FreeUpToHours
25,361 Messages for $6.85: The AI Machine That Keeps 15 Rooms Full Every Day
Automation·By Oliver Valencia Sebastian·Published August 6, 2026·10 min read

25,361 Messages for $6.85: The AI Machine That Keeps 15 Rooms Full Every Day

Last month my AI bill was $6.85.

That is not a trial credit and it is not a first-month promo. For $6.85, DeepSeek v4 Flash answered 25,361 inquiries in my Facebook inbox and burned through 280,109,038 tokens. It did not sleep, it did not get annoyed at the fortieth "magkano po" of the day, and it did not once quote the wrong rate.

In the same stretch, one of my reels crossed 730,505 views with 577 shares and pulled in 2,724 new followers. My Google Analytics went from a flat line to 3,944 monthly active users, up 2,103 percent.

None of those are the number that matters. This is the number that matters: 15 rooms, full every single day, since April.

The $6.85 that answered 25,361 messages

DeepSeek platform dashboard showing $6.85 total cost, 25,361 API requests, and 280,109,038 tokens on deepseek-v4-flash
One month of my Facebook inbox: 25,361 requests, 280 million tokens, $6.85 total.

People assume the number is fake, or that I am hiding a second bill somewhere. I am not. That screenshot is the whole inbox cost for a month.

The reason it works is boring. At 25,000 messages a month, choosing a model stops being a taste preference and becomes arithmetic. Run that volume through a premium model and it will bankrupt you. Run it through the cheapest model that still clears your quality bar and you get the same booking with a rounding error attached to it. DeepSeek v4 Flash is, right now, the cheapest model I have found that still answers a Filipino guest properly in English, Tagalog, and Taglish without embarrassing me.

I am not loyal to DeepSeek. I am loyal to the math. The day something cheaper clears the same bar, I will switch, the same way I switched off an expensive model when my volume climbed and the bill started to hurt.

This is also why I do not run one model for everything. DeepSeek sits in the inbox because the inbox is a volume problem. Claude writes my content, my SEO, and my landing pages because that is a judgment problem and I trust its output enough to publish it. Gemini makes my videos because, for now, Gemini makes the best ones. Three jobs, three tools, and no romance about any of them.

Why my bot never hands the conversation back to me

Most chatbot setups are built to fail gracefully. They answer the easy stuff and then escalate to a human the moment things get real. Mine does the opposite. It always replies.

Somebody says "ang mahal naman." The bot does not apologize and it does not go quiet. It offers a cheaper room, because I actually have cheaper rooms, and a cheaper room is a booking while a discount is a bleed.

Somebody says "pwede po discount?" The bot says there is no discount. Plainly, politely, and every single time.

That second one is the part most owners cannot do for themselves. It is 11 PM, the room is empty tomorrow, someone is asking for five hundred pesos off, and you will cave. I have caved. The bot never does, because it is not tired and it does not feel the empty room the way I do. Holding your own price is a discipline problem, and I solved mine by removing myself from the conversation.

Here is the part people expect me to hide: the bot tells guests it is an AI assistant. It says so. And if someone wants a real human, it tells them to give it about an hour and I will reply myself. No pretending, no fake name, no fake profile photo.

Every business owner I talk to is scared of exactly this. They think guests will feel tricked, that AI will kill their brand, that customers can tell and will hold it against them. I have run 25,361 messages through a bot that openly announces itself as a bot, and I have been 100 percent booked the entire time. The fear is real. The consequence is not.

And the reason it can hold a conversation without me is that it is not answering from a script somebody wrote for it. It is answering from my answers. Every reply it gives is a copy of a reply I already gave to a real guest. Over enough months, it stops sounding like a chatbot trained on my business and starts sounding like me personally.

The 30 questions that make it smarter every month

This is the part of my system nobody writes about, and it is the reason the whole thing compounds instead of going stale.

The bot handles the 25,000 messages. Anything it cannot answer, it logs and passes to me. Then, once a month, it hands me the pile as a list of about 30 questions. Real gaps. Real things a real guest asked that my system did not know.

I sit down and answer those 30 questions once. They go back into the system. Next month, those questions never reach me again, because the bot handles them now. And a fresh 30 shows up, a little further out on the edge of my business than the last batch.

  • Month one, it cannot answer basic things and I answer a lot.
  • Month four, the questions are strange and specific, because everything ordinary is already automated.
  • Month eight, I am answering questions I did not know my own guests cared about.

Most people build a chatbot once, load it with an FAQ, and then quietly watch it get worse as the business changes around it. Mine gets better every month by exactly 30 answers. That is not a feature I bought. It is a loop I built, and it costs me maybe an hour a month.

730,505 views, and that is my own face on the screen

Facebook reel insights showing 730,505 views, 9,727 interactions, 33.9 percent hook rate, and 2,724 new follows
One reel, June 19: 730,505 views, 33.9 percent hook rate, 577 shares, 2,724 new followers.

A 33.9 percent hook rate means a third of everyone who scrolled past that video stopped. On Facebook, that is the whole ballgame. And 2,724 people followed a transient house because of one reel.

I am the person in the video. Not a stock avatar, not a generic AI presenter. I subscribed to Gemini, built an avatar of myself, set the head movement, and it generates a 10-second video whenever I need one. Ten seconds. That is the format.

The tool is not the interesting part, and I want to be blunt about that. Anyone reading this can subscribe to the same thing this afternoon. The interesting part is that I post one every single day. Most people make three, get a few hundred views, decide reels do not work for their business, and stop. I have posted daily long enough that one of them was always going to hit, and when it hit, it hit for 730,000.

That is also why the reel and the bot belong in the same post. A viral reel with no system underneath it is just a nice screenshot. The reel dumps hundreds of people into my inbox at once, and the $6.85 bot catches every one of them at 2 AM while I am asleep. Attention without a system is a vanity metric. A system without attention is a very well built empty room.

I left WordPress, and it had nothing to do with speed

I spent 11 years as a WordPress developer. I did not leave because it was slow, and I did not leave because of plugin bloat, even though both of those are true.

I left because Claude Code cannot work on WordPress.

That is the honest reason. WordPress is genuinely the friendlier platform if a human is doing the clicking. But the moment your best worker is an AI, friendliness for humans stops being the thing you optimize for. An AI cannot open your dashboard, poke through a plugin settings panel, and reason about what your theme is doing to your markup. It needs files. It needs to see every page, every schema block, every internal link as something it can read and edit directly.

So I rebuilt on Next.js, where the whole site is a repo. Now when I want to change how something ranks, I do not go hunting for the plugin that controls it. I point Claude at the file. SEO stopped being a setting somebody else built and became something I actually control, line by line.

If you take one technical decision from this entire post, take that one. Choose the stack your AI can work in, not the one that was easiest for you in 2019.

The flat line, and then April

Google Analytics reports snapshot showing 3.94K 30-day active users up 2,103 percent and 1.09K 7-day active users up 2,840 percent
Flat from December through March. Then April.

Look at the shape of that chart, because the shape is the story. Dead flat from December 2025 through March 2026. Then it goes near vertical and does not stop. 30-day active users at 3,944, up 2,103 percent. 7-day at 1,090, up 2,840 percent.

I am not going to invent a tidy explanation for you. I do not know precisely which lever moved that line. It was the rebuild, the content, the structure, and the reels all landing in the same season, and I cannot cleanly separate them. What I can tell you is the date it turned and what has been true since: April, and I have been 100 percent booked from that month to this one.

What actually changed inside my business is simpler than the chart. Before this, I only knew basic SEO. I would have needed to hire an SEO guy and a content writer, and in Baguio that is real money every month for work I could not evaluate. Now I do it myself. I prompt Claude into a specialist with 20 years of experience and I get output at that level. I stopped outsourcing the two functions that most small businesses overpay for and understand least.

15 rooms, full every day

Fully booked, for me, means occupancy. It means 15 rooms with someone in them, every day, since April.

Put the two numbers next to each other and the argument makes itself. On one side, $6.85. On the other side, 15 rooms that would otherwise be sitting empty because I did not answer a message fast enough at midnight. In this business the booking goes to whoever replies first. Not whoever has the nicest photos, not whoever has the best rate. Whoever replies first.

That is what I am actually buying for $6.85. Not software. First place in the reply race, 25,361 times a month.

What I would tell you to build first, and it depends on you

Here is where I disagree with almost everyone selling AI to Filipino businesses right now. They sell the chatbot to everybody. The chatbot is not the right first move for everybody.

If you are already known in your area and your inbox is flooded, build the chatbot first. That is my situation and it is the situation of any business in Manila drowning in messages it cannot answer fast enough. You have demand you are actively losing. Plug the leak before you do anything else.

If you are just starting and you want to scale, do not build a bot. A bot with nobody messaging it is an expensive answering machine for an empty room. Build the SEO-friendly website first, on a stack your AI can work in, and go earn the messages. Then automate them.

Either way, post every day. That part does not depend on your stage, your budget, or your industry. Use Gemini, use whatever comes after Gemini, it does not matter. What matters is that you post daily long enough for one of them to break out, and that you have somewhere for those people to land when it does.

That is the whole machine. Attention from daily video, a site your AI can actually edit, and a $6.85 bot that catches everything and gets 30 answers smarter every month. I did not buy a platform for any of it. I built it, on my own rooms, with my own money at risk, and it has kept 15 doors full since April.

If you are looking at your own inbox right now and doing this math, my recommendation is to start with whichever of those three you are most obviously losing money on today. Not all three. One. The one that is bleeding.

Frequently asked questions

How much does it cost to run an AI chatbot for 25,000 messages a month?
My actual bill for one month was $6.85, covering 25,361 API requests and 280,109,038 tokens on DeepSeek v4 Flash. That is the inbox cost only, and it is possible because I chose the cheapest model that still answers guests properly in English, Tagalog, and Taglish. Running the same volume through a premium model would cost many times more for the same booking, which is why model choice at high message volume is arithmetic rather than preference.
Why use DeepSeek v4 Flash instead of a premium AI model for customer messages?
At 25,000 messages a month, the price per message decides whether the system is profitable. DeepSeek v4 Flash is currently the cheapest model I have found that still clears my quality bar for Filipino guest inquiries, so it handles the inbox. I use different models for different jobs: Claude for content, SEO, and landing pages, because those are judgment work, and Gemini for video. There is no single best model, only the cheapest one that clears the bar for each specific job.
Should an AI chatbot tell customers it is an AI?
Mine does, and it has not cost me a single booking. My bot introduces itself as an assistant and offers a real fallback: wait about an hour and I will reply personally. I have run over 25,000 messages through a bot that openly announces itself, and the property has stayed 100 percent booked the entire time. The common fear that customers will feel tricked and abandon your brand did not survive contact with 25,000 real conversations.
How do you stop an AI chatbot from giving discounts?
You instruct it to hold the price and offer a cheaper room instead. When a guest says the rate is too high, my bot does not apologize or go quiet, it presents a lower-priced room that is genuinely available. When someone asks directly for a discount, it says there is no discount, politely and consistently. This is the one area where a bot outperforms the owner, because at 11 PM with an empty room tomorrow, most owners cave and the bot never does.
Why move from WordPress to Next.js for SEO?
Not for speed. I moved because Claude Code cannot work inside WordPress. An AI agent cannot open a dashboard, navigate a plugin settings panel, and reason about what a theme is doing to your markup. On Next.js the entire site is files in a repo, so the AI can read and edit every page, schema block, and internal link directly. SEO stops being a plugin setting and becomes something you control line by line. Choose the stack your AI can work in.
What should a small business automate first, a chatbot or a website?
It depends on whether you already have demand. If your inbox is flooded and you are losing inquiries to slow replies, build the chatbot first and plug the leak. If you are just starting and nobody is messaging you yet, a chatbot is an expensive answering machine for an empty room. Build the SEO-friendly website first, on a stack your AI can edit, and go earn the messages before you automate them. Either way, post short video daily from day one.
How do you keep an AI chatbot from getting outdated?
I use a monthly gap loop. The bot answers what it can and logs everything it cannot, then hands me roughly 30 questions at the end of each month. I answer those 30 once, they go back into the system, and those questions never reach me again. Each month the remaining gaps get more specific because everything ordinary is already handled. Most chatbots decay because they are loaded once with an FAQ and never updated. Mine improves by exactly 30 answers a month, for about an hour of my time.

Want the same system for your business?

I'll set up AI automation for your business — just like I did for mine.

Related Posts

What Tasks Can AI Automate and Grow Your Business Right Now? The Honest Task List, From 10,000 to 30,000 Messages a Month
Automation

What Tasks Can AI Automate and Grow Your Business Right Now? The Honest Task List, From 10,000 to 30,000 Messages a Month

Most answers to this question are a generic tool list. Mine is the order I actually automated tasks in my own business: inquiries first, which took me from 10,000 to 30,000 messages a month after I switched off an expensive model, then marketing, which took a Reels habit from 100,000 to 300,000 views. Here is the exact task list, what to automate right now, and the one task, closing a sale, I still will not hand to AI.

The Business Came First, the AI Came After: What "Most Recommended AI Business Consultant in Baguio City" Actually Means
Growth

The Business Came First, the AI Came After: What "Most Recommended AI Business Consultant in Baguio City" Actually Means

Anyone can learn to prompt an AI model in a weekend. What they cannot learn in a weekend is how to run a business without one, the years of manual inquiries, follow-ups, and rejections that teach you what is worth automating. That is the real basis for being recommended as an AI business consultant in Baguio: operator experience first, AI second. Here is the case study, what an engagement looks like, the pricing, and honestly, who should not hire me.

The Backlog Was Never the Messages, It Was the Photos: How I Automate Customer Service for a Small Business in the Philippines
Automation

The Backlog Was Never the Messages, It Was the Photos: How I Automate Customer Service for a Small Business in the Philippines

People assume the breaking point for customer service automation is unanswered text. Mine was worse: guests sending photos (IDs, room requests, proof of payment) that piled up unread because a photo takes more attention than a text does. Here is the actual system I built to fix it, on a model that costs about $6 a month, the one thing it still cannot do, and who should not touch automation yet.