We get the question from clinics, lawyers, accountants and public organisations: can we use AI without sending our data to a cloud abroad? The answer is yes — with an important clarification about what "AI" means in that sentence.
What works well locally
Transcription is the clearest example. Speech to text runs excellently on an ordinary modern computer, including in smaller languages, and the audio never has to leave the machine. We have built and use a tool for exactly that ourselves.
Classification, summarising and extracting information from documents are also within reach. A local model can read a case file and pull out names, dates and amounts, or sort incoming mail by topic, without anything leaving the building.
What still lags
Tasks needing broad knowledge and long reasoning — writing a thorough legal assessment, holding a long conversation with context, solving something that takes many steps — there the large cloud models are still noticeably ahead. The gap shrinks every six months, but it is there.
That is not an argument against local AI. It is an argument for being precise about which task you want solved before deciding how.
Long reasoning
Tasks needing many steps in a row, where one wrong step topples the rest.
Broad general knowledge
Questions that assume knowledge of everything outside your own material.
Very long documents
Cases running to hundreds of pages, where the whole thing has to be held at once.
Open-ended conversation
Dialogue across many turns where tone and context have to hold throughout.
What it takes
Less than most people expect. Transcription and the lighter language tasks run on a recent Mac or a PC with a reasonable graphics card. If you want to run larger models for several users at once, you need a dedicated machine, and then you are looking at an investment on the scale of a solid server — not a data centre.
What gets overlooked most often is maintenance. Better models arrive every quarter. Someone has to keep an eye on that and swap them in, or in two years you will be running on something hopelessly out of date.