How a company with three people could carry ten million in revenue

A calculation, not a case study

Ten million euros in annual revenue. Three people. For anyone who learned to run a company in the 20th century, that calculation does not work. More revenue needs more hands. That was true for a long time.

One thing first, so no wrong impression sticks: we have no client list of such companies and no study on them. What we have are projects in which individual areas are built the way this article describes. A support desk that answers most requests without a human. A bookkeeping process that assigns receipts on its own. This article puts building blocks like these together into a model and works out what might come of it. Not as a promise, but as a thinking aid.

The revenue is the least important number here. The interesting question is which work in a company has to be done by people at all and which does not.

Why companies are as large as they are

The economist Ronald Coase described in 1937 why companies exist. Buying every single task on the market is too expensive. Every purchase costs search, negotiation, coordination and control. So we bundle work into organisations. The higher those costs, the larger the sensible company.

The reverse holds too. When coordination and handover get cheaper, a smaller company can carry the same work.

That is exactly where software comes in. In this article an agent is a program that uses a language model to handle one bounded task from start to finish. Read a request, fetch the matching answer from the knowledge base and send it, for example. It needs no onboarding, no desk and no handover to a colleague. It runs when a task exists and not otherwise.

To be honest, AI has not abolished coordination costs. It has lowered them for part of the work: the part that can be standardised and whose result can be checked. For everything else the costs remain. That is why three humans are left in this model and not zero.

What changed between 2022 and 2026

In 2022, growth meant new positions. Anyone who wanted to double sales hired people, waited months for onboarding and started over when someone left. Growth was tied to people and people are slow to find and expensive to keep.

In 2026, growth in standardised tasks can mean: more calls to the language model. The cost of those calls has fallen sharply since 2023 and it grows with volume, not in steps like headcount. An agent does not resign and does not get sick.

But it also does not get better on its own. That is a common misconception. Every new model version behaves differently and has to be tested against your real cases before it goes live. More on why automation only holds with that discipline in The Great Filter 2025.

How the work is split across four areas

The model has no departments. It has four areas in which agents carry most of the routine. What the agent takes over and what stays with a human looks different in each one.

Area	The agent takes over	Stays with a human
Sales	Researching target companies, pre-qualifying requests, keeping the CRM current, follow-ups	The conversation with the customer, the price, the close
Fulfilment	Onboarding emails, collecting documents, executing standard deliverables, checking results against criteria	Special cases, complaints, anything without a clear check
Bookkeeping	Creating and sending invoices, matching incoming payments, assigning receipts, reminder stages	Approval, the accountant, decisions on unclear entries
Support	First answer from the knowledge base, diagnosis from logs and old tickets, churn signals	Upset customers, goodwill, anything with legal effect

Sales

The sales agent is not a single bot but a chain. One step searches for companies that match the target customer. One step reads the company’s website and prepares an approach. One step holds the first written conversation and collects budget and timeline. One step keeps the CRM current and flags deals where nothing is moving.

Two limits belong here. Cold email outreach is legally restricted in Germany and an agent does not change that. And the reply rate depends on how well the agent really understands the company. What works in projects is in our analysis of sales automation.

Fulfilment

After the close, fulfilment takes over. For an agency that means creating content, setting up campaigns, writing reports. For a shop: processing orders, shipping, returns. A checking step compares every result with defined criteria. Whatever fails or has no criterion goes to a human.

That last step is the most important one in the whole model. An agent without a check produces errors in series on a bad day and nobody notices.

Bookkeeping

Create invoices from contract data and send them on the due date. Match incoming payments with open items. Assign receipts to accounts. Reminders in stages. Prepare the monthly close for the accountant. All of this is standardisable and easy to check, which is why bookkeeping is usually the area where automation holds fastest. The accountant stays. The agent prepares, it does not decide.

Support

The first answer comes from the company’s knowledge base. A second step searches logs and old tickets for patterns in more complicated cases. A third one spots customers whose behaviour points to churn and puts them in front of a human. Why a simple chatbot with canned answers is not enough for this is in Death of Chatbots.

Which technology runs it

The model does not run on one big platform but on building blocks that can be replaced individually. No per-seat licences, payment by usage.

Language models. Models from Anthropic and OpenAI for tasks that need understanding. Smaller, cheaper models for bulk tasks like sorting and extraction.

Knowledge base. A vector database such as Qdrant on your own server. It stores text in a way that lets you find similar content even when different words are used. Every agent reads from the same knowledge base. What that means for a company’s memory is in The Immortal Company.

Orchestration. n8n for processes that can be described as a chain of steps. Custom code in Python or TypeScript where several agents interact. Everything versioned, tested and monitored. How to run n8n yourself is in our guide to self-hosting n8n.

Servers. Own servers at Hetzner in Germany. Full control over the layer you run yourself. The privacy assessment still depends on what you process and which models you call externally.

Website. Astro with statically generated pages. Fast, low-maintenance, few dependencies. Legacy is a liability, in the frontend too.

Reporting. For every agent: throughput, error rate, cost per transaction. The person controlling operations sees in one view whether the company is running.

The three roles that remain

The model is not a cost-cutting exercise. It is a concentration. The three roles are not interchangeable.

Direction. One person decides on markets, prices and partners and holds the conversations with the important customers. Those are few decisions per week, but the ones the value of the company depends on. More in Radical Focus Culture.

Building. One person builds and maintains the agents. They write the processes, test new model versions, fix errors in the instructions. Their work is not programming in the narrow sense but designing the machine that runs the company.

Control. One person watches the reporting, takes the cases the agents hand over and corrects mistakes. Every correction flows back into the instructions as an example. How often they have to step in depends on the area: rarely in bookkeeping, more often in support.

These three do not work more than others. They work on different things. Their impact is not measured in hours but in how many transactions get better through a single correction.

What the model depends on in practice

This is where the honest list belongs, because most attempts do not fail on technology.

Standardisation. The model holds for service businesses, agencies and shops with recurring processes. In project work where every order is different, the share of human work stays high.

Data quality. An agent answering from an incomplete knowledge base answers confidently and wrongly. Maintaining the knowledge base is an ongoing job and nobody budgets for it at the start.

Checkability. Every area needs a criterion that separates a good result from a bad one. Where there is none, there is no unsupervised agent either.

Concentration risk. Three people means: if one person is out, a third of the company is missing. The person who builds the agents is the hardest to replace.

Law. Where an agent talks to customers, disclosure duties apply. What the AI Act requires is in AI Act Article 50 still applies. And the agent changes with every model version, it does not automatically improve.

Whether you end up with ten million and three people or two million and five depends on your starting point. The direction is the same in both cases: fewer positions for tasks software can do in a checkable way.

What you can check

Do not start with the company. Start with one process.

First: write down the processes where your people read, interpret and forward data. Invoices, requests, orders, applications. Second: take the one with the highest volume and the lowest complexity. Third: measure the current state before you build anything. How long does one transaction take, how many are there, how many go wrong. Without that number you cannot say later whether it paid off. Fourth: build a single agent with a checking step and a human who receives the exceptions. Only when that runs cleanly for a month does the next one follow.

If you want to know which process should be your first, write to us. How we build such processes with n8n is under n8n automation.