Skip to content
MakerGain
← Back to project

Launched at one in the afternoon, first payment thirty minutes later: Chatbase turned company documents into a support agent that can look up orders

Edited by MakerGain

Yasser Elsaid launched Chatbase in 2023: upload a document and get a customized support bot; within 30 minutes of launch, there was a first payment. The product expanded into a customer support agent across chat, email, and voice that can look up orders, update records, and hand off to a human. The company is self-funded, with Basic at $150/month and Business at $500/month; compared with SiteGPT on action capabilities and message quotas.

Yasser Elsaid chose one in the afternoon to put Chatbase live. At the same time, he sent a tweet and put up a pricing page. Thirty minutes later, Stripe popped up a notification: the first paying customer had arrived. Ten minutes after that, the second; about an hour later, the third.[1]

He later said that at that moment, he knew he had to stop everything else in his life.

At the time, what Chatbase did was so small it hardly looked like a business: upload a PDF and get back a custom ChatGPT trained on your own material.[2]

But the problem it aimed at was not small. Companies always had piles of documents, FAQs, and policy pages that no one finished reading, and support staff repeated the same things over and over. Throw the material in and let the machine take the first round: that was the easiest step to think of, and the easiest to validate the same day.

Now the pitch on the same website has taken on a different face.

From talking to taking action

According to Chatbase’s official website, it is a conversational AI agent for customer experience that can be built, deployed, and optimized across chat, email, and voice; its answers are grounded in “approved knowledge,” and it can also look up orders, update records, create tickets, and trigger workflows; when a human really needs to take over, the full conversation, context, and customer information are handed over together.[3]

The words “approved” and “trigger” cannot be skimmed past. The former says answers are required to stay on approved knowledge; the latter says it is not just a search box that talks.

Crossing from “answering questions” to “changing things” is a different order of risk. If it gets one sentence wrong, an apology will do; if it changes one order record wrong, an apology will not. That is also why “carry the context over during handoff” deserves its own line in the product description—what the AI hands over cannot be a conversation that has been cut off.

The return process on the official site is an example of this shift: first ask the customer what they want to return, request an order number or email, look up the order details, confirm which item is being returned, and then check whether it is within the return window; when a human needs to take over, the full conversation, context, and customer information are handed over together.[3]

See the gap first, then wait for the models to catch up

Elsaid did not work backward from “I want to build a customer support bot” to Chatbase.

He studied computer science in Canada, and the default path around him was to get an internship at a big California company—they called it Cali or Bust. He got one, did a stint at Tesla and at Facebook, and then quickly found that it was not quite the life he wanted: too safe, too structured. The difference between five years later and ten years later might just be going from an E3 engineer to an E6 engineer—higher pay, higher status in the company.[1]

What really drew him in was the group of indie developers on Twitter—have an idea, sit down and build it, throw it out to the world, and see if anyone will pay for it. He built a few small projects himself, and it was in those projects that he saw a gap: however strong a model was trained on general data, it knew nothing about a specific person or a specific company. The first example he gave was a book—upload a textbook and then talk to it.[1]

In 2022, before ChatGPT had been released, he stopped going to class and started working on this. The models were not strong then, and many use cases could not be fully served; they could only be brought to “good enough, demonstrable, and worth it to some customers.” What he bet on was not how good the model was at the time, but that the model would definitely get better—as long as he built the scaffolding around the model, he and his customers would benefit together on the day the model got stronger.[1]

Self-funded, and an uncomfortable order

Chatbase has never raised outside funding. Elsaid’s reasoning is direct: being self-funded means being accountable only to customers and the team, with control in his own hands; funding, by contrast, rewrites “what counts as success”—the moment a term sheet is signed, an outcome that might originally have counted as a win suddenly looks too small.[1]

The cost is that every expense has to be justified by ROI first, with an eye on the cash in the account, and efficiency pushed as high as possible. He admits that for a long time he carried the mindset most common among self-funded founders: everything had to have positive ROI, everything had to save money, everything had to minimize risk.[1]

Then he turns it around and says this is exactly the mistake self-funded founders make most often—keeping that mindset for too long. Chatbase really accelerated after he began taking risks: when the costs still did not add up, he first hired people he considered expensive but good, and first ran business experiments whose short-term math did not work. The order is uncomfortable, but he says revenue started climbing from that point.[1]

This is exactly the opposite of the intuition that “self-funded means conservative”: saving money is the starting point of being self-funded, not its conclusion.

Revenue, pricing, and a line that cannot be erased

The numbers climbed steeply. He says that in the first month MRR was $3,000, in the second month it was close to $40,000, then $60,000, and 117 days after the first tweet, ARR reached $1M.source[1] He also said that after revenue covered model costs, the company became profitable and has remained profitable.[1]

A line has to be drawn here: ARR is not profit, and neither is MRR. The above is the founder’s own account; the inference costs, labor costs, gross margin, or net profit for the corresponding period were not disclosed, so how profitable “making money” actually was remains unclear.

The post-scale numbers come from two other places. Supabase’s customer story says that as of early 2026, Chatbase had more than 8,000 paying customers, over $10,000,000 in ARR, and 18 people, 11 of them engineers.source[2] The body of Stripe’s customer story says it reached $10,000,000 in ARR in March 2026.[4]

The public pricing page (as of November 14, 2026) divides plans into three tiers by team and usage: Basic at $150/month, Business at $500/month, and Enterprise with custom pricing; annual billing gets another 20% discount, and the page does not give the actual total annual charge. When the quota is not enough, auto-recharge is $40 per 1,000 message credits, and the credits do not expire.[5]

Image: source Chatbase official pricing page[5], retrieved on 2026-11-10; original interface preserved.

The positioning on the page follows the same line: Basic is for small teams just launching their first agent, Business is for teams whose customer base is growing, and Enterprise is for companies deploying AI agents at scale.[5]

In this arrangement, three groups are clearly separated: the ones actually talking to the agent are the customer’s own customers; the ones changing instructions, writing flows, and watching analytics in the backend are the customer’s support or operations team; and the one paying for Chatbase is that company. The deliverable is not a customer support service built for you, but a set of tools that lets customers build agents themselves, plus a bill charged by message credits.

There is a counterintuitive approach to customer acquisition. Elsaid says that of the people the team proactively contacted, about 80% had already visited the website, signed up, or churned, rather than being cold leads who had never heard of the company; the follow-up was personalized outreach to this group that already had awareness.source[1] This explains whom they contacted, and it also shows that the founder attributes the high conversion of personalized follow-up to the fact that the other party already had awareness; but specific conversion rates and channel costs were not disclosed. More than 8,000 customers is also just a total; it does not show how much comes from large customers versus small ones.

Delete a document, and the vectors have to follow

There is a kind of trouble that does not make it into marketing copy, but it determines whether customers dare to hand over their data: if a customer deletes a document, or deletes an agent outright, are the vectors that were split out from the documents still there in the backend?

Chatbase moved its vector embeddings from Pinecone to Supabase’s Postgres + pgvector. Yasser says that a few years ago Pinecone was the default way to do RAG chat, but their path was the opposite—not splitting things apart more and more, but consolidating them more and more. The direct benefit of the move: when a customer deletes a document or deletes an agent, the related vectors are cascade-deleted in the same database; if the vectors lived in another service, the application would first have to delete in Supabase, then call the vector service, retry on failure, and also reason through what to do when “one side succeeds and the other fails.”source[2]

This is not a selling point that can go into ad copy, but it answers everyday questions like “who cleans up when data is deleted, and who handles the aftermath when something goes wrong.”source[2]

An alternative

Compare the same task: feed in the content a company already has and get back a bot that handles the first customer contact.

SiteGPT’s Starter plan is $59 per month, or $468 per year (equivalent to $39/month), including 1 chatbot and 1 member; the monthly quota is converted by model, up to 4,000 GPT-4.1-mini messages or 400 GPT-4.1 messages, and with a stronger model it drops to one-tenth.source[6] The capabilities listed on its website are: train on a web address, file, or raw text, embed on any site, hand off to a human, collect leads, view analytics, and connect tools such as Zendesk.[7]

Both companies tier their plans by message quota and bundle seats with the tier, and both let a company exchange its own content for a bot that handles the first customer contact. The differences are how the quota is converted, how many seats are included, and the fact that Chatbase has written actions such as order lookup and record updates into the product itself—its official website lists “taking action” as a separate capability; what SiteGPT’s official website describes is still mainly answering questions, handing off to a human, and collecting leads.[3] source[7]

What this business is really selling

What it sells is the judgment of “which step can be handed off and which step must stay with a human,” plus a mechanism for cleaning up the aftermath for customers—deleting cleanly, handing off completely, and having someone there when something goes wrong.

Continue reading

AI search
He Treated Reddit as a Survey. 135,000 People Used It for Research. In 2025, It Stopped at a License

Fed built GummySearch into a tool for audience research using public Reddit discussions: it searched and tracked community conversations by categories such as pain points and people seeking solutions. After commercial licensing talks with Reddit failed, he chose to stop sales on November 30, 2025, and gradually shut down the site rather than continue operating without a license; paying users kept transitional access, and the free tier retained only login, export, and account deletion.

AI supportAI searchAI commerce
Customer Support Bots Start Looking Up Orders for You, and Chatwith Bills “Talking” and “Acting” Separately

Chatwith bills customer-support answers and external actions separately: messages use credits, while actions successfully executed through MCP are billed separately; plans start at 9 US dollars, and an extra 1,000 MCP actions cost 6 US dollars. The founder says he put about 21,000 US dollars of AppSumo revenue into Google Ads in 2024 and at one point a Google Ads bill went unpaid; TrustMRR shows MRR of 3,622 US dollars and 42 active subscriptions in November 2026.

AI audioAI productivityAI search
You Ask Claude What the Client Said Last Week, and It Digs Through Your Voice Notes

Speechyou connects voice notes and meeting transcripts into Claude, letting users search, quote, and jump back to the original audio; the connection uses OAuth 2.1 and is read-only. The annual prices on the two pages are $67 and $100, while monthly is $15 on both. The founder self-reports MRR 1800$; the TrustMRR record shows $1,351 and 104 subscriptions, and the one-person team has listed the product for sale at $35,000.

AI supportAI productivity
Card first, then entry: ChatDash has turned payment collection for AI voice agents into an annual-fee business

ChatDash turns a white-label AI agent dashboard for agencies into a business selling customer slots annually: customers must first add a card before entering the dashboard, and Stripe automatically handles subscriptions and usage-based billing; SaaS Mode is limited to a Growth add-on and supports only Retell voice and OpenAI text. As of 2026-11-14, the TrustMRR page reports cumulative revenue of $603,545 and MRR of $17,167.