Frequently Asked Questions

An LLM (large language model) is, in plain terms, a program trained on a huge amount of text that has learned to predict which word comes after another. It sounds simple, almost silly. And yet that’s where ChatGPT, Claude, Gemini and the whole wave of tools that write you an email or summarize a contract in two seconds come from.

Think of it as your phone’s autocomplete, but trained on an entire library instead of your last few messages. The difference in scale is so big that the result stops looking like autocomplete and starts looking like conversation.

Why now and not ten years ago? Because three things had to come together at once: huge amounts of available text data, computers powerful enough to process it (the famous GPUs, which until recently only mattered to gamers), and a change in the model’s internal design, the so-called “transformer,” which came out of a Google paper in 2017. Without all three pieces together, this would still be a lab experiment.

One thing people assume that isn’t true: the model doesn’t “know” anything the way a person does. It doesn’t look anything up in a database when you ask it something — it generates the answer word by word based on learned patterns. That’s why it sometimes makes things up with total confidence (what’s known as a hallucination), and why it’s best treated not as an oracle but as a tool whose work needs checking.

For a small business this translates into very concrete uses: answering common customer questions, drafting first versions of texts, summarizing long documents. Nothing like science fiction — just a very fast, slightly absent-minded employee whose work you need to review.

If you’re wondering whether it’s worth looking into for your business, the question I’d ask isn’t “what can AI do,” it’s “which repetitive task eats up the most time every week.” That’s usually where the answer is.

→ Direct link to this question

I mostly get asked this by clients who are wondering whether they really need to hire an agency or whether they can “just do it with AI.” I get it — anyone who’s watched ChatGPT write a function in seconds has asked themselves that question.

Generative AI writes code reasonably well when the problem is well-defined and looks like something that already exists thousands of times on the internet: a contact form, a function that sorts a list, a typical component. That’s where it genuinely saves time, and at the agency we use it constantly for exactly that.

What it doesn’t do well yet is understand the full context of a real project: why that API returns data in such a strange format, what decision from two years ago still constrains how a given component can be touched, or what’s going to happen to performance when that client goes from 50 to 5,000 users. That requires someone who understands the business behind the code, not just the code.

There’s an example I use a lot: asking an AI to “fix a bug” in a large project without giving it any more context than the error message. Sometimes it gets it right. Many times it proposes a fix that makes the symptom disappear but leaves the real problem intact, or worse, introduces a new problem somewhere else in the code that nobody reviewed.

So, does it replace a programmer? Not the programmer who knows what they’re doing. It can replace, over time, the most mechanical and repetitive part of the work, and in fact it’s already doing so, which changes which skills are worth having. A developer who knows how to review, fix and decide which AI-generated code is worth using is worth more now than three years ago. One who just copies and pastes without understanding, isn’t.

If your business is considering doing without developers because of this, the question I’d ask is who’s going to take responsibility when something breaks in production on a Saturday night. That part, for now, is still human.

→ Direct link to this question

Before getting into this topic I had to learn it myself, because I had no idea what actually happened to what you write in an AI chat. The short answer is: it depends, and most people don’t read the part that it depends on.

When you type something into a free, consumer-facing AI tool, in many cases that conversation can be used to keep training the model, unless the tool itself offers (and you turn on) an option to prevent it. In business plans or in the APIs we use to build things for clients, it’s usually the opposite: by default it isn’t used to train anything, it just stays as part of the service you requested.

The rule I apply, no exceptions: never paste into an AI chat anything you wouldn’t paste into an email to a stranger. Passwords, customer data, contracts with confidentiality clauses — all of that stays out. Not because I assume bad faith from the company that makes the AI, but because any system can have a security flaw, and the less sensitive data circulating, the less there is to lose if something fails.

For a business that wants to add AI to its website or customer service, there’s a question almost nobody asks and should: where does the data live while it’s being processed? If it’s a tool with servers in the European Union and a data processing agreement (which the GDPR requires), things change quite a bit compared to copying and pasting customer data into the public-facing website of just any chatbot.

And then there’s the obvious thing we forget: if your own company doesn’t allow sharing certain information with any external provider, you shouldn’t share it with an AI either just because it looks like “just a tool.” It’s one more external provider, with its own terms of service and fine print like any other.

What I learned researching this is that the real risk usually isn’t in the AI itself, but in not reading which plan you’re using and what it’s configured for.

→ Direct link to this question