There is a lot of hype about Large Language Models (LLMs) and "AI Agents." But if you strip away the Silicon Valley jargon, the architecture of modern AI looks remarkably like a system legal professionals use every single day: the relationship between a Client, an Instructing Solicitor, and a Barrister.

Every lawyer in Australia understands how a brief to counsel works. A client comes in with a problem. The solicitor listens, gathers the facts, applies their knowledge of the law, compiles a brief, and sends it to a barrister. The barrister reads the brief, analyses the issues, and delivers an opinion.

That process — client to solicitor to barrister — is almost a perfect analogy for how modern AI systems work. Let me explain.

The Prompt Is the Client’s Instructions

When a client walks into a solicitor's office and says "I have a contract dispute, I need advice," that is the prompt. The client knows what their problem is, but they don't know the exact legal mechanics required to solve it. They hand you raw, unstructured information. In AI terms, a basic user typing a question into a chatbot is just a client presenting a raw problem.

The quality of the client's instructions matters enormously. A client who can articulate the key facts clearly, identify the relevant documents, and explain what outcome they are seeking gives the solicitor everything they need to work with. A client who gives vague, incomplete, or inaccurate instructions creates problems that flow through the entire process.

The same is true of AI. The prompt — the instruction you give to the AI — determines the quality of what comes back. A precise, well-structured prompt produces precise, useful output. A vague or poorly framed prompt produces vague or misdirected output.

The Solicitor Is the Agent

An AI agent is not the LLM itself. The agent is the layer that sits between the user and the LLM — the system that takes the initial instruction, processes it, gathers additional information, and constructs a complete, organised input before anything is sent to the model.

An AI agent is not just a passive chatbot; it is software designed to do work on your behalf. It takes the client's raw prompt and acts as the Instructing Solicitor.

When a solicitor receives a client's instructions, they do not simply forward those instructions to the barrister unchanged. They:

  • Take the client's account and interrogate it for gaps
  • Apply their own knowledge of the law to frame the issues correctly
  • Gather the relevant documents, pleadings, and evidence
  • Prepare a chronology
  • Identify the parties and witnesses
  • Research applicable case law and legislation
  • Anticipate the counterarguments
  • Draft a cover letter setting out precise questions for counsel to answer

The solicitor takes that initial prompt, uses their own professional knowledge, and compiles a comprehensive, structured Brief to Counsel. What the barrister receives is not the raw client instructions. It is a carefully constructed brief; organised, indexed, and framed to give the barrister exactly what they need to produce a useful opinion.

The AI agent does the same thing. It takes the user's prompt, enriches it with additional context, retrieves relevant information, structures the input, and sends a complete, organised package to the LLM. The LLM then reasons over that package and produces its response.

The Brief Is the Context Window

A barrister can only work with what is in the brief. They cannot go outside the brief to find additional facts. If a critical document was not included, if a key date was wrong, if a witness's statement was omitted, the opinion will reflect those gaps.

An LLM works the same way. It can only reason over what has been placed in front of it, what is called the context window. The context window is everything the model can "see" at the moment it produces its response. It includes the prompt, any documents that have been provided, prior conversation, and any additional information the agent has retrieved and included.

A complete, accurate, well-organised brief produces a reliable opinion. An incomplete or inaccurate brief produces an unreliable one.

The Solicitor’s Legal Knowledge Is the RAG Database

A solicitor does not start each matter from scratch. They bring to every engagement years of accumulated legal knowledge — how contract law works, what the standard clauses mean, which cases are likely to be relevant, what courts in this jurisdiction tend to do with these issues.

This background knowledge is what allows the solicitor to transform a client's raw instructions into a properly framed brief. The solicitor augments what the client has told them with what they already know.

In AI systems, this function is performed by what is called Retrieval-Augmented Generation (RAG) and vector databases. These are repositories of documents, case law, legislation, firm precedents, and practice guides that the agent can search and retrieve from at the moment it is constructing the brief for the LLM.

When the agent receives a prompt about a contract dispute, it searches its databases for relevant material — applicable legislation, similar cases, standard contractual terms — and it incorporates that material into the context it sends to the LLM. The model then reasons over both the user's specific instructions and the retrieved background knowledge together.

The solicitor augments the client's instructions with their legal expertise. The agent augments the prompt with retrieved knowledge. The result in both cases is a richer, more complete input and a more reliable output.

The Barrister Is the LLM

A large language model (LLM) is the engine at the centre of AI tools like ChatGPT, Claude, and Gemini. But before a barrister can read any brief, they must first be trained: years of university, textbooks, historical judgments, statutes, and academic commentary. They absorb the foundational patterns of legal language, legal reasoning, and legal argument before they ever appear in a matter.

An LLM is built the same way. Before it ever sees your prompt, it undergoes massive pre-training on vast datasets of text — legislation, case law, books, articles, contracts, correspondence — across millions of documents. From that training it develops a sophisticated ability to reason, synthesise, and produce written output.

Think of the LLM as a very senior barrister. One who has read everything. Tens of thousands of cases. Every major textbook. Legislation across dozens of jurisdictions. Years of pleadings, judgments, and academic commentary. They sit in chambers, waiting for the brief to arrive.

When you send this barrister a brief, they don't look anything up in the moment. They draw on everything they have already absorbed during their career and their training and they produce an opinion — a view of how a court would likely decide the matter, based on the information they have been given.

When an LLM processes your brief, it does not rely on a single stream of data; rather, it performs a sophisticated dual-lookup, seamlessly blending its broad foundational memory with your specific files. To generate an opinion, the model actively retrieves and reasons over the information it has from its pre-training phase (its embedded legal "education" containing tens of thousands of cases and statutes) in conjunction with the hyper-specific text provided in the prompt and augmented by RAG (the precise client facts and localized firm precedents). It treats the background legal principles absorbed during pre-training as the analytical framework, while treating the prompt and RAG context as the strict evidentiary boundaries of the current case. The model reasons across both layers simultaneously to ensure that its structural legal logic aligns perfectly with the unique, concrete facts you have placed on its desk.

The Knowledge Cutoff

Legal training is not a static event. A barrister who qualified twenty years ago and has not kept up with developments in their practice area cannot advise competently on current law. They will reason fluently and confidently, but from an outdated foundation.

An LLM has the same limitation. Every model has a knowledge cutoff date — a point beyond which it has no awareness of new cases, new legislation, or new developments.

When AI labs (like Anthropic) train a baseline model, they compile a massive snapshot of data up to a specific date (e.g., January 2026). Once training begins, that "book is closed":

The 4.7 Barrister: Understands the world perfectly up to the moment its training data was frozen. It has a rigid "knowledge cutoff." It does not know about any cases filed, laws passed, or news events that happened after that date.
The 4.8 Upgrade: When Anthropic releases an upgrade like Claude Opus 4.8, they haven't just gently fed it a few new articles while it was running. They have gone back to the servers, bundled a fresh, updated massive dataset (including newer court cases and recent data), and run an entirely new, multi-million dollar training cycle to produce a brand-new, distinct "brain".

Why Jurisdiction Matters

Imagine an Australian solicitor preparing a brief for a complex breach of contract dispute in the Federal Court of Australia and mistakenly sending it to an American barrister trained exclusively in US law. The American barrister reads the brief. They understand the English language, the concept of a contract, the layout of a chronology. When they deliver their opinion, they apply US Uniform Commercial Code principles, cite Delaware case law, and reference American procedural rules.

The output is beautifully structured. It reads as authoritative. And it is completely useless — if not dangerously wrong — for a matter being heard in Sydney or Perth.

This is exactly what happens when a lawyer uses an AI model that was not trained predominantly on Australian law, or whose training predates recent legislative changes or significant judgments in the relevant practice area. Raw analytical power means nothing if the underlying training is misaligned with the jurisdiction and the current state of the law.

The Opinion Is Not Infallible — an "Hallucination" Is Just a Poorly Prepared Brief

A barrister's opinion is a prediction. It is a considered, expert view of how the law applies to the facts as presented, but it is not a guarantee of outcome. The opinion is only as good as the brief. If the facts were wrong, if key documents were missing, if the issues were incorrectly framed, the opinion may be wrong too.

Exactly the same caveat applies to LLM output. The model produces its best reasoning over the information it has been given. It does not know what it has not been told. It cannot verify facts independently. It can be confidently wrong.

In the AI world, when a model gives a confidently incorrect answer, the term people use is hallucination. In the legal world, we would simply call it the natural consequence of a poorly prepared brief.

This is why, in AI as in law, the quality of the input is everything:

  • The client's instructions must be accurate and complete.
  • The brief the solicitor compiles must be accurate and complete.
  • The model must be trained on the right jurisdiction and current law.

If any of these conditions are not met, the barrister's opinion may lead the client in the wrong direction.

What This Means for Lawyers Using AI

Lawyers are already expert at the skills that make AI work well. You know how to frame issues precisely. You know how to identify the relevant facts and discard the irrelevant ones. You know how to construct a logical, organised presentation of a problem. You know how to ask the right question.

Those skills — the skills of a good instructing solicitor — are exactly the skills of a good AI user.

The lawyers who will get the most out of AI are not necessarily those who are the most technically sophisticated. They are the ones who bring the most rigour to their instructions, who treat the prompt the way they would treat client instructions to counsel, who understand that the quality of what comes out depends entirely on the quality of what goes in, and who ask the right questions about whether the model they are using is trained on the right law for the matter at hand.

You already know this. You have always known this. You just knew it as briefing counsel.