Local model is a phrase that sounds technical and mostly is not. Stripped of jargon, it means this: the AI software lives on a computer you control, and the thinking happens there. For an owner with one business that is a custody decision. For an owner with a family of businesses, it is also an architecture decision, because it decides whether running them on one system means pooling what each of them knows. On the Logix pattern, it does not.
The travel, reversed
Every AI model is, at bottom, a large file of learned patterns, plus a program that runs it. When you use a big cloud service, that file sits in a data centre, and your words travel to it. You type a question, the question crosses the internet along with anything you attached, a machine far away reads it, and an answer crosses back. Convenient, fast, and for many jobs perfectly fine. It also means the contents of that exchange exist, at least for a while, on someone else's equipment.
A local model reverses the travel. The model file sits on a machine in the business's office, or on a dedicated machine run for that business, and the files never make the trip. A recording is transcribed where it was recorded. A customer list is cleaned on the premises. A private document is summarised in the building. The results come back, and the only copies are the ones that business keeps.
Each business keeps its own walls
This is where the multi-business framing matters. The Logix ledger, set out elsewhere on this blog, is deliberate about what is shared and what is never shared. Structure, numbering, and queue discipline are the same across the family. Data, brand, and knowledge are always each business's own.
Local models are how that ledger holds for AI work. A model placed inside one business works on that business's recordings, records, and files, inside that business's walls. The standard for how the work is done is shared, so the owner can compare and oversee the family honestly. The data the work is done on is not shared, pooled, or copied across to make the system function. Adding a business to the family installs the standard again, with that business's own model placement and its own empty, private records. It never inherits a sibling's material.
Which jobs suit a local machine
The jobs that suit local placement are probably the majority of what a business actually does with AI. Transcription is steady, repetitive, and private. Summaries, record cleanup, first drafts from the business's own notes, answers drawn from its own documents: steady, repetitive, and private. In our own operation, meeting and call recordings are transcribed by a local transcription model on our own hardware, and local language models take the steady internal work of turning transcripts into summaries and action items, tidying records, and drafting from notes. Voice is generated locally where a voice is needed.
What may still use the cloud is everything that is not sensitive: public drafts, general research, the jobs where the biggest models earn their keep and nothing private is in the prompt. The line is drawn workload by workload, and any business on the system can be shown exactly where its work falls on it. The stack is also built only from models whose licenses permit commercial use, because some well-known open models do not, and a business cannot build on software it is not allowed to keep.
The honest trade-offs
Anyone selling local models should name the costs. The largest cloud models are still ahead on the hardest reasoning: a local model of practical size is a very capable worker, not the smartest mind on earth, and for a genuinely difficult problem the cloud may simply do it better. Hardware is a real cost: a machine that runs models all day is not free, and someone has to maintain it, update it, and notice when it stops. Local can also be slower. A data centre answers in a blink because it is a data centre; a box in an office answers at the speed of a box in an office.
Against that, the ledger has entries owners feel immediately. No per-minute or per-seat meter running on every transcription, summary, and cleanup job. No third-party copy of a business's records waiting to be included in someone else's bad headline. Work that continues when the internet does not. And custody that can be explained in one sentence to any customer who asks where their information goes: it stays here, in this business, and it does not travel to its sister businesses either.
Local where it counts
For most owners the answer is not local everything. It is local where it counts. The recordings, the records, the private files: local, inside the business they belong to. The public and the generic: wherever it works best. That is the line the framework builds on, and it is a line a business can only draw if local is genuinely available to it.
Local is not nostalgia for a pre-cloud world. It is a decision about custody, made by an owner who understands that each business's customer information is not an input to be handed out or pooled. It is the business itself, in written form, and it stays home.
See the framework behind the placement. The framework sets out what is standard across a family of businesses and what stays inside each one's walls. Read the framework →