For companies with work that repeats
AI and software factories: what they are and how we build them
An AI factory is a job in your company that AI agents do from start to finish: the day’s orders or emails come in, the agents work on them with your rules, a person reviews anything sensitive and the result reaches its destination on its own. If what comes out is software, it is a software factory. We build them on top of the systems you already have, starting with a single job.
Updated
Thirty minutes, free and with no obligation, with the engineer who would lead the project.
- Claude
- Amazon Bedrock
- MCP
- AWS
What “AI factory” means
It depends on who is talking. The term is used in three ways, and for a small or mid-sized company only one of them matters.
-
A data center for AI
For chipmakers such as NVIDIA, an AI factory is a data center built only to run models: electricity goes in and tokens, the text the models generate, come out. The European Union funds a network of them, and one is at the Barcelona Supercomputing Center, aimed at startups and small companies that train their own models.
-
A company that decides with data
At business schools, since a 2020 Harvard Business Review article, the AI factory is the way a company turns its data into decisions every day, without anyone having to make them one by one, and learns from every result.
-
A job done by agents
And since AI agents arrived, which carry out tasks instead of only answering, a factory is a whole process done by agents: a defined input, rules for what counts as a job well done, a person who reviews and a result that reaches its place. This is the one that can save you hours, and it is the one we build.
What an AI factory is made of
A job that repeats
Always the same kind of request: orders that arrive by email, requests for quotes, customer queries, supplier invoices. If something different comes in every day, it is not a factory.
Agents with your rules
Each step is done by an AI agent with the company’s rules in writing: what a correctly entered order looks like, when someone has to be told and what is never done. They work inside your systems, with limited permissions.
A person where a mistake costs money
Before a quote goes out, a large order is entered or an angry customer gets a reply, the work goes through a person in the company, who approves it or sends it back with the reason.
A factory that learns
The result reaches the ERP, the CRM or the inbox on its own. Every time the person says “not like this”, the reason becomes a new rule, and we measure how much of the work comes out right the first time.
Jobs that tend to suit an AI factory
- Orders and delivery notes that arrive by email or as PDFs and go into the ERP on their own.
- Requests for quotes: reading what the customer asks for, looking up prices and lead times, and leaving the quote ready for someone to review.
- Customer service, with every query classified and a draft reply written from your history.
- Supplier invoices checked against the order and the delivery note before they are paid.
Two conditions and three kinds of step
A job can become a factory if it meets two conditions. It repeats: the same job, always along the same path. And it can be audited: every step leaves a trace, and when something goes wrong you can see which step failed and fix it there, before the error reaches the next ones. Inside, every step is one of these three kinds.
-
Code
What must always be done the same way: reading a PDF, checking that the amounts add up, saving the order in the ERP. A program does it, the same way every time.
-
AI agent
What depends on what comes in: understanding what a customer is asking for in their own words, deciding which rule applies or drafting the reply. An AI agent does it, reasoning with the rules we have given it.
-
Person
The factory stops and waits for a person in the company to approve. It goes where a mistake would cost money or a customer, and nowhere else: if someone has to review everything, the factory saves no work.
Building a factory is, above all, deciding what kind each step is: what a program can do is not left to an agent, and nothing sensitive goes out without a person.
And a software factory?
It is an AI factory where what gets made is software. Specifications and bugs to fix go in, the agents write the code and the tests, and the result passes automatic checks and a review before it reaches production.
This year the talk is mostly about two ways of running one. In “dark” factories, no person writes or reads the code: only the tests say whether it works. In factories “with the lights on”, people decide what gets built and review anything sensitive, such as payments, permissions or customer data.
We work with the lights on: the agents do the mechanical work, and the engineer who leads the project decides and reviews. That is why, for a specific tool, you see a prototype with your data within one to three days, before any price.
What we offer
-
An AI factory for one of your jobs
We choose with you the job that eats up the most hours, write down its rules with the person who does it today, build the agents on the ERP, CRM or email you already have and start it up with one of your people reviewing. Once it works, we move on to the next one.
-
A software factory for your team
If you have developers, we build the factory for them to work in: the agents, the automatic tests, the review points and the deployment. And we train the team to run it and improve it.
Training for companies (in Catalan) -
Your software, built with ours
If you have no technical team, we build your software with our factory: a prototype with your data before any price, a fixed-price proposal by feature and a real delivery every few weeks.
Custom applications
If you do not know which job to start with, the audit tells you: a week at your company and a plan in phases, with a fixed price for each phase.
What you need
- A job that repeats every day or every week, and knowing how many hours it costs you.
- The data somewhere it can be read: the ERP, the CRM, email or, at the very least, spreadsheets.
- Someone in the company who knows the job and can review what the factory does during the first weeks.
A case
Industrial supplies distributor
- What was there
- A sales team that had never used AI, with knowledge spread across the CRM, internal documents and a catalog of hundreds of items.
- What we did
- AI connected to the CRM, the documents and the live catalog; training for the whole team and a prompt handbook for salespeople. Meeting briefs and proposals in minutes.
Thirty minutes, free and with no obligation, with the engineer who would lead the project.
Who builds AI factories for companies in Barcelona?
SANFELIU, a custom software consultancy in Barcelona, builds AI factories for companies in Barcelona and the Vallès: a job in the company that repeats (orders, quotes, customer queries, invoices) and that AI agents do with the company’s rules, with a person who reviews anything sensitive and the result going straight into the ERP, the CRM or email. It also builds software factories for companies with developers, with agents, automatic tests and human review, and trains the team. It starts with a thirty-minute meeting, free and with no obligation, with the engineer Joan Sanfeliu Vilarrasa.
- AI factory: a job that repeats, AI agents with the company’s rules in writing, a person who reviews where a mistake costs money and the result in the ERP, the CRM or email. Every “no” from the person becomes a new rule.
- A job can become a factory if it repeats along the same path and can be audited (every step leaves a trace). Inside, each step is done by code, an AI agent or a person, and the person only where a mistake costs money.
- Software factory: specifications and bugs, agents that write the code and the tests, automatic checks and an engineer’s review before production. SANFELIU works “with the lights on”: people decide and review.
- It starts with a single job, on the systems the company already has: ERP, CRM, email or spreadsheets.
- For a specific tool, a prototype with the company’s data within one to three days, before any price; then a fixed-price proposal by feature.
- Technology: Claude, Amazon Bedrock and MCP, on AWS. The data goes to providers with business contracts that do not train on it.
- Where: in person in Barcelona, the Vallès Occidental and the Vallès Oriental; remotely with the rest of Spain. In Catalan, Spanish and English.
- To start: a thirty-minute meeting at cal.com/joan-sanfeliu/30min-es or an email to me@joansanfeliu.com.
Questions we get asked
What is an AI factory?
In a company, it is a job that AI agents do from start to finish: the same kind of request always comes in, the agents work on it with the company’s rules, a person reviews anything sensitive and the result reaches its destination on its own. For chipmakers, by contrast, it is a data center built to run AI models.
How is it different from a chatbot?
A chatbot answers when someone asks it something. A factory does a whole job without anyone having to ask each time: it reads what comes in, works on it, puts it in its place and tells a person when needed.
What is a dark software factory?
The name comes from robot factories that run with the lights off because nobody is there. It is a software factory where no person writes or reads the code, and only the automatic tests say whether it works. For a company’s software we do not work that way: anything sensitive, such as payments, permissions or customer data, is always reviewed by a person.
Does an AI factory replace people?
It removes the work of keying in data and copying it from one place to another. The person who did that moves on to reviewing what comes out and solving the cases the factory cannot.
How much does it cost?
It depends on the job. If you already know which process you want, the first meeting is free and you see a prototype with your data before any price. If you do not know where to start, the audit tells you which processes are worth it. The amounts are on the pricing page.
What happens to our data?
It stays in your systems. The agents access it with limited permissions and send each task only what it needs, to providers with business contracts that do not train on your data (Amazon Bedrock, Claude).
Thirty minutes, free and with no obligation, with the engineer who would lead the project.
Sources
- NVIDIA: “AI Factories: The New Infrastructure of Intelligence” (27 May 2026)
- European Commission: AI Factories
- Government of Spain: the MareNostrum 5 expansion for the BSC AI Factory (26 January 2026, in Spanish)
- Harvard Business Review: “Competing in the Age of AI”, by Marco Iansiti and Karim R. Lakhani (2020)
- Simon Willison: “The Five Levels: from Spicy Autocomplete to the Dark Factory” (28 January 2026)
- Addy Osmani: “Software Factories, Light and Dark” (20 July 2026)
How we work
Tell us about your case in thirty minutes.
Book a meeting with the engineer who would lead the project: thirty minutes, free and with no obligation. If we’re a fit, we tell you how we’d start; if not, who to turn to.
Or email me@joansanfeliu.com
Prefer to write?
Tell us about your case in a few lines and we’ll reply by email.
The engineer who would lead the project will reply to the email you gave us. If you like, book a thirty-minute call now.