You’ve almost certainly used one by now, even if you didn’t think of it in these terms. You typed a question into ChatGPT, asked Copilot to draft an email, or watched a colleague use AI to summarise a long document. What you were using, in each case, was a large language model.

This article explains what that actually means, in plain terms, and what it does and doesn’t make sense to use one for if you run a financial advice firm.

What is a large language model?

A large language model (LLM) is a type of AI system trained on an enormous amount of text. The training process, simplified to its essentials, involves the system learning statistical patterns across billions of words: what tends to follow what, how ideas connect, how questions are typically answered. The result is a system that can generate fluent, contextually appropriate text in response to a prompt.

It is not thinking. It is not reasoning in the way a human reasons. It is doing something much more specific: predicting what a useful or accurate continuation of a given input looks like, based on patterns it has absorbed. That distinction matters, and I’ll come back to it.

The “large” in large language model refers to scale. Modern LLMs contain hundreds of billions of parameters (the numerical values that encode the patterns learned during training), and they were trained on text drawn from books, websites, academic papers, code repositories, and other sources running into the trillions of words.

What can an LLM actually do?

In practical terms, an LLM can do several things well:

  • Drafting. Given a brief, it can produce a first draft of an email, a letter, a report, a policy summary, or almost any other document type. The draft will usually need editing, but it gives you something to work from rather than a blank page.
  • Summarising. Feed it a long document and ask for a summary. It will produce one quickly. This is one of the most reliably useful applications in a financial advice context.
  • Answering questions about documents. When you constrain an LLM to a specific set of documents, rather than letting it draw on its general training, it can answer questions about those documents with reasonable accuracy and low rates of invention.
  • Reformatting and restructuring. Converting a table into a narrative, restructuring a set of bullet points into prose, translating something from technical language into plain English.
  • Generating options. Brainstorming subject lines, alternative phrasings, agenda items, or frameworks. The output often isn’t perfect, but it gets you unstuck.

What an LLM cannot reliably do

This is at least as important as the list above.

An LLM does not know what it doesn’t know. When it lacks the information needed to answer a question accurately, it may generate a plausible-sounding answer that is wrong. This is called hallucination, and it is a structural property of these systems, not a bug that will simply be patched away. For most consumer applications it’s a manageable inconvenience. For a regulated firm producing suitability letters, file notes, or compliance documentation, it is a material risk if the output isn’t reviewed by a human before it goes anywhere near a client.

An LLM also does not have current knowledge unless it is explicitly given it. Most models have a training cutoff, after which they don’t know what’s happened. If you ask an LLM about a regulatory development from last month without providing the actual source text, you may get an answer that is outdated or invented.

Finally, an LLM is not a compliance engine. It can help you draft, summarise, and structure. It cannot tell you whether a recommendation is suitable for a specific client. That judgement belongs to a human professional, and nothing in this article or anywhere else changes that.

The smart intern

One framework I find useful is thinking of an LLM as a smart intern [1]: sharp, motivated, fast, and genuinely capable, but with zero knowledge of your business on day one. You wouldn’t ask a new intern to sign off a suitability letter unsupervised. You also wouldn’t refuse to delegate any work to them because they weren’t a chartered planner. You’d brief them carefully, check their output, and give them progressively more responsibility as you understood what they were good at.

That’s roughly the right relationship to have with an LLM. Use it for the tasks where it can genuinely reduce your time or improve your output. Keep a human in the loop on anything regulated or client-facing.

An LLM is a capable first drafter and a useful summariser. It is not a decision-maker, and treating it as one in a regulated context is where firms get into difficulty.

Where LLMs are practically useful for an advice firm

Drafting client communications. An LLM can take a set of bullet points about a client’s situation and produce a draft communication quickly. The adviser edits, personalises, and approves. The gain is in reducing the time from “I know what I need to say” to “there is a document I can work with.”

Summarising long documents. Annual reports, fund factsheets, regulatory consultation papers. An LLM can reduce a forty-page document to a one-page summary in under a minute. The summary needs checking, but for initial triage it is a significant time saving.

Answering questions about your own documents. This is the application that is most relevant to regulated firms right now. When you constrain an LLM to a specific corpus of documents you control, such as your firm’s process documents, your file notes, or a client’s policy terms, and ask it questions only about those documents, the rate at which it invents things drops substantially. One implementation of this architecture (NotebookLM, when constrained to uploaded documents) has been cited as achieving below 1% hallucination rates on document-grounded queries [2]. That changes the risk profile meaningfully. The AI is no longer drawing on general training; it’s indexing your own material.

Internal knowledge bases. Building a searchable, conversational interface to your internal procedures, templates, and process documents. Instead of a new hire spending an hour hunting through a shared drive, they ask a question and get an answer with a reference to the relevant document.

Preparing for meetings. Feeding an LLM a client’s file summary and asking it to identify questions worth exploring, or gaps in the information gathered.

What this means for how you route work through AI

A useful starting point is to be explicit about which tasks you are using AI for and what the human review step is for each.

First, map where text-heavy work happens in your firm. Where do advisers or admin staff spend time drafting, reformatting, or summarising? Those are the candidate tasks.

Second, categorise by risk. A first draft of a client newsletter is low risk. A first draft of a suitability letter that goes to compliance review before anyone sees it is medium risk. A final suitability letter going directly to a client with no additional review is a task an LLM should not be completing autonomously, full stop.

Third, constrain the AI to your own material where accuracy matters. The architecture that routes an LLM’s responses through documents you own and control (your process notes, your file records, your templates) is more appropriate for regulated use than asking the model to rely on its general training. The hallucination risk is lower, and the outputs are auditable.

Fourth, document what you’ve deployed. The FCA and Bank of England issued a joint statement in 2025 signalling that AI governance for UK-regulated firms is moving from voluntary best practice toward regulatory expectation [3]. Knowing what AI tools your firm uses, for what purpose, and what the human oversight process is, is the baseline. If you can’t answer those questions now, start there.

A note on the technology moving quickly

The thing about LLMs is that the tools built on top of them are changing fast. Some of the niche applications that were useful twelve months ago have been absorbed into larger platforms. Some of the standalone tools being heavily marketed to professional services firms right now may not exist in their current form in two years. That’s not a reason to avoid the technology. It is a reason to be deliberate about what you depend on and to treat AI tool selection with similar due diligence to any other critical supplier relationship.

The underlying capability, drafting, summarising, answering questions about documents, is real and useful. The tool you use to access that capability is a separate question, and one worth revisiting periodically rather than deciding once.

What you can do this week

Most firms don’t need a formal AI strategy to start using these tools sensibly. They need one person to spend an afternoon experimenting, and then to share what they found.

Pick one text-heavy task that happens repeatedly in your firm. Draft an email, summarise a document, or pull key points from a piece of regulation. Run it through one of the mainstream tools (ChatGPT, Claude, or Microsoft Copilot if you’re in a Microsoft 365 environment). Evaluate the output honestly against the time it took.

That’s enough to start forming a view about where AI genuinely helps in your specific context, as opposed to where it’s more interesting than useful.

If you’d like to think through what this looks like for your firm specifically, a discovery call with Cordrey Consulting is a reasonable place to start.

This article is for informational purposes only and does not constitute regulated financial advice or a compliance opinion. Consult a qualified compliance professional for advice specific to your firm.

Sources

  • [1] Gadoci Consulting, “The ‘smart intern’ metaphor for LLMs” (2026-05-23). Vendor-sourced framework for framing LLM capability to business users.
  • [2] Google, NotebookLM product documentation and Google Workspace Studio integration notes (2026). Vendor-sourced. Sub-1% hallucination rate claim relates specifically to document-grounded queries where the model is constrained to uploaded source material.
  • [3] Bank of England and Financial Conduct Authority, AI and machine learning in UK financial services: joint statement (2025). Primary regulatory source for FCA/BoE AI governance expectations for UK-regulated firms. Available at fca.org.uk.