A blog post like this is half the story. I’d like to see the results. Did your brother get more business? What were the failure modes? Did customers care if it was a bot or not?
Maybe I am in the minority here, but I appreciate the new crop of LLM based phone assistants. I recently switched to mint mobile and needed to do something that wasn't possible in their app. The LLM answered the call immediately, was able to understand me in natural conversation, and solved my problem. I was off the call in less than a minute. In the past I would have been on hold for 15-20 minutes and possibly had a support agent who didn't know how to solve my problem.
Also I bet the LLM didn't speak too fast, enunciate unclearly, have a busted and crackly headset obscuring every other word it said to you, or have an accent that you struggled to understand either.
I was on the wrong end of some (presumably) LLM powered support via ebay's chatbot earlier this week and it was a completely terrible experience. But that's because ebay haven't done a very good job, not because the idea of LLM-powered support is fundamentally flawed.
Amazon support does this pretty well with their chat. The agent can pull all the relevant order details before the ticket hits a human in the loop, who appears to just be a sanity check to approve a refund or whatever. Real value there.
My big question is. Why has the company and their development process failed so horribly they need to use LLM instead the app? Surely app could implement everything LLM can too.
I had a similar situation with a chatbot: I posted a highly technical question, got a very fast reply with mostly correct data. Asked a follow-up question, got a precise reply. Asked to clarify something, got a human-written message (all lowercase, very short, so easy to distinguish from the previous LLM answers).
Unfortunately, the human behind it was not technically-savvy enough to clarify a point, so I had to either accept the LLM response, or quit trying. But at least it saved me the time from trying to explain to a level 1 support person that I knew exactly what I was asking about.
Agreed; they're far better than the old style robots, which is what you'd have to deal with otherwise.
More generally, when done well, RAG is really great. I was recently trying out a new bookkeeping software (manager.io), and really appreciated the chatbot they've added to their website. Basically, instead of digging through the documentation and forums to try to find answers to questions, I can just ask. It's great.
i genuinely don't get the point of this. isn't it easier to have a native chat interface? phone is a much worse UX and we simply use it because of the assumption that a human is behind it. once that assumption doesn't hold - phone based help has no place here.
I had a recent experience with the Lowes agent today. It was pretty decent! Until I asked "how many of that item is available", and it didn't know how to answer that (It was a clearance item). At least when I asked to talk to a human I got one in a few seconds.
> This isn’t a generic chatbot. It’s a custom-built voice agent that answers his phone, knows his exact prices, his hours, his policies, and can collect a callback when it doesn’t know something.
This is bad engineering. Step 1 should not be "start building the thing". The first thing you do (after understanding the customer's problem) is look for existing solutions already available and evaluate them.
I just called, chat bot is not even used yet. This is the worst tech demo ive seen on HN. New coders will be replaced by AI because they rather write articles and be streamers.
> This influencer was casually talking about how they learned to code and were now making a bajillion dollars. Two thoughts hit me instantly: 1.What is this "code" they’re talking about? and 2. How do I earn a bajillion-dollars?
I wish this person all the best, but most of this is incredibly out of alignment with how I wish programming / engineering culture would evolve.
"hundreds of calls per week" because the mechanic is under the hood every day seems a little off. Hundreds seems way too excessive. Perhaps an exaggeration to highlight the problem's seriousness?
With that kind of volume, I think even before AI could have helped, why not hire some staff and potentially even a receptionist. Given the volume, this seems like an easy choice.
I used to work as a service advisor - or as the article says, receptionist. This system will not work as described for several reasons.
1. Unless you have a recent job that matches the exact same repair/service, you have incorrectly estimated the cost of the repair. In some states, this matters a lot and will cost the shop money. Unless your LLM only quotes for labor in sane amounts for diagnostic and nothing else, you’re only adding noise. This is a disservice to the client and the shop owner. The client now has an inaccurate quote for work and the shop will get a reputation for being inaccurate in quoting work.
2. Let’s say that you manage to get the exact same job twice. Your machine now needs to source parts. Parts may have been in stock yesterday. The might be out of stock now. If they are in stock, you need to retotal the price since prices are dynamic. Did you teach the agent how to source parts? What rules does it have for sourcing used parts?
3. New jobs can’t be quoted. Even if you taught the machine how to calculate book time and margins, it still has to find the right parts. If your shop does high end work, you know how much of a pain in the ass this is. Also remember that some work requires nonobvious parts - like fluids if you need to remove a part in the way of your goal.
4. The only area I see this being useful in is pickup. The shop can mark a car as done and the LLM can call to inform the client that they can come at a preset, unchanging time to get the vehicle. If the vehicle is staying overnight, the LLM can call with a progress update.
Finally, I’d like to note that this sort of dev work goes beyond hubris. It’s dangerous. The more we assume we know without verifying, the greater the risk. In this case, the dev is risking someone else’s livelihood.
>and he’s losing thousands of dollars per month because he misses hundreds of calls per week. He’s under the hood all day. The phone rings, he can’t answer, the customer hangs up and calls someone else. That’s a lost job — sometimes a $450 brake service, sometimes a $2,000 engine repair — just gone because no one picked up.
How much does it cost to have an outsourced receptionist? Even if it is 500 a month if we are really talking about thousands of dollars per month lost your ROI is still crazy.
At the moment I'm pretty inclined to hang up if I feel I'm wasting my time with a robot.
But maybe soon we will not even realise we speak to a robot, given the current speed of ai development.
I wonder how that will erode trust in calls. I moved from cold emailing and cold LinkedIn to cold calling because of the massive amounts of ai spam I have to compete with. But maybe cold calling will die soon as well if the robots emerge.
Is RAG even necessary here? Minimal information like a couple of price list with job times and opening hours should easily fit into any context window, right? It's not like he's dumping entire service manuals into the vector database here...
Our local air conditioning/heating/plumbing place started routing their frontline answering service to an AI call service. It has an odd uncanny valley feeling to it that I simply don't trust. I chased down other points of contact until I got ahold of a person instead. I'd trust a phone tree with a voicemail box more.
I'm not part of the target market but I hate talking about technical and financial things on the telephone even to a human being. I much prefer being able to put my request in writing and get a written response plus a phone call from someone who has already been able to evaluate my request and pick holes in it and advise me what to do next.
That way there are fewer misunderstandings. This is how arranging servicing with Tesla works (at least in Norway and the UK). I use the app to describe the problem, add photographs if i think they would help. Then if it clear wat is to be done the app will notify me that the estimate is ready for approval, otherwise I'll get a message in the app or a phone call or both.
Reading this thread has me wondering if a smarter voicemail inbox system would be a viable and useful alternative.
Something like leave a voicemail and get a ticket created in the system to log the call for triage once the owner's hands are free.
The system could be configured to text a link to an intake form or quote calculator or calendly or whatever if intent is successfully detected.
I wouldn't mind this as a caller because I'd hear a typical phone tree setup, have a callback option, __and__ have a digital action I can take without being annoyed by a fake receptionist.
I imagine a certain type of business owner would appreciate the more streamlined workflow, plus having more control without the AI risks.
In the UK all of my local garages miss calls. If you go down there it is easy to see why, they are booked solid and have people turning up out-of- the-blue with emergency work. There is no use for an agent. If you want to get your car fixed you have to work around them.
In the article the person claims their brother misses a lot of calls because they are too busy to take them. If they are too busy to take calls, how are they going to fit more work in?
Then there is the luxury element, luxury services answer with humans. The end.
> Wired up Claude for response generation — The retrieved documents get passed as context to Anthropic Claude (claude-sonnet-4-6) along with a strict system prompt: answer only from the knowledge base, keep responses short and conversational, and if you don’t know — say so and offer to take a message. No hallucinations allowed.
Claude will hallucinate anyway, sometimes.
I don't think there's any way around this other than a cli or MCP that says "press the 'play prerecorded .WAV file button that says the brake repair service info and prices.'"
No idea what luxury is doing here, but if I get an LLM receptionist, that ain't it.
This isn't to disparage the project - I think this sort of usage will become very common and a decent standard that produces good consumer surplus in terms of reduced costs etc. Especially impressive is that it's a DIY family-first implementation that seems to be working. It's great hacker work.
But be warned it will erode - in general - the luxury previously associated with your brand, and also turn some customers away entirely.
If you didn't have a sibling to do it for you free/cheap, I wonder how many months of a human receptionist (or service) the fee to build (and maintain) such a thing would cover.
> When a caller asks something that isn’t in the knowledge base, the AI doesn’t guess.
I've seen a number of instances of this type of thing in wild, and under the hood it's usually some prompt that asks the llm to gauge confidence in its own answer. And from my admittedly naive understanding of how this all works, this seems extremely unlikely to be accurate. But I'm curious if that's still the case or if I'm operating under pre-2026 information?
Fair warning to those out there: I've had terrible experiences with AI receptionists so far, to the point that I refuse to do business with anyone who uses them.
I went through hell on a home remodel project 6 months ago around this stuff. I got a quote from a reputable plumber and went to schedule the rough-in session. An AI receptionist answered, got confused during the scheduling flow and could not understand my address, asking me to repeat it over and over. And it couldn't forward to me to human.
If I'm paying you tens of thousands of dollars for remodeling work, I damn well better be able to get in touch with you. I found a different contractor and never looked back.
I build software for contractors (plumbers, electricians, HVAC repair, etc) and they're some of the fastest adopters of these systems. I believe YC has even invested in a few.
Regarding the AI receptionists, from the calls I've listened to, there's still a bit of the uncanny valley/overlapping speech issues that I'm unsure are ever fixable just due to latency.
But for low margin businesses like contracting and (I imagine) auto repair where labor is your most expensive cost, these owners are doing anything they can to reduce their overhead.
I think the idea and write up is neat, but I question if the brother ever thought of, I don’t know, hiring a receptionist so he wasn’t losing “thousands of dollars per month”? Maybe I’m oversimplifying the problem.
The other aspect of this I have questions about is the price quoting. Knowing the standard price list makes sense, but how can you trust that a caller knows what they need without looking at their car? Wouldn’t this likely result in quotes for work that may or may not even be needed?
Our local Subaru dealership has an option to use an AI assistant when you call to set up an appointment. I tried it, it worked perfectly, better than a human, honestly. Likewise our local Taco Bell is uses AI ordering and it works great as well. I'm sure they have their issues, but I'm losing nothing by not talking to a human in this situation, and there's always a human available if I need one.
Honestly great work, but this is very much something where the results matter more than the product. It ends without a single comment about whether it worked in Production.
> "and he’s losing thousands of dollars per month because he misses hundreds of calls per week. He’s under the hood all day. The phone rings, he can’t answer, the customer hangs up and calls someone else"
I'm not familiar with american pay amounts, what's the wages for a receptionist like? If you just hire one to make sure you don't lose thousands of dollars a month in jobs, will you still be making a profit?
My mechanic has had an earpiece thing. After three rings it rings in his ear. Few customers realize the voice booking thier appointment is often elbow-deep in an engine.
He told me once that he would never have a dedicated receptionist. He has his mechanics rotate between the shop and the desk. They have the skills to give advice and prices over the phone, sometims ordering parts before the customer ever arrives.
I built something similar and learned a lot about how useful a “receptionist” (advisor) is vs something that just immediately sends the user to the shop asap
The foundational product truth is you win business by saying “oh you’re car is broken, come in right now” then an actual tech or manager does the sale
They admit a raw LLM would be dangerous and then proceed to use RAG... How is this any better? You cannot allow an LLM to generate the final outbound message if you are liable for what it says.
LLM to understand the question? Yes. Generate SQL maybe with Embeddings to look up answers? Yes. Generate the final response? No.
Think about scaling this as you're building, your brother is just your first customer, make sure your service works with any number of customers out of the gate. I should be able to sign up for your service, point it at my website to ingest all my information, and have it ready to go.
It’s a luxury to hire a human… but needed for these type of high value customers… just have the receptionist do other tasks a human can do while between calls… a clean shop is a functional shop ;)
This felt oddly ai-authored, like the kind of “brain dump then ask ChatGPT to sort it into a post”. Probably because of too many bullet points and bold emphasis
Thanks for sharing the journey. What did you do in terms of security for the receptionist? I suspect someone can trick the agent through things like prompt injection.
Ignore the expected negativity, many here have not used the latest gen of voice agents in development. Even if used as a router , prefer that to waiting to get through
A bunch of comments around the theme of "Is it worth it for _this_ use case?" misses the point of "Which of these techniques/lessons from the blog can I use for _my_ use case". I found the blog post a good starting point for some agentic stuff I want to do for my use case. Good on you for experimenting and good luck!
319 comments
I was on the wrong end of some (presumably) LLM powered support via ebay's chatbot earlier this week and it was a completely terrible experience. But that's because ebay haven't done a very good job, not because the idea of LLM-powered support is fundamentally flawed.
When implemented well it can work great.
Unfortunately, the human behind it was not technically-savvy enough to clarify a point, so I had to either accept the LLM response, or quit trying. But at least it saved me the time from trying to explain to a level 1 support person that I knew exactly what I was asking about.
More generally, when done well, RAG is really great. I was recently trying out a new bookkeeping software (manager.io), and really appreciated the chatbot they've added to their website. Basically, instead of digging through the documentation and forums to try to find answers to questions, I can just ask. It's great.
If my mechanic answered with an LLM I’d take my car elsewhere.
> This isn’t a generic chatbot. It’s a custom-built voice agent that answers his phone, knows his exact prices, his hours, his policies, and can collect a callback when it doesn’t know something.
This is 2026's most generic chatbot.
> This influencer was casually talking about how they learned to code and were now making a bajillion dollars. Two thoughts hit me instantly: 1.What is this "code" they’re talking about? and 2. How do I earn a bajillion-dollars?
I wish this person all the best, but most of this is incredibly out of alignment with how I wish programming / engineering culture would evolve.
With that kind of volume, I think even before AI could have helped, why not hire some staff and potentially even a receptionist. Given the volume, this seems like an easy choice.
1. Unless you have a recent job that matches the exact same repair/service, you have incorrectly estimated the cost of the repair. In some states, this matters a lot and will cost the shop money. Unless your LLM only quotes for labor in sane amounts for diagnostic and nothing else, you’re only adding noise. This is a disservice to the client and the shop owner. The client now has an inaccurate quote for work and the shop will get a reputation for being inaccurate in quoting work.
2. Let’s say that you manage to get the exact same job twice. Your machine now needs to source parts. Parts may have been in stock yesterday. The might be out of stock now. If they are in stock, you need to retotal the price since prices are dynamic. Did you teach the agent how to source parts? What rules does it have for sourcing used parts?
3. New jobs can’t be quoted. Even if you taught the machine how to calculate book time and margins, it still has to find the right parts. If your shop does high end work, you know how much of a pain in the ass this is. Also remember that some work requires nonobvious parts - like fluids if you need to remove a part in the way of your goal.
4. The only area I see this being useful in is pickup. The shop can mark a car as done and the LLM can call to inform the client that they can come at a preset, unchanging time to get the vehicle. If the vehicle is staying overnight, the LLM can call with a progress update.
Finally, I’d like to note that this sort of dev work goes beyond hubris. It’s dangerous. The more we assume we know without verifying, the greater the risk. In this case, the dev is risking someone else’s livelihood.
>and he’s losing thousands of dollars per month because he misses hundreds of calls per week. He’s under the hood all day. The phone rings, he can’t answer, the customer hangs up and calls someone else. That’s a lost job — sometimes a $450 brake service, sometimes a $2,000 engine repair — just gone because no one picked up.
How much does it cost to have an outsourced receptionist? Even if it is 500 a month if we are really talking about thousands of dollars per month lost your ROI is still crazy.
But maybe soon we will not even realise we speak to a robot, given the current speed of ai development.
I wonder how that will erode trust in calls. I moved from cold emailing and cold LinkedIn to cold calling because of the massive amounts of ai spam I have to compete with. But maybe cold calling will die soon as well if the robots emerge.
I appreciated your post and have some takeaways around text formatting for TTS in my own projects. Thanks!
That way there are fewer misunderstandings. This is how arranging servicing with Tesla works (at least in Norway and the UK). I use the app to describe the problem, add photographs if i think they would help. Then if it clear wat is to be done the app will notify me that the estimate is ready for approval, otherwise I'll get a message in the app or a phone call or both.
Something like leave a voicemail and get a ticket created in the system to log the call for triage once the owner's hands are free.
The system could be configured to text a link to an intake form or quote calculator or calendly or whatever if intent is successfully detected.
I wouldn't mind this as a caller because I'd hear a typical phone tree setup, have a callback option, __and__ have a digital action I can take without being annoyed by a fake receptionist.
I imagine a certain type of business owner would appreciate the more streamlined workflow, plus having more control without the AI risks.
In the article the person claims their brother misses a lot of calls because they are too busy to take them. If they are too busy to take calls, how are they going to fit more work in?
Then there is the luxury element, luxury services answer with humans. The end.
> Wired up Claude for response generation — The retrieved documents get passed as context to Anthropic Claude (claude-sonnet-4-6) along with a strict system prompt: answer only from the knowledge base, keep responses short and conversational, and if you don’t know — say so and offer to take a message. No hallucinations allowed.
Claude will hallucinate anyway, sometimes.
I don't think there's any way around this other than a cli or MCP that says "press the 'play prerecorded .WAV file button that says the brake repair service info and prices.'"
luxuryis doing here, but if I get an LLM receptionist, that ain't it.This isn't to disparage the project - I think this sort of usage will become very common and a decent standard that produces good consumer surplus in terms of reduced costs etc. Especially impressive is that it's a DIY family-first implementation that seems to be working. It's great hacker work.
But be warned it will erode - in general - the luxury previously associated with your brand, and also turn some customers away entirely.
> When a caller asks something that isn’t in the knowledge base, the AI doesn’t guess.
I've seen a number of instances of this type of thing in wild, and under the hood it's usually some prompt that asks the llm to gauge confidence in its own answer. And from my admittedly naive understanding of how this all works, this seems extremely unlikely to be accurate. But I'm curious if that's still the case or if I'm operating under pre-2026 information?
I went through hell on a home remodel project 6 months ago around this stuff. I got a quote from a reputable plumber and went to schedule the rough-in session. An AI receptionist answered, got confused during the scheduling flow and could not understand my address, asking me to repeat it over and over. And it couldn't forward to me to human.
If I'm paying you tens of thousands of dollars for remodeling work, I damn well better be able to get in touch with you. I found a different contractor and never looked back.
Here’s the video: https://youtu.be/QmH9b27xm6k
It was very impressive at that time. They did raise money after that pitch, but they ended up pivoting (multiple times). They IPOd in 2017
Regarding the AI receptionists, from the calls I've listened to, there's still a bit of the uncanny valley/overlapping speech issues that I'm unsure are ever fixable just due to latency.
But for low margin businesses like contracting and (I imagine) auto repair where labor is your most expensive cost, these owners are doing anything they can to reduce their overhead.
The other aspect of this I have questions about is the price quoting. Knowing the standard price list makes sense, but how can you trust that a caller knows what they need without looking at their car? Wouldn’t this likely result in quotes for work that may or may not even be needed?
> "and he’s losing thousands of dollars per month because he misses hundreds of calls per week. He’s under the hood all day. The phone rings, he can’t answer, the customer hangs up and calls someone else"
I'm not familiar with american pay amounts, what's the wages for a receptionist like? If you just hire one to make sure you don't lose thousands of dollars a month in jobs, will you still be making a profit?
He told me once that he would never have a dedicated receptionist. He has his mechanics rotate between the shop and the desk. They have the skills to give advice and prices over the phone, sometims ordering parts before the customer ever arrives.
The foundational product truth is you win business by saying “oh you’re car is broken, come in right now” then an actual tech or manager does the sale
Worth checking out at: https://wrenchdesk.ai
* i'd love to hear a sample/customer call. Even if it's just a test
* a blog without rss? How can i subscribe for part 2?
LLM to understand the question? Yes. Generate SQL maybe with Embeddings to look up answers? Yes. Generate the final response? No.
However, does the regular "joe/jane" feel the same way? I imagine my mom or dad would most likely not notice or care if they did.