Large language models are the dominant form of useful AI that professionals encounter today, and the current generation of these models has capabilities that did not exist several years ago. Understanding these capabilities in their current form is useful for a practitioner deciding how to incorporate AI into their work.
Current LLMs handle complex language tasks with a level of quality that makes them usable for serious professional work. They draft documents that require substantive editing rather than reconstruction. They analyse texts in depth, extracting specific information, comparing documents, and identifying patterns that would take a human reader substantial time to surface. They follow multi-step instructions, producing outputs structured the way the practitioner specified. They hold conversations across many exchanges, retaining context within a single session and adapting their responses to the practitioner's feedback. These capabilities are not perfect, and they are substantial enough that firms across every professional domain are actively reorganising work to take advantage of them.
Current LLMs are also multimodal. The models the professional is likely to be using can process not only text but also images, audio, and in some cases video. A practitioner can upload a photograph of a document and ask the model to extract information from it. They can upload a screenshot of a chart and ask for an interpretation. They can upload an audio recording and ask for a transcript or a summary. The model handles these non-text inputs by converting them into representations compatible with its internal mechanisms and then processing them alongside the text of the request. Multimodal capability means that AI tools are increasingly useful across the full range of material a professional works with, not just the subset that happens to be in plain text form.
Current LLMs can also be connected to external systems in ways that extend their useful capabilities significantly. A model on its own can only work with what is in its context window at a given moment. A model connected to tools can search the web, retrieve documents from a firm's document management system, query databases, call calculators, and execute operations in other software. When a professional uses an AI tool that can search the web, retrieve information from their calendar, or draw on their firm's internal documents, they are using an LLM that has been given access to external tools and is calling them as part of producing its responses. This tool-using capability, which has developed rapidly since 2023, has moved AI tools from being knowledgeable-but-isolated conversational partners to being assistants that can act in the professional's actual information environment.
The rate of capability improvement over the past several years has been faster than in any previous period of AI development. Models released in a given year have been meaningfully more capable than those released twelve months earlier. This pace may or may not continue, and the implication for the professional is that the AI tools available today are more capable than the ones that would have been available when any currently published guidance was written, and the tools available a year from now will be more capable still. Building professional practice around the specific quirks of a particular model's current behaviour is therefore less useful than building practice around the underlying mechanisms, which have been stable across this period of rapid development and are likely to remain stable as models continue to improve.