AI already works. The challenge is operating it.
Between asking ChatGPT a question and having AI run part of your operations there's a chasm. This series explains what's in between.
You use ChatGPT. Your competitors use ChatGPT. Over 60% of freelancers already use some kind of AI tool. The models — GPT-4, Claude, Gemini — reason at levels that would have seemed like science fiction three years ago.
We're not debating whether AI works. It works. That debate is over.
The question is different: can it work every day, doing real work, in your business?
Because there's a huge gap between asking ChatGPT something and having AI run part of your operations without you pushing it. Between a one-off query and a system that records invoices, matches payments and prepares documentation — correctly, day after day.
That gap is what almost nobody talks about. And it's what this series is about.
The chasm between using and operating
If you search "invoice n8n chatgpt" you'll find hundreds of ready-made workflows to connect your email to a spreadsheet via GPT-4. "AI-Powered Automated Invoice Processing." The idea is good — genuinely. There are people doing very interesting things with these tools.
But there's a huge distance between a demo that works with a clean sample invoice and a system that works every day with the real invoices from your business. The blurry ones, the incomplete ones, the ones that arrive via WhatsApp at seven in the morning from the passenger seat of a van.
To understand that distance, it's worth looking at the only place where AI already operates with real autonomy: software development.
In software, AI has covered enormous ground in barely two years. It started by replacing Stack Overflow searches — copy a question, paste the answer. Then it moved to pasting code snippets into ChatGPT for corrections. Then came intelligent autocomplete: Copilot suggesting the next line as you type. Then you'd ask for entire functions and it would generate them. And now there are agents — Claude Code, Codex — that operate on their own: they read your project, decide what to do, write code, run tests, fix errors, and repeat. For hours. Without intervention.
Two years. From "better search engine" to "autonomous agent."
How was that possible? Because the world of software had spent decades building — without knowing it — exactly the conditions AI needs to operate: automatic verification (tests, compilers), everything digital and accessible (code is text in a repository), native environment (AI lives inside the tool the programmer already uses), and many valid solutions (it doesn't need the answer, just one that works).
It wasn't the model. GPT-4 and Claude are the same for everyone. What made the leap possible was the ecosystem that already existed around them.
And now the obvious question: does your business have any of that?
What you discover when you try to make it actually work
Probably not. And when you try to get AI to operate in a real business, you discover why each of those conditions matters.
Who tells you if you got it right? A programmer writes code and the test says "this fails on line 42." The AI fixes it and starts again — that's why an agent can run for hours on its own. In your business, the AI records an invoice and... who tells it the VAT is wrong? Who tells it that supplier doesn't belong to this company? There's no test. No compiler. That has to be built.
Does it know anything about you? Cursor indexes your entire repository. The AI sees your code, your patterns, your architecture — and gives answers that fit your project. Without that, the same model performs 70% worse. In your business, without knowing who your suppliers are, what your tax ID is, what your spending patterns look like, the AI gives generic article-style advice. The difference between useful AI and AI that does real work comes down to one thing: whether it has access to your data or not.
Is it where you work? In software, the AI sits inside the IDE — the tool the programmer already uses. No friction. In your business, your day-to-day is WhatsApp and the phone. AI lives in a browser. If it's not where you work, it doesn't exist for you. 82% of micro-enterprises say AI "doesn't apply to their business." It's not that it doesn't apply — it's that they can't see it where they operate.
Does it interpret or calculate? Code either compiles or it doesn't. 2+2 is 4, always. With an invoice, you feed it the PDF and it reads it perfectly: it tells you the VAT, the withholding tax, the supplier. But when it records it, the total is wrong. Because reading "15% withholding" is one thing and calculating base + VAT - withholding = total is something else entirely. AI interprets. It doesn't calculate. And on an invoice, a single cent matters.
Does it stay on track? Software agents work because every step is verified. If they drift, the test pulls them back on track. Without that safety net, AI starts a conversation well, with a clear focus, but as it goes on it loses context, accumulates irrelevant information, drifts away from the instructions you gave it. It's not a bug — it's the nature of generative systems.
If any of this sounds familiar, it's because managing AI is quite similar to managing people. A new employee also arrives without context. Also needs you to explain how your business works. Also drifts if you don't give them feedback. Also interprets things in their own way.
People are more probabilistic than deterministic — and we've spent centuries learning to work with that. With AI the challenge is similar: not eliminating uncertainty, but learning to manage it.
Where we're coming from
We've walked that software path ourselves. We're a technology company that uses AI operationally every day. We've lived the transition from Stack Overflow to autonomous agents from the inside. We know what building blocks are needed because we use them — and we're using them right now.
And we're trying to build those same building blocks for small businesses. Verification, context, infrastructure where the user already works. Real companies, real invoices, real accountants. The problems we've described above aren't things we've read about — we've lived them. One of our agents created a supplier called "NIF/CIF: not visible" because it interpreted a label from a bad scan as the company name. Another recorded an invoice under the wrong company because it didn't have enough context to distinguish two businesses in the same PDF.
We're not telling these as anecdotes. We're telling them because each of those failures points to a challenge that anyone trying to operate AI in a business will encounter.
What you'll find here
This series will explain, one by one, the real challenges of bringing AI to a small business. Each article covers one concept — technical, yes, but explained for someone who runs a business, not someone who writes code.
What you won't find: hype, miracle tool lists, or promises to automate your business in a day.
What you will find: real problems, real decisions, and what we're learning about how to solve them. Written from the experience of people who are building this and using it every day.
We believe that understanding how to operate AI will be a fundamental skill in the coming years — as important as learning to use a computer was in its time. But this time the tool doesn't always do the same thing. Sometimes it drifts. Sometimes it interprets instead of calculating. Sometimes it doesn't know things it should know.
The businesses that learn to work with that first will have an edge.