How he sells AI agents
for 2750 euros plus a subscription

1. Who Carson is: a solo designer whose agents squeezed out websites

Carson is a solo web designer from Belgium. The host opens the episode with exactly that framing: he is not a startup founder or a venture-story founder, but a person with an already working design agency who started offering agent setups to his existing clients (00:00).

A key detail Carson repeats later in the middle of the conversation: the design agency did not go away, it still runs and still takes his time (16:11). This is not 'sold the business and moved into agents', it is a parallel line on top of an existing client base.

Analysis: it is the existing base that makes the model possible. First sales do not come from cold traffic but from people who know Carson from websites and trust him. They know him, watch his videos, so buying from him is easier (06:07).

2. Pricing: about 3000 dollars per setup, 250 euros a month for support

The host gives the figure in dollars: about 3000 per setup and around 300 a month just to manage it (00:07). Carson clarifies in euros: 2750 per setup (08:07).

On the monthly part Carson is precise: 250 euros a month, and the sum already includes a 20 dollar OpenAI subscription (08:15). The client pays 250, while a third-party retainer of 20 sits inside.

3. Reselling subscriptions: earning on the gap between 5x and 20x

A separate revenue layer is reselling OpenAI subscriptions. A client who needs a 5x account is sold an upsell, while Carson has 20x accounts, and he earns on the difference (08:22).

Analysis: this is a margin add-on. Carson holds higher-limit accounts and spreads them across clients, collecting the full upsell price from each. Three people on one 20x account (08:30) is arbitrage on limits.

4. The market: zero competition in the country and marketing not yet started

One of the strongest facts of the episode: in Belgium Carson literally has no competitors. The host states it in the intro (00:12), Carson confirms it later, in the context of media plans (26:08).

The audience is local business owners who speak Dutch. 5000 Instagram followers (00:20), all Dutch, all local entrepreneurs who themselves ask how to build an agent (00:23).

And most importantly: he has not started marketing yet (00:29). The entire inbound flow is an organic side effect of content about Claude and ChatGPT that he made for himself, not for sales.

5. Origin: n8n, why he did not sell, and the OpenClaw failure

The story starts with ChatGPT appearing about three years ago, when everyone began using it for emails and similar tasks (00:44). Then Carson and his team started implementing n8n solutions for clients, but deliberately did not sell it as agents (00:56).

The reason for not selling early is purely engineering: everything broke. A Windows update crashed the agent, and you could not sell that to a client. He deliberately waited for reliability instead of selling something raw.

The next stage is OpenClaw, which Carson tried, like many others (03:15). His review is blunt: it was a pain.

After that he went back to Claude (03:29), and then came to Hermes, which already existed but for him 'appeared' right then (03:32).

6. Infrastructure: VPS did not fit, an old gaming PC as the first 24/7 agent

Carson then describes his search for hardware. Everyone was buying up Mac Studio (03:47); he tried a VPS but it did not work the way he wanted (03:51). In the end he took an old gaming PC and ran the first 24/7 agent on it (03:58).

Analysis: the typical evolution of an enthusiast searching for a setup by trial. The VPS failed on feel, while the home PC gave control and constant operation.

7. The talk with Mikail and the move to Orgo

The turning point is a talk with Mikail, the co-founder. Carson told him he saw a video about a remote agent and had to find it (04:14). He went back to the iPad, where he watches AI content every evening, found in the history a video with the host and Greg and realized it was about Orgo (04:14).

Then he suggested Mikail try Orgo: now you can manage agents from anywhere and build an agent that manages the rest (04:38). He shut down his home desktop and moved to Orgo with his full setup (04:57).

A week after moving to Orgo he decided to build a business (05:04). Analysis: the decision came not from a business plan but from technical confidence that the setup works and is manageable.

8. Sales: inbound via ManyChat and a conversation about client pains

Asked what he sells in the first five minutes of a conversation, Carson answers matter-of-factly: he uses ManyChat like everyone and people just text him that they need an agent (05:30). But then comes the key correction: people do not know what an agent is, at least in Belgium and the Netherlands (05:35).

So the sales mechanic is not an agent demo but a free call and listening. They book a meeting, it is free, and Carson listens to what is needed and what problems the company has (05:41).

The answers are all the same: automate the routine, too much work, a messy inbox (05:48). Only after that does Carson talk about a 24/7 agent that is better than ChatGPT and Claude (05:56).

9. The sales funnel from inbound message to installed agent

Let us assemble the funnel from what Carson describes in pieces across different parts of the conversation. It is not linear and almost not automated: an inbound message, a free call, recording via Fireflies, gathering pains, on-site setup, an overnight task, a follow-up in a month.

10. Differentiation against ChatGPT and Claude: memory, Honcho, cloud, support

When a client says 'I already have ChatGPT or Claude', Carson answers with a set of differences: memory (06:29), Honcho (06:41), cross-platform (06:47), cloud as security (06:56), support (07:03).

The second is Honcho, which Carson calls a deal breaker. By his account, Honcho forgets nothing and knows what you are saying (06:41).

The third is cross-platform: you can write from Telegram, WhatsApp, iMessage and return to Hermes, and the agent remembers what was said in the Telegram chat (06:47).

The fourth is cloud as a form of security: the agent lives in the cloud, not on a personal computer, and has no access to all your files (06:56). The fifth is support: when something breaks, they help remotely (07:03). The sixth, from the host, is the absence of lock-in to a single model and stack, with the Jev example (07:45).

11. On-site setup in person and why you need to watch the client's day

Carson stresses that the setup happens at the client's home (08:33). The reason is not ritual but observation. It matters to them to see what the client does during the day, so they can implement several aha moments right away (08:41).

Analysis: the on-site setup is reconnaissance. In two or three hours Carson gathers the client's real work scenarios, which cannot be pulled from a form. This makes the agent relevant from day one.

12. The overnight task as a technique: the client wakes up to a finished result

The second required element of the setup is the overnight task. Carson deliberately implements something the agent does at night, so that in the morning the client wakes up and sees a finished result (08:51).

The host calls it a sticky moment that should be given to every client (09:46). The overnight task solves the first-impression problem: the client should not have to imagine the value, they should see it in the morning.

13. A client case: she paid for setup without knowing why

The first client story is a woman, an old client of Carson. She said she wants to start using the agent, will pay, just do the setup, though she does not know what she will use it for (09:03).

Two days later they implemented an integration with the CRM Alphas Profits, a large system, and the client came back with the reaction that the agent can do everything (09:23).

Analysis: the sale happened before the value was understood. Trust in the person replaced understanding of the product, and the client only saw the payoff after the integration. The model runs on warm contact, not on calculation.

14. The Derek case: the client took the offer for a scam

The second story is stronger than the first. A morning meeting, the client has not signed up yet. He sells parts for Posh bars and prepared five pages of what he does during the day and where he wants help (10:18). At the end of the meeting Carson said the agent can do that and that it can be implemented right at setup, which takes two to three hours (10:40).

Derek's reaction was disbelief: he could not believe it was possible (10:45), and Carson says outright that the client took it for a scam, as if they just wanted to sell him something (10:58).

In fact the task was simple: scraping websites and working with a warehouse API (11:03).

15. The stack: a verbatim inventory from Orgo to file skills

Carson walks through the stack in order, and this is the densest technical part of the episode (11:17-13:33). It starts with the Jeff agent, which is connected by API to Orgo (11:31).

What Jeff can do: spin up a computer in Orgo, set up the Hermes agent, implement skills and guardrails (11:34). Then come the Hermes terminal with an OpenAI subscription (12:05), the desktop via Telegram and BotFather (12:13), agent mail, which is not given to clients (12:16), Obsidian and Dropbox by default (12:26), Tailscale as secure access (12:37). It closes with Composio (12:47), Honcho (12:52), Latitude (13:04), files and skills (13:24).

16. Stack diagram: Orgo, Hermes, interfaces, Composio, Honcho, Latitude

The host confirms the stack lines up: Composio handles connectors and authentication, Obsidian is the knowledge base, Honcho is the memory layer, Latitude is observability (13:35-15:46).

17. The Jeff agent: the one who spins up Orgo and builds a client in 22 minutes

Jeff is Carson's separate agent, living in Hermes on his desktop but also in Orgo (11:19). He can spin up a computer in Orgo, set up the Hermes agent, and implement skills and guardrails (11:34).

The speed of this process is impressive: Carson says that the day before, Jeff assembled a full setup in 22 minutes (31:07).

Analysis: Jeff is a meta-agent that automates product delivery. While manual setup takes two or three hours, the internal build of the infrastructure has shrunk to 22 minutes. That is the scaling lever.

18. Guardrails and client notifications on failure

Guardrails are implemented at the build stage. Their purpose: when something breaks, the client gets a message in Telegram or the desktop app instructing them to contact support (11:51).

Analysis: this is managing client expectations. The client is not left in silence when something breaks; they get a route to a solution, which for local business lowers anxiety and holds trust.

19. A warning about Composio: do not use your own account across different clients

One of the most practical tips of the episode comes straight from Carson: never use your own Composio across different projects (14:08).

The problem is that the client cannot click a Composio link and make the connection themselves if the account belongs to someone else (14:20). The right path: let the client create their own Composio account (14:30).

The host agrees he does the same and promises to pass this to the Composio team as feedback (14:35). Analysis: separating accounts gives both working links and incident visibility per client (14:43).

20. Honcho: memory without per-workspace cost accounting

Honcho is the persistent memory layer, and Carson warns about its weak spot: it is hard to tell how much a specific client's service costs (12:52).

The host clarifies Honcho's role: it is not a knowledge base but short-term memory. The agent does not search the entire Obsidian vault for an answer, it has this fast memory (15:21).

21. Latitude as observability: free up to 100 thousand credits

Latitude is the observability layer that shows what the agents do. Carson estimates they have about 5 clients on Latitude (13:07), and that the plan is free up to roughly 100 thousand credits (13:10).

The host states the purpose: Latitude is your observability, you know what the agents are doing (13:16), and you get clear signals about the behavior of all agents (15:39).

22. Fireflies: meeting recording and early pain discovery

A separate tool in the sales process is Fireflies. It records the meeting, and it is present at the very first meeting (19:20). Because of this the client describes their own problems, and if Carson finds something suitable he implements it in the agent right away (19:36).

Analysis: Fireflies turns a conversation about pains into the source data for configuration. It links sales and delivery: what is said on the call becomes the agent's first task.

23. Speed and slowness: complaints from email and skill checks

The only notable negative feedback is slowness (24:09). The reason is technical: after connecting to Composio the agent is forced to check the email and the skills (24:23).

The host answers that such things speed up over time, and notes that Hermes suffered from the same problem as OpenClaw, namely harness bloat (24:42). He says the founder recently stated that Hermes will be slimmed down, made faster and lighter, closer to a PI-type agent (24:50).

24. Fulfillment: he does setups himself, teaches the co-founder from next week

Carson answers the fulfillment question honestly: the design agency is still on him, he has clients and tasks for them, though much is already automated (16:11). But he only has 24 hours a day, seven days a week (16:19). So for now he does the setups himself, and he will start teaching the co-founder next week (16:25).

Then the capacity arithmetic: if you calculate, you can do three setups a day, but realistically two a day, and with two people that is four a day (16:36).

Analysis: the model stays manual at the contact level, but with two people the arithmetic gives four setups a day. At 2750 per setup that is large turnover, if the inbound flow holds. The manual labor ceiling is the bottleneck.

25. The retention problem: the client works with the agent for a month and starts building alone

One of the most honest parts of the episode is Carson admitting the retention problem. People use the agent for about a month and are already building something themselves (16:57).

A month later a follow-up meeting is held to ask how things are going, whether the client is having difficulties and whether they can be helped with anything (17:07). But Carson sees that people are already building dashboards and the like themselves (17:18). So he and the co-founder need to sit down and think about how to do this better and how to earn more (17:22).

Analysis: the agent is so good that the client stops needing the contractor before the retainer pays off. The flip side of the model: the faster the client masters the tool, the faster they drop off the retainer.

26. The installation cycle step by step

Let us assemble the cycle from Carson's description: everything is already spun up in Orgo, but the connections on site remain. First Tailscale, then Composio, then the Hermes desktop, about an hour, longer if there is an IT team (18:01). Then Telegram (18:17), then Outlook, email skill, morning briefing and watching the client's day.

27. Setup tools: Outlook, email skill and morning briefing

After the connections, Carson asks the client what they work with and starts with Outlook: in Belgium and the Netherlands most people use Outlook, not Gmail (18:21). From this comes an email skill everyone needs (18:30), then the morning briefing (18:34).

Then the agent manages the mailbox: it decides which emails need a reply, remembers that you sent an email to a client and are waiting for a response, and reminds you of it (18:46).

Analysis: the set of base skills is the same for every client. This lets them sell a standardized product, while personalization sits on top, from watching the owner's specific day.

28. The management dashboard: clients, tokens, models, incidents, Orgo status

Carson shows an internal dashboard. It only holds clients who pay for support (26:23). You can see all clients (26:29), how many tokens they use, which model they chose, Astra or 5.6 (26:35), and incidents (26:45).

The dashboard highlights in orange on high token spend (26:55). The most important field is incidents: the client Sophie's Orgo disk is almost full, and once it fills up the agent stops working (27:01).

Analysis: the dashboard is not reporting analytics but a response console. It catches incidents before the client notices them. That is what makes the subscription meaningful: the client pays not for code but for someone watching.

29. The observability loop with auto-repair via Jeff

Carson's key innovation is not just the dashboard but auto-repair. Jeff, who does the setup, sees all incidents and can fix them (27:15). He has a list of problems he has already solved, so he knows what to do, and either finds a solution or messages Carson on Telegram (27:24).

30. The task board: backlog, to-do, busy, done, including agent work

Besides the dashboard there is a task board. Carson or the co-founder create a task (27:41), and the columns are backlog, to-do, busy and done, with work done by the AI agent marked separately (27:50).

Incidents from the dashboard show there too: Carson uses too many tokens, and the Orgo disk is overflowing (28:01). You can click an incident and Jeff will find a solution (28:12). That is what 'fix it' means (28:17).

Analysis: the board is a single point where manual tasks and incidents meet, and where one click triggers auto-repair. It is an operating system for the service, not just a to-do list.

31. What fills the disk: git repositories and chat history, expanding to 15 GB

The host notes: the disk size in Orgo can be increased by raising a slider (28:20). Carson adds that space is taken by GitHub repositories, where people save all their files, and by chat history (28:29).

The solution is simple for now: Carson tells Jeff that the disk is full, and he brings it back to 15 gigabytes (28:47).

32. Client notifications: an email when Orgo or OpenAI goes down

The Orgo status is connected to the dashboard, probably via API, Carson does not remember exactly (29:05). If something is wrong with Orgo they get an email, and the client also gets an email when Orgo or OpenAI stop working (29:12).

The client notification text: sorry, we have a problem with this or that, we will send a new email once it is fixed (29:18).

The host calls the whole setup elaborate and admits he draws inspiration for his own internal dashboards (29:26). He says Carson has plenty of observability, alerts and watchdog mechanisms for failures, and that Carson is clearly preparing to scale (29:41).

33. The golden image: an Orgo template, then Jeff assembles everything

The host asks about the golden image, the Orgo template. Carson answers that they used the Orgo template at the start (30:26) and admits it would be worth digging deeper (30:31). But now there is the Jeff agent, living both on the desktop in Hermes and in Orgo (30:36).

Only two actions are required from Carson: make the OpenAI connection and hand over the workspace API from Honcho (30:52).

Analysis: the golden image is not a disk image but Jeff himself. The template as an artifact turned out less important than the agent that deploys the infrastructure. The two manual inputs, OpenAI and Honcho, are the boundary of automation.

34. Pre-built agents by profession and the minimal viable agent (MVA)

Carson then describes the next step: pre-assembled agents for a specific profession. He uses Belgian bookkeeping as an example. In Belgium every transaction must be linked to an invoice (21:00); an electrician has many, because they buy a lot, and every three months it must be filed with a government body (21:11).

Even for Carson himself this is a full working weekend, because you have to sit at the computer and link everything by hand (21:19). So they are building an agent connected to the bookkeeping system that does this every day and asks questions when something is unclear (21:30).

Bottom line: they are looking toward pre-assembled agents that can be sold to electricians or similar companies, a small agent for a very specific task (21:45). The host calls it a minimal viable agent and proposes the term MVA (22:01).

35. Clients: architects, electricians, large companies and a restaurant in Antwerp

Asked who his client is, Carson answers in one word: everything (31:31). Then he lists: architects, electricians, large companies (31:32).

And a separate big case: a very famous restaurant in Antwerp is starting to work with them (31:41). A dashboard connected to the agents is being built for it, to see revenue across the different restaurants and the people who work for them (31:45).

Analysis: the client base is broad, from solo owners to corporate groups. The product is not tied to one industry, and the restaurant case shows movement toward dashboards and larger deals.

36. Lead generation: content about Claude and ChatGPT, not about agents

Carson's lead generation is counter-intuitive. He still makes videos about Claude and ChatGPT, not about agents, sells nothing on Instagram and shows nothing (25:23).

And people still ask about agents (25:33). Then he answers that they can work together and gives a calendar link for a free call (25:39). The flow: two or three meetings a day, about ten in the pipeline (25:50).

At the same time Carson does not want to grow too fast, because they are still working on the support system (25:56).

37. Positioning as the only one in the country and media plans

Carson explains why he is not rushing: everything is nearly ready to grow bigger and attract media attention, because nobody in Belgium is doing this (26:06). He wants to do it right and at the right pace (26:13).

Analysis: the positioning is built on uniqueness within the country. While there is no competition, Carson has the luxury of choosing his pace. But the advantage is fragile: once the market is noticed, the window closes.

38. Unit economics of one client: 2750 plus 250 a month against costs

Let us lay out the unit economics from the transcript data and note what is missing. Revenue: 2750 one-time and 250 a month (08:07, 08:15). Costs: OpenAI at 20 in the retainer (08:15), client tokens (26:35), Honcho with no per-workspace breakdown (12:52), Latitude free up to 100 thousand credits (13:10), setup time two or three hours (10:40).

The transcript does not say how many clients are needed for a target income. Our arithmetic: 100 active clients give 25000 euros a month in retainers plus one-time setups. Carson does not name this figure. What is missing: token cost, Honcho cost per workspace, churn, acquisition cost.

39. Risks of the model

Let us break down the risks that follow from the transcript's facts. First: the manual setup ceiling, Carson names the limit as two setups a day per person (16:36). Second: a small retainer, 250 against 2750 one-time, and clients drop off once they start building themselves (16:57).

Third risk: vendor dependency. The stack rests on Orgo, OpenAI, Composio, Honcho, Latitude, and an Orgo or OpenAI outage already requires emailing clients (29:12). Fourth: the agent's access to client data, it is connected to the CRM, bookkeeping and email (09:23, 18:46). Fifth: the agent as its own DevOps, and if Jeff himself breaks, manual labor and Telegram escalation remain (27:35).

40. Comparison table: their stack versus ours

Let us compare Carson's stack with what we have. We have a terminal, browser, memory, skills, a scheduler, observability via cron and watchdog, boards and an MCP server. What we lack: cloud machines per client, ready connectors, a separate memory layer and external observability as a service.

41. What carries over to us and what does not

Some of Carson's techniques carry over directly, others are tied to his local context and do not. Let us divide honestly.

42. Our candidates for a minimal agent per niche

Let us apply Carson's MVA logic to our tasks. The requirement for a candidate: one specific, repeatable pain that the agent closes and that can be sold to many. Four candidates stand out from our directions.

Analysis: Carson's closest analog is the pairing of compliance and reporting, because there is a hard deadline and repeatability. Lead generation is closer to scraping and API, the reader is closer to the briefing. All four fit the formula 'one pain, one agent, low entry price, retainer on top'.

43. A 30-day plan

Three concrete steps, each with an owner and an artifact. They follow from Carson's techniques that carry over to us and test the model on our ground.

44. Open questions

Let us gather what the transcript does not answer but that is critical for transferring the model.

• Retention. Carson admits churn after a month (16:57), but gives neither a rate nor a customer lifetime.
• Token cost per client. It shows in the dashboard (26:35), but the transcript gives no specific amounts.
• Honcho cost per workspace. Carson says outright that there is no breakdown (12:52).
• Speed. Slowness remains an open problem, and the promise to make Hermes lighter is so far just a statement (24:42).
• Who owns the client and the data. The agent is connected to the CRM, bookkeeping and email (09:23), but there is no discussion of data ownership or the client relationship.

45. Verdict and our pricing hypothesis

Carson's model is not a technological breakthrough but the packaging of a ready stack (Orgo, Hermes, Composio, Honcho, Latitude) into a service with support, sold in a warm market with no competition (00:12, 13:35). The real innovations are the overnight task as a first impression (08:51) and the Jeff agent, which builds and repairs the infrastructure of other agents (27:15, 31:07).

The weak spots are honest too: the retainer is small, the manual labor ceiling is close, retention is fragile, and the design depends on vendors. This is a business built on speed and reputation, not on protected technology.

Our pricing hypothesis: replicate Carson's structure but strengthen the weak link. Bet on a higher retainer with stronger support value and a product format for a niche (MVA) instead of one-off builds. Setup stays comparable to 2750, and the retainer goes above 250 thanks to measurable returns: Carson has a low retainer and churn, a sign of underpriced support.