Skip to content
MakerGain
← Back to project

Teach the bot to admit uncertainty first: a side project bringing in about \$500 a month

Edited by MakerGain

Andy and his two cofounders at London's HeyDeacon built Deacon for founders and small teams still answering support themselves, starting with their own use. When it cannot answer, it admits that and asks visitors for an email address. Only after an address is actually provided is the conversation marked unread and the owner emailed; human replies can enter the knowledge base. The team focuses on the workflow around the model. Knowledge needs manual recrawling or publishing-script triggers to update, while changing models does not require rebuilding the product.

Andy is one of the three cofounders of London-based HeyDeacon. The other two were running several other software products, and both were receiving more customer messages than they could handle. Many were simple, repeated questions. They first built a bot to answer for themselves, then thought others might need it too. That became Deacon, an AI support agent for founders and small teams still answering support messages themselves (Starter Story interview[1]).

The most interesting part is what it does when it cannot answer, not how many things it can answer.

Admit it when you do not know

According to the official description, when a visitor asks a question on a website, Deacon looks only in that company's content: web pages, uploaded PDFs, Word documents and other files, pasted text, and answers written by you and your team. It usually links to the page it used, can reply in the visitor's language, and can read screenshots the visitor pastes into the chat (official FAQ[2]).

If the documents do not cover the question, it says it does not know and asks for an email address. Once the visitor leaves that address, the conversation is marked unread and the owner receives an email. A visitor explicitly requesting a human also triggers an email notification. The owner replies in the dashboard, and the reply appears directly in the visitor's chat. Clicking “answer this” for a question adds the answer to the knowledge base, so the agent can answer the next person itself (official FAQ[2] · [1]).

Image: Conversation interface shown on Deacon's website, with a human taking over an unanswered question. From the official product page[2], obtained on 2026-11-18.

The knowledge base does not update itself

That loop sounds convenient, but it exchanges one problem for another. The official guide is explicit: Deacon does not reread the website by itself. Changed pages require a manual recrawl or an API trigger from your own publishing script. Each page can be recrawled at most once in 24 hours; a crawl covers up to 100 pages. Login, cart, checkout, and account pages are never crawled (user guide[3]).

The “knowledge base,” then, is documentation that someone must keep current. It can be recrawled manually or triggered through the API by a publishing script, and team members with seats can maintain it too. The saving is in repeated answers, not maintenance. Old pages mean old answers.

He is changing the rules around the model

Deacon was originally made for the other two cofounders' products; the team then began telling people about it directly (interview[1]). Early versions had gaps in understanding customer intent, so the team changed the logic and how unrecognized questions were handled (same source).

Andy calls that layer a harness around the language model: find answers in the documents, say when it does not know, hand over to a person, and swap in a new model when models improve. In his account, the bet is on the workflow around the model, not on one model being smarter than another. Changing models does not require rebuilding the product.

The interview does not identify the model provider, hosting arrangement, or conversation-data retention period, so the harness is presently a direction, not an established moat.

A monthly price on the page, a year charged upfront

According to the pricing page on November 17, 2026, the displayed annual-billing prices were: Free at $0/month, with 50 answers a month—about 25 conversations—1 seat, and no credit card required; Starter at $40/month, but charged as $480 for a full year; and Growth at $120/month, charged as $1,440 upfront (pricing[4]). All three have identical features, differing only in answer allowances, knowledge-base size, monthly readable-page allowances, and seats. There are no overage charges or automatic upgrades, and all plans use the same model (About[5]).

That detail deserves a pause: the price says “month,” but the bill is annual. Choosing annual Starter requires $480 upfront, not $40 every month. Those who do not want to pay can use the ongoing free plan.

The owner receives emails at 80% and 100% usage. After the allowance is exhausted, visitors see a brief message and are asked for their email address, without an additional charge. The owner can still read and reply; automated answers resume in the next billing month or after an upgrade. The free plan has no end date. An account closes only after a continuous year of inactivity, with a warning beforehand (pricing[4] · [5]).

It does not take action in other systems either: no refunds, plan changes, or order lookups. It declines visitors' account-specific questions and has no ready-made helpdesk or Slack connection (official FAQ[2]). Coding agents can use MCP to read customer feedback and page errors and propose code fixes awaiting your approval (pricing[4]). The direction is consistent: organize the problem and hand it to a person. Zendesk handles the same website-support task but with a different action scope: its AI agents can look up order status across systems, update account information, and execute approved actions, preserving context for human handoff when needed (Zendesk's official explanation[6]). Deacon explicitly does not look up orders, change plans, or issue refunds; account questions go to people (official FAQ[2]).

$500 does not tell us much

The only public financial information is self-reported: the project began as a side business in June 2026 and was bringing in about $500 a month by the September 23 interview. That was not described as profit and was not audited. Costs, gross margin, churn, and customer count are undisclosed, so an outsider cannot calculate whether the business is profitable.

At that interview, they had not held a major public launch. They were telling people directly, with Product Hunt planned. The first paying customer came through organic search: someone running a small SaaS in Japan who needed help with customer requests. Growth activity included finding founders on Reddit, Hacker News, Indie Hackers, LinkedIn, and X; creating SEO comparison pages, platform pages for tools such as Lovable/Bolt/v0, and how-to guides; joining free software directories; and showing in videos and images how the agent answers, cites sources, and hands off to humans (interview[1]). All of that is self-reported, without channel-level conversion or revenue data. A channel appearing in the story does not mean it generated money.

Its own FAQ groups Fin and Decagon with companies that already have support teams and helpdesks, says Chatbase and Crisp are closer comparisons, and positions Deacon for founders and small teams with neither (official FAQ[2]). This is a vendor's comparison, not an independent test. It at least states where it wants to stand: not replacing a support department, but serving people who have no support desk to begin with.

Andy has been a software engineer for ten years. In the interview, he says his biggest mistake was not starting earlier and advises building for a specific person who needs it now. Deacon first served the two cofounders, so whether the problem was real was never their concern. The answer-first, hand-off-when-stuck workflow also closely copies the actions of a founder answering messages personally.

It defines two failures simply: admit an unknown answer and hand it over; stop at the usage limit, ask for an email address, and do not charge. For a founder answering support and worrying about bills, one missed answer is a nuisance; one invented answer could mean a customer who never comes back. Andy sells not answers, but the moment an unanswered question reaches the right person.

Continue reading

AI support
Human handoffs are free: Fini puts its AI-support bill on “actually resolved”

Fini's AI support agent handles chat tickets inside a customer's existing helpdesk. Behavior trees and a deterministic executor read plan and account state node by node, then call customer system interfaces to perform actions such as canceling subscriptions. Human handoffs are not billed; by default, only end-to-end resolutions without human takeover count as billable tickets. The team cleared its first product and rebuilt after four weeks embedded with a fintech support team, choosing not to replace existing helpdesks and requiring human approval before publishing knowledge corrections proposed by background AI.

AI support
The database crashed in his first week: Zipchat wants every store's support ready out of the box

Zipchat is an AI support agent for ecommerce. When Ruslan took over in January 2024, the team struggled to account for its MRR and customer sources, and the database started crashing in the first week. He first fixed the books, using Mantle to make MRR visible and collecting feedback on replies. His blog reports a 25% MRR increase about two months later, reaching US$20,000 for the first time. The product claims 30-minute setup, but brand rules, helpdesk connections and answer corrections still require store-by-store configuration.

AI commerce
He asked if people would pay before building LeadFlow. He started by making cold email easier to set up

Before building LeadFlow, 21-year-old German university dropout Adrian Berisha asked people in his marketing network and independent founders whether they would pay for a cold email tool that was extremely simple to set up and could get emails into the inbox. After hearing that people would pay, he put campaign configuration, mailbox setup, AI personalization, reply analysis and CRM handoff in one workflow. He used his own product to run cold email campaigns, booked meetings over LinkedIn and helped early customers get set up through a done-with-you service.

AI images
Pykaso turned “make the next image look like the same person” into a business

Pykaso AI fits between defining a virtual character and publishing content. It helps creators make subsequent images look like the same person. The team bootstrapped the company and combined training and generation in one platform from its first version. To promote it, the team pursued organic growth on social media and reached out to creators. Archived pages differ on how Pykaso measured usage and how it charged.