On 31 July 2026, OpenAI announced that ChatGPT had passed one billion active users. Some of your clients already use it to rough out a question about a retirement plan, a life-insurance contract or the tax treatment of an investment.
Behind ChatGPT, but also behind Claude, Gemini and Mistral, sits the same technical object: an LLM, a large language model. The term is everywhere and almost never explained.
Understanding what an LLM is isn't an engineer's curiosity. It is what lets you know when to rely on it in your practice, and when to be wary of it.
So what is an LLM, exactly, and why can it produce an answer that is perfectly fluent and perfectly wrong?
In short. An LLM is a program that predicts the most probable next word from a piece of text. It generates the most plausible answer, not the truest one. That is why it excels at rewording, summarising or simplifying, yet can state an error with the same confidence as a fact.
Contents
What is an LLM, concretely?
An LLM is a statistical model that predicts the next word in a sentence.
The acronym stands for large language model. Its principle fits in one image: it is autocomplete pushed to the extreme. You give it the start of a text, and it computes the most probable continuation, one word at a time, based on everything it read during training.
It does not "understand" in the human sense. It does not reason about true and false. It evaluates probabilities over sequences of words, split into units called tokens, and returns the most likely continuation.
Its fluency comes from there: it has seen so many well-formed sentences that it produces new ones that look coherent.
This mechanism explains both its strength and its weakness. A tool that can imitate any writing style is extraordinarily versatile. But a tool that computes the plausible has, by design, no guarantee of stating the true.
How does an LLM learn during training?
During training, an LLM ingests vast volumes of text and retains statistical regularities, not verified facts.
In practice, it is shown billions of sentences and asked, over and over, to guess the masked or next word. Every time it errs, its internal parameters adjust slightly.
At the end of the process the model has memorised not a tidy knowledge base, but correlations between words: which terms go together, in what order, in what context.
The architecture that makes this learning effective, the Transformer, was described in Google's 2017 research paper "Attention Is All You Need." It is the architecture that powers ChatGPT, Claude, Gemini, and Mistral today. The "T" in ChatGPT.
One point is worth keeping in mind, because it keeps coming back in marketing pitches: a model learns forms of language, not a reference truth. The difference between training a model, tuning it for a domain and connecting it to verified sources is covered in a dedicated article on what training, fine-tuning and RAG actually mean.
Why does an LLM generate an answer instead of looking for the truth?
An LLM produces the most probable answer; it does not consult a base of truths.
This is a difference in nature from the tools a wealth advisor is used to. A simulator applies a calculation rule. A regulatory database returns an existing text. An LLM, by contrast, builds its answer on the fly by chaining the most plausible words. It has no internal notion of a source, and no mechanism that checks whether what it writes matches a fact.
That does not mean it is always wrong. On heavily documented, stable topics, the most probable answer is often the correct one. But on a precise, dated or technical point, such as a tax ceiling or a rule that has just changed, nothing guarantees that the plausible matches the exact.
An LLM can be connected to a source to look information up at the moment of answering rather than generating it from memory. That is a distinct mechanism, which we will detail, and it changes a great deal for advice.
[[SIMULATEUR]]
What is an LLM good at?
An LLM excels at language tasks where factual accuracy is not the critical point: rewording, summarising, structuring, translating, simplifying.
These are real, legitimate uses in a practice. Turning meeting notes into a readable summary, condensing a long document, rewriting a technical explanation in plain language for a client, drafting a first version of an email: on these tasks the tool saves measurable time. Studies run by OpenAI in 2025 point to 40 to 60 minutes saved per day on office tasks among regular users.
The rule that must frame these uses is simple: on this ground, the AI prepares and the advisor stays in control. A draft remains a draft until a professional has reviewed and validated it. To compare tools for this kind of use, our overview of the AI tools available to a practice goes into detail.
Caution begins the moment you leave language behind and move into figures, rules and recommendations.
Why can an LLM be credible and wrong at the same time?
Because it is optimised to produce plausible text, not accurate text: a confident tone is never a guarantee of accuracy.
This is the phenomenon known as hallucination. The model phrases a wrong answer with the same fluency as a right one, because nothing in its mechanics tells the two apart. To a reader, a well-turned statement inspires trust, which makes the error all the harder to spot.
Regulators have acknowledged it. France's financial markets authority, the AMF, notes that conversational generative AIs, naming ChatGPT, Mistral and Gemini, are not finance-specialised AIs and do not replace a professional's advice: nothing guarantees the products they suggest are suitable. An MIT-Stanford study from August 2026 goes further and shows that an AI's answer to a wealth question varies with how the question is phrased. We analyse it in the article on why a chatbot is not enough to build advice.
The issue is not theoretical for a practice. According to the AMF's 2025 savings and investment barometer, published in December 2025, 11% of French people already say they use AI before making an investment, a share that rises to 19% among the under-35s. Your clients are asking these questions of an LLM, with or without you.
What an LLM is not
An LLM is neither a database, nor a search engine, nor a piece of software that acts on your behalf.
It is not a database: it does not store exact records you can query, it reconstructs a probable answer. It is not a search engine: by default it does not fetch a page or a verified source, it generates text. And it is not a tool that executes actions inside your software.
This last distinction is the most important one, because the vocabulary keeps the confusion alive. An agent is the layer above the LLM. An LLM produces text, nothing more. An agent is an LLM that can use external tools, whether connected directly or not, and can therefore act rather than merely answer.
In practice it picks the right tool, uses it and chains the steps to complete a task, reading a file, running a simulator, updating the CRM. The LLM is the brain that phrases, the agent adds the hands that execute, and in both cases the advisor validates. That is the whole point of a dedicated article on the AI that acts, the agent connected to the firm's tools.
Which leaves a practical question, once the mechanics are clear: between ChatGPT, Claude, Gemini and Mistral, which AI should you choose as a wealth advisor?
Conclusion
An LLM is a machine for producing plausible language. That is what makes it valuable for rewording, summarising or simplifying, and risky as soon as a figure, a tax rule or a recommendation is involved. Tone says nothing about accuracy, and the tool has, on its own, no way to verify what it states. It must be shifted from a probabilistic to a deterministic approach by relying on specific tools.
The real question is therefore not whether to use AI, since the shift is already here. It is which of these tools to use, and for exactly which tasks. That is the starting point of the next article in the series.
FAQ
Is an LLM the same as AI?
No. Artificial intelligence is a broad field. An LLM is one specific category: a model specialised in language that predicts and generates text. All LLMs are AI, but not all AI are LLMs.
Is ChatGPT an LLM?
ChatGPT is an interface that gives access to an LLM. The model is the engine; ChatGPT is the application you talk to. One model can be offered through several interfaces, and one interface can draw on several models.
Can an LLM be wrong?
Yes, and often with confidence. Because it generates the most probable answer without checking a source, it can produce a false but credible statement, called a hallucination. On a precise tax or legal point, its answer must always be verified by a professional.
Does an LLM know my client data?
No, not by default. An LLM only knows what it was trained on and what you give it in the conversation. What happens to the data you enter depends on the tool and its hosting, a topic to settle before any professional use.
Can you trust an LLM for a tax calculation?
No, not on its own. An LLM generates a plausible result, it does not run a controlled calculation. For a tax figure you need a tool that applies a deterministic rule and cites its source, not text generated on the fly.
Sources
OpenAI, audience communications: more than one billion active users announced on 31 July 2026; 900 million weekly active users in February 2026 (reported by Reuters).
Autorité des marchés financiers (AMF), "Using artificial intelligence to invest: what to watch out for", amf-france.org (accessed September 2026).
Autorité des marchés financiers (AMF), 2025 Savings and Investment Barometer, December 2025.
Ashish Vaswani et al., "Attention Is All You Need", 2017 (founding paper of the Transformer architecture).
Choukhmane, de Silva, Lin, Akuzawa (MIT Sloan, Stanford GSB, NBER), working paper on AI-generated financial advice, August 2026.