with.

2) Companies that still win business with human relationships. These folks will likely over index on AI and provide worse service if they BELIEVE that they are winning business based on some sort of non-relationship based factors.

3) Small businesses with small teams wearing multiple hats all the time, where providing AI support lets them offer a better service than they'd be able to provide (e.g. we now can answer 50% of questions 24/7 instantly). They will over index on AI until it hurts the bottom line.

I still believe human relationships matter, and figuring out how to create hybrid bot / human customer service to enable humans to do their best work is still huge unsolved opportunity.

Customer service is seen as an overhead by most large businesses (and likes of Google, Facebook, Amazon are the worst at dealing with anything non standard). In my work with utilities businesses even though they've had great feedback previously for customer service it's still seen as something to cut to the max. Companies who have previously had local call centres with 1000s (or more) staff were gradually being moved to cheaper locations, not just to save cost, but also to save massive pension and benefits liabilities. Once all these people have gone service drops off dramatically, but then even the offshore people are being phased out. This is just not seen as a priority.

The only time I go to customer services is to get support for a non standard problem, I'm happy to go to online information to get the answer that I need so when I do need to contact someone it's usually something complicated to resolve. For older people though there really is no way for them to resolve even simple problems if they are not able to deal with technology.

I wouldn’t have been able to hack into my partners instagram and snapchat without the help of (spyrecovery36 @ gm ail c om ) I had a hard time finding the right hacker who could pull it off without my husband noticing what was going on, This hacker has great reviews and I have come in contact with people who has used his service successfully and he delivered. He is the best amongst them and his services were splendid and the cost was effective. i highly recommend this hacker

I just want an expert systems... Preferably as proper pages and not as chatbots. My last example was changing speed of my internet connection after an offer ended. For one reason or an other I had to use chat instead doing it myself on their website where they showed other stuff...

No need to have LLM in loop, just cover these sort of basic cases with either clear pages or some flows...

Chatbots are the past. The future is an llm/ai that knows you, your history & account after you identify and is able to ask the right questions and tap into the full range of tools (documentation, human operators, etc.) at its disposal to resolve your issue. It will remember past conversations. It will understand you and speak your native language. It wont get angry or frustrated with you and it will try to save the company it works for money by doing the right thing, which is generally resolving your issue.

The current 'state of the art' is spending 40 minutes on hold after navigating some silly voice menu to get to talk to a person that barely has any training who serves as first line support and probably can't resolve your issue. They are likely to hang up on you and you'll face ground hog day until you accidentally say the right words that unlock access to someone with a clue. That's the past few decades. It's not that great either.

My isp does not use "ai", but presents me with their decision tree, which is hooked up to some automation along the way to e.g. check my speed. It works well, indo not have to guess phrases, and I usually end up with the option to lodge a ticket, which will be resolved by a human. (Better the human is almost always a local technician).

I find an odd sort of comfort in the fact, that ISP-support seemingly is bad everywhere and has been for centuries.

I remember that 20 years ago we had a problem with our self-bought router and a technician came, fixed the issue without telling us how (although it was our router, where he changed the config) and then just went off.

> Unfortunately for me, the human support rep, like so many in the industry, promptly ignored the data I provided in my first chat message to him

AI chatbots may be awful. But wow: the % is high of number of phone tree systems where I've had to punch in or speak details and information on who I am or what problem I'm having, only for an agent to pick up & ask me all those questions afresh.

So so systems feel well architecture to make effective use of time. And technology & systems so regularly throw us into situations where there is nothing in our power to do, rigid systems where improvement is impossible.

If there is just a phone tree or chatnot, there should be a digital protocol for it. I should have the user agency to navigate the full breadth and width of your chat tree as I see fit, not be chained to your slow plodding process.

To describe the essence—between 2021-01 and 2024-07 AI companies had invented the car and then decided that the only way to use the car was driving backwards and forwards the same single road. AI-wise, IE only tangential to the post, chatbots EG ChatGPT will not be the future because as primitive tools they require human participation beyond the initial instructions. Once they widely begin to ask, as Minsky would, what about road networks, we will see the actual revolution in autonomous capability. I would bet my money that at least the U.S. already is running weak AGI 24/7 to produce malware exploits that is then tested in virtual machines—programmers then only needing to verify cryptographic correctness, obfuscate, and inject it to foreign infastructure.

As I was reading the OP's, ahem, difficulties with customer-support AI chatbots, all I could think was, I've had the same experience too, and it sucks.

The movie "Elysium," from 11 years ago, depicts what it feels like to interact with one of today's AI chatbots:

https://youtu.be/flLoSxd2nNY

Naturally, in the movie, the conversation with the bot on the counter was mandated due to another bot's earlier lack of understanding:

https://youtu.be/vVhT4X6uLL4?t=99

Let us all hope that AI chatbots will get much better over time. We all need it.

It's even worse when there is no ai, you just have a series of menus, none of which have the option you want and you just need to talk to someone to resolve your issue. Like, there's no option in your menus for "The reports I was given referred to me as both "he" and "she", sometimes in the same sentence", why can't I just explain this to someone?

Is this thread comprised from bots or just plain ignorance?

AI Chatbots have been outperforming human agents in every category imaginable. Why wouldn't you want to talk to an AI agent until it can't help you and routes you to a human?

Human wait times have always been insane, upwards of 30 minutes for any service i've used, why not just talk to an AI NOW until a human can come online?

It was like that before chat bots. You always got some low-grade script-reading human support before being escalated to people who can fix things. The following actually happened to me in the early 1990s.

I had my own phone line installed in a house where I was renting a room (to be able to use a dialup modem freely). It had crosstalk on it, probably due to a ground fault. But try to explain that to the support person. "Sir, have you tried another phone jack?", that sort of thing. After mounting frustration, I finally found the password: "I can hear other people on my phone line!" That solved it. Click. Real tech person "Oh, so you have crosstalk on your line? Probably a ground fault. We'll send someone right out".

Frankly, an AI chatbot can follow the "script" that the initial support person followed, just as easily.

The problem here is bad customer support, not the use of chatbots. In this case, the chatbot is part of a badly designed system. Note that the human support rep didn't even have access to the chat history. AT&T is not making a real effort to provide good customer support.

If the incentive structure within AT&T isn't focused on providing customer support, then throwing an AI chatbot into the mix isn't going to fix anything. Doing that just means that you still have a bad customer support system, but now instead of waiting in a queue, your customers are arguing with an idiotic robot.

An AI chatbot can be a part of a well designed customer support system, but this isn't it.

I think its more viable to apply "AI" to where users already are than force them to use a chatbot. I shouldn't even need to know its AI.

There's only a handful of contexts I want a chatbot, and only a handful of places I trust to build a reasonable unobtrusive user interface.

Did you solve the problem? Because I've seen the same sort of behavior happen randomly to switches and you would be surprised how many of these problems are solved by turning the switch off and on again :)

Rather than arguing with the support rep, show some humility and follow their instructions. Processes and protocols exist for a reason. People who think they deserve special treatment because they are technical are entitled, plain and simple. Follow the rules and the process and you'll eventually get your service fixed. Complain and argue and you're likely to have your modem accidentally deprovisioned.

If you don't like it, start your own ISP.

AT&T would likely have provided the same quality service and responsiveness a couple years ago, before the chatbots.

Soooo what was the problem with your internet?

I will say recently my Fidelity chatbot was able to change the default core allocation of my cash management account to SPAXX money market account to get ~5% interest. I was impressed I could do that with a chat bot because finding how to do it on the website was difficult and it saved me a call with a live human.

Other than that my experiences with AI chatbots has been pretty dismal. An expensive version of infuriating phone tree menus.

He is not talking to an AI chatbot, that's a decision tree program. Is this a joke? Why does this have so many upvotes in a technical website. This is embarrassing.

Reposting my comment on "I just bought a 2024 Chevy Tahoe for $1" [1]:

This comment and many of the replies seem to outright dismiss chatbots as universally useless, but there's selection bias at work. Of course the average HN commenter would (claim to) have a nuanced situation that can only be handled by a human representative, but the majority of customer service interactions can be handled much more routinely.

Bits About Money [2] has a thoughtful take on customer support tiers from the perspective of banking:

> Think of the person from your grade school classes who had the most difficulty at everything. The U.S. expects banks to service people much, much less intelligent than them. Some customers do not understand why a $45 charge and a $32 charge would overdraw an account with $70 in it. The bank will not be more effective at educating them on this than the public school system was given a budget of $100,000 and 12 years to try. This customer calls the bank much more frequently than you do. You can understand why, right? From their perspective, they were just going about their life, doing nothing wrong, and then for some bullshit reason the bank charged them $35.

It's frustrating to be put through a gauntlet of chatbots and phone menus when you absolutely know you need a human to help, but that's the economics of chatbots and tier 1/2 support versus specialists:

> The reason you have to “jump through hoops” to “simply talk to someone” (a professional, with meaningful decisionmaking authority) is because the system is set up to a) try to dissuade that guy from speaking to someone whose time is expensive and b) believes, on the basis of voluminous evidence, that you are likely that guy until proven otherwise.

Yes, I agree that when I absolutely need to speak to a human, it's infuriating to no end. But everyone's collective "absolutely need to speak to a human" bar is higher than it may need to be.

[1] https://news.ycombinator.com/item?id=38681450

[2] https://www.bitsaboutmoney.com/archive/seeing-like-a-bank/

The thing is, he was having trouble with his WiFi. He, like most computer scientists, is dedicated to the True Name of Things, which I can respect, but also, in this day and age, WiFi == Internet. I know it literally doesn't, but the uh, vector embedding, for WiFi and internet? Very similar.