Every day, in a company, someone pastes some data into an AI. A list of customers to reorder. A document to summarize. A file to "have read" by ChatGPT or Gemini to extract something.
It's convenient. It works. And almost no one stops to think about where that data ends up.
Because the point isn't whether AI is useful. It is, absolutely. The point is another: not all AI needs you to send it your data.
There is a category of models — small, specialized — that do very useful work while staying on your device. They don't go up to the cloud. They don't "see" anything on a server. One of the most used is called NER. And understanding it helps you reason better about all the AI you use.
What a NER model is
NER stands for Named Entity Recognition. In simple words: it's a model that reads a text and understands what the words are.
Not just "this is a capitalized word," but:
- this is a person
- this is a place
- this is an organization
Take the sentence: "Marco Rossi works in Florence for Ferrari." A NER model tells you: Marco Rossi is a person, Florence is a place, Ferrari is a company. And it does this from context, not from a list.
Why lists aren't enough (and context is)
The naive approach is: "I'll make a list of names and search for those." It doesn't work. And the reason is simple: the world's names don't fit in a list.
There are Italian, Polish, Turkish, Chinese, Spanish names. Surnames you've never seen. Cities you don't know. Companies born yesterday. A list is always incomplete. And filling it with everything creates errors: "Como" is a city, but also a piece of a sentence in Spanish; "Marino" is a surname, but also a town.
NER solves the problem differently: it doesn't memorize words, it recognizes patterns. It has "read" millions of sentences during training and has learned how people, places, and companies behave inside a sentence. That way it recognizes a name it has never seen, in any language, simply because it behaves like a name.
It's the difference between memorizing a vocabulary and learning to read.
How it works, no jargon
A NER model is, essentially, a file: a network of numerical "weights" that represent what the model has learned. When you give it a text:
- it breaks it into pieces (words, and sometimes pieces of words);
- for each piece it calculates how much it "resembles" a person, a place, an organization, also looking at what comes before and after;
- it returns the entities with their label.
That's it. No magic, no consciousness. A pattern recognizer that's very good at one specific task.
And that's exactly its strength: it does one thing, and does it in little space. A NER model weighs a few dozen megabytes. A large conversational model weighs thousands, and for that reason has to live on someone else's servers.
The point that changes everything: it can run locally
A small and specialized model like NER can be executed entirely on your device — inside the browser, on your computer. Concretely this means that:
- the model (its "weights") downloads once, like any app;
- from there on it works locally;
- the text you give it to analyze never leaves your machine.
It's not a promise of trust. It's architecture. There is no server to send the data to, because the computation happens where you are.
With cloud AI, instead, the flow is the opposite: your data leaves, arrives at a server, is processed there. You can trust the vendor's policies, but the data has left. And "left" is a word that, with customer and employee data, weighs a lot.
Why a B2B company should really care
It's not a technical matter. It's a matter of responsibility. Every time a collaborator pastes a file with names, emails, contracts, or personal records into a cloud tool, they are sending personal data out of the company. Often without knowing it.
The GDPR doesn't look at good intentions. It looks at where the data goes. And the paradox is obvious:
- if you want to use AI to work on your data, you're tempted to upload it to an online AI;
- but uploading it is exactly what, for confidentiality and compliance, you shouldn't do.
Local models break this paradox. They give you AI capability without the data transfer. Not on everything — but on many concrete, repetitive tasks, yes.
When it's worth it
It's not for everyone, always. But there are clear signals. It's worth looking at these models when in your company:
- you often work with files full of personal data — customer records, CVs, contracts, CRM exports — and you'd like AI help without sending them out;
- you have to share data with third parties (agencies, consultants, vendors) and you need to "clean it up" first;
- you operate in a sensitive or regulated sector, or the DPO / legal office has already warned about using ChatGPT with company data;
- you have repetitive and well-defined text tasks — extracting names, classifying, tagging, sorting — on significant volumes;
- you want to feed a process with AI, but compliance blocks you on "uploading online."
The common thread: data that should stay in-house, plus a repeatable task. That's where a local model stops being a technical curiosity and becomes a work tool.
A practical signal not to ignore: if someone on the team has already started pasting company files into ChatGPT "because it's convenient," the problem already exists. The question isn't whether it will happen. It's how to give them a safe alternative.
And in practice, how do you do it?
Here the right question is only one: who has to do it? Because there are two levels, very different.
Level 1 — Direct use, no IT required. There are ready-made tools that run the model inside the browser. You open a page, the model downloads once — like when a site loads a font or a library — and from there it works on your computer.
- Do you need to install anything? No.
- Is there an interface? Yes: a normal web page. You upload the text or file, you get the result.
- Do you need to call IT? No, anyone can do it.
- Does the data leave? No: it stays in your browser.
It's the right path for most people and daily tasks.
Level 2 — Integration into systems, with a developer. If instead you want to embed the recognition into your processes (the CRM, a management system, an automatic flow), a developer is needed there. They will use open and free libraries — the best known are spaCy and Hugging Face Transformers — and download the model from Hugging Face, a kind of public repository of AI models, the "GitHub" of artificial intelligence: thousands of ready-made models, free, in many languages.
- Do you need to install anything? Yes, but the technician does it, once.
- Is there an interface? No: it's code that runs inside your systems.
- Does the data leave? No: it runs on your servers or your computers.
In both cases the principle doesn't change: the model comes to you, your data doesn't go to it.
And, at the decision level, the important thing is that it isn't a huge project. A ready NER model downloads for free, weighs little, and for basic use requires neither budget nor infrastructure. The real cost isn't technical: it's deciding to handle confidential data with a tool that doesn't send it out, instead of the usual convenient copy-paste.
When local makes sense and when the cloud does
It's not a war. It's a task-by-task choice.
- Local is best when the task is specific and the data is sensitive: recognize entities, classify, extract, mask, sort. Small, fast models, on your device.
- Cloud is best when you need the generalist power of large models — open reasoning, complex writing, broad summaries — and the data isn't confidential.
The practical rule: the more sensitive the data and the more defined the task, the more local is the right choice.
And it should be said honestly: local has a price. Small models are less "general," they need to be downloaded once and run on your computer's power, so on huge files they are slower. In exchange, your data stays yours.
The question to ask
Useful AI isn't just the one that answers best. It's also the one that doesn't ask you to give up control of your data to work.
A NER model is a small example of a bigger principle: most of the value of artificial intelligence can stay close to you, not on a distant server.
So, next time you're about to paste a file into an AI, one question is worth asking: does this data really need to leave the company to be processed?
Very often, the answer is no. And understanding that, today, is already a competitive advantage.
