Skip to content
MakerGain
← Back to cases

Seven AI employees with names, and 50 hours of work a month

Edited by MakerGainPublished: Article updated:

Marblism founder Ulric Musset once hired Sarah to handle his inbox, calendar, and email, alongside a LinkedIn ghostwriter and an SEO expert. He cloned these virtual-assistant roles into named AI employees handling jobs such as social media, outreach, and writing. The product shifted from building web apps for people to putting AI to work inside their companies. People assign tasks in a room, gather and inspect outputs, and can stop agents when something is wrong. Billing is based on monthly AI work rather than message counts, with no conversion table from hours to emails or articles. The reported $250k ARR has not been clearly attributed to a particular business.

Marblism founder Ulric Musset described a starting point in his letter on the website[1]: he once paid $2,000 a month for an administrative assistant named Sarah to handle his inbox, calendar, and email replies. He also hired a LinkedIn ghostwriter who grew his audience to 10,000 followers or more, plus an SEO expert to publish articles that could rank. He says he “cloned” all these virtual-assistant roles into AI Employees.

There are seven names on the product’s pricing page[2]: Eva handles administration and email; Sonny handles social media; Stan finds leads and does outreach; Penny writes SEO articles; Rachel answers front-desk calls; Walter builds websites; and Linda is a legal assistant. Every plan includes all seven. The FAQ says, “Plans start at $24/month on the yearly option,” and the page also advertises a 7-day money-back guarantee.

Image: From the official product demonstration[3], obtained on 2026-11-18; official presentation retained as shown, for product illustration only.

Giving all seven AI employees names was probably not arbitrary. You send a brief in a conversation or @ Stan to contact a batch of leads, rather than “call the outreach module.” Packaging features as jobs makes assigning tasks and checking output feel more like managing a small team.

It was originally a different product

On September 17, 2024, Musset posted a Launch HN[4] on Hacker News. Marblism then helped people build web apps: describe what you wanted, and AI would build the data model, generate an app with login, payments, and email, then generate frontend pages. You could test it online, change the code yourself, and deploy it with one click when satisfied. At the time it could generate apps such as SaaS products, marketplaces, and social networks. It did not support Chrome extensions or mobile apps and was not suited to small games like Snake or Tetris. Asked whether it could connect to an existing PostgreSQL database and generate an interface from it, he said not at the time, but that it was a good idea and technically possible.

That was a product for building other people’s companies. Today’s Marblism puts AI to work inside your own company. The two share a domain, but almost no workflow.

The product behind the $250k ARR remains unresolved

The headline of the Indie Hackers story[5] reads “Killing his AI product after growing it to $250k ARR within a year.” Its introduction attributes that figure to the old AI engineer product. But in the body of the same article, the founder writes “We quickly hit $250k ARR” only after introducing the AI employee suite—administrative assistant, social media, SEO blogging, and sales assistant. The public material does not align on whether the figure belongs to the old product or the current AI employee business.

It cannot stand in for today’s revenue, nor is it used here to calculate anything. What is interesting is the pivot itself: from selling “help building an app” to selling “doing a job for you.” The first delivers an app that can launch; the second delivers daily tasks. What is sold has shifted from making something to keeping work going.

People still inspect the work, and the allowance has no conversion table

The Rooms announcement[6] puts multiple AI employees and human colleagues in one conversation. Ulric described the hassle of his previous workshop: he spent an entire morning pasting the same brief into separate chats with Penny, Sonny, and Stan, then copying Penny’s article to Sonny so the posts would match. This time, he opened a room and sent one message: Penny would write a blog post, Sonny would adapt it into LinkedIn and Instagram posts, and Stan would find 200 local agency owners and send invitations. Requests are automatically routed to the appropriate agent by role, or an @ can specify one. Agents can hand work to one another, outputs are collected in the Work tab, and one click can stop all agents.

In that workflow, the person has not been removed, just moved further along: assign work, inspect outputs, and stop things when something looks wrong.

Billing is also a product choice. The pricing page[2] says each plan includes 50 hours of AI work per month, reset monthly. Only advanced AI operations consume hours; chat and quick tasks do not. Extra hours can be purchased at any time after the allowance runs out. Selling “work” rather than message counts is conceptually closer to hiring, but it also leaves a question: how many emails, outreach messages, or articles do 50 hours buy? The page has no conversion table.

So the $2,000 human assistant and the seven-agent bundle starting at $24 should not simply be equated. One is a defined monthly salary; the other is an allowance measured in hours without a fixed email or article conversion. The $24 is a starting price under annual billing. With monthly billing, the current plan card shows $44/month.

Acquisition and cost share a ledger, but only half is visible

Marblism’s partner program[1] is a public acquisition channel: the entry Ambassador tier pays a 30% lifetime commission; Growth rises to 40% after 10+ conversions, also for life. The website reports[1] 7,000+ active partners and $200K+ paid to partners. The terms[7] are more specific: commissions are settled monthly, with a $50 minimum and smaller balances carried over. Commissions may be withheld or clawed back for refunds, cancellations, chargebacks, suspected fraud, or prohibited promotion.

“Lifetime” deserves a pause. It permanently attaches a revenue share to each renewal. That is a payment structure, not evidence of acquisition cost. The $200K+ is the website’s own reported amount paid, not total customer acquisition cost, much less an income statement. Outsiders cannot see what remains after AI inference and hosting, payments, support, and commissions.

The website[1] also reports the trust of 60,000 businesses and a 20-person team across 5 countries. The YC company page[8] records a 2024 founding, the W24 batch, a team size of 5, London, and Active status. Both sets of figures remain as stated, without merging them.

For comparison, Teamday[9] is in the same category with different conditions: $99/month covers an entire AI marketing team, provided you connect your own Claude/ChatGPT account, whose AI usage is billed separately at that account’s rates. Without an AI account, users do not take the $99 subscription; they use prepaid credits starting at $20 that never expire, charged per use at model prices plus 25%. One starts at $24 with an hourly allowance; the other ties a $99 monthly fee to a user-supplied model account. Each has its own arithmetic. Neither can be declared better value for every customer.

Return to Sarah and her $2,000. The real bet is not whether AI can reply to emails, write articles, or make calls. It is whether someone will hand over an inbox, an outreach list, and a front-desk phone, then get used to inspecting the results in the Work tab.

Continue reading

AI productivity
Over 180,000 views, 1 February sale: AI Renamer leaves batch-renaming confirmation to people

Ozgur Ozer's AI Renamer gives content-based names to opaque filenames such as DSC_0412.JPG and scan_20250312.pdf. For manual batches, people inspect the suggested names before pressing Rename All; nothing changes on disk beforehand. Ozer had always given away his small apps and first used Polar to solve payment collection because Turkey had no Stripe support. His first post exceeded 180,000 views but produced 1 sale that month; sales began coming after local support and a Product Hunt launch.

AI productivity
After closing a product nobody wanted, they broke entrepreneurship into “the next step”

David Heikka and his brother Felix launched their first AI form product and found zero demand. They tried ads, landing-page changes, renaming, and added features, then closed it after six months. They went on to build Buildpad, breaking business creation into stages that indicate what to validate next. Its early and current billing mechanisms differ, and the product was renamed aicofounder before changing back to Buildpad.

AI productivity
It writes your emails, but stops at LinkedIn

According to its website’s workflow, DeepReachAI finds people, researches recent developments, writes messages, classifies replies, and books meetings for cold outreach, while excluding automatic LinkedIn sending, WhatsApp scraping, and calls without consent. In an IndieHackers post, founder Aryan Agrawal said he had expected personalization to be the hook, but the compliance boundary—no automatic LinkedIn sending and no WhatsApp scraping—was what made people connect accounts. The homepage says Orio can send a first email through a connected account, without identifying the account provider.

AI productivity
An assistant cost thousands of US dollars a month. He put AI to work and reached $10k+ MRR

Because an assistant would cost thousands of US dollars a month, Elie Steinbock built an AI email assistant instead. It connects to an inbox, sorts messages, and drafts replies using email history, calendars, and customer records; a person still presses Send. On Product Hunt launch day, he emailed 20,000 newsletter subscribers and posted a video on his YouTube channel with 2,000 subscribers. He estimated that these extra efforts may have added fewer than 50 votes—a judgment, not a measured result.