Skip to content

Practical guide · October 30, 2026

A local AI assistant for a Mac that keeps client data private

How lawyers, accountants and consultants can run an AI assistant on a Mac where client documents never leave the computer: what to install, which models fit 16 GB and 32 GB, and what to check first.

Why client data and cloud AI do not mix easily#

A cloud AI assistant works by sending your text to a provider's servers. For a recipe that is fine. For a contract, a tax file or a client's financial statement it raises a question you have to answer before you paste anything: who is allowed to see this text, and where does it go?

The question is not hypothetical. A 2025 study by KPMG and Melbourne Business School of 48,000 people in 47 countries found that almost half of employees admit to uploading sensitive company information to free public AI tools (Melbourne Business School release). A professional bound by confidentiality cannot treat that as normal.

One way to avoid the question is a model that runs on your own computer. The text is processed on the Mac and is not sent to a provider. This article shows how to set that up with Growth OX, and where it falls short.

What "local" means here#

Growth OX runs local models through Ollama, a separate program that must be installed and running on the same Mac. With a local model, the app sends no text over the internet. Your workspace, drafts and notes are also stored on your computer, and the licence server never receives them.

Downloading the model and installing updates still use the internet. Only the work you give the model stays on the machine. If you also connect a cloud provider for other jobs, that provider receives the text of those jobs. You choose which model does which job, so a client file can stay with the local one.

Set it up in five steps#

  1. Install Ollama and open it. If Ollama is pointed at another computer, Growth OX will not install a local model, because a local model only runs on this computer.
  2. Measure your Mac. In Growth OX choose Measure my computer. It reads memory, free disk and graphics, and downloads nothing.
  3. Read the verdict next to each model: fits, borderline or does not fit. Models that fit are marked recommended.
  4. Install one. Accept the licence, then choose Install. The download shows its progress and can be stopped.
  5. Try it on a harmless document, not a client file. The app runs three short jobs (a Turkish summary, an English translation and a tool call) and reports whether each worked and how fast. Then you choose the jobs the model takes.

A model that cannot call tools is installed but gets no assistant jobs. The app tells you which ones can.

Which model fits which Mac#

The app decides from your own measurement, but these are the starting points in its model list. All are licensed Apache-2.0.

Local models in the Growth OX list, by Mac memory (download sizes as listed)
Mac memoryModelDownload sizeNote
16 GBQwen3 8B5.2 GBRecommended by the app on a 16 GB Mac
16 GBQwen3 4B2.5 GBLighter, faster, less capable
24 to 32 GBgpt-oss 20B14 GBToo large for a 16 GB Mac in the app's check
24 to 32 GBQwen3 30B (A3B)19 GBA mixture-of-experts model; all of it is loaded
48 GB and moreQwen3 32B20 GBFor larger Macs

Speed is not promised in advance. After the trial the app shows a measured speed in tokens per second for your Mac.

What a local model does not do#

  • It is slower and weaker than the largest cloud models. Expect it to handle summaries, drafts and translations; check its work on anything that matters.
  • It does not make you compliant. Whether a tool is allowed for client work depends on your professional rules, your client's terms and your jurisdiction. Ask your regulator or bar association if you are unsure.
  • It can be wrong. You stay responsible for the output, as the AI use terms say.

Approvals stay in force with a local model. Posts, replies, e-mails and spending wait for your yes, so a draft prepared from a client file is never sent by itself. See Approvals.

When the Mac is busy#

A local model uses a large share of the Mac's memory while it runs. Growth OX has three profiles for how much it may use: Safe, Controlled and Experimental. Safe leaves the computer plenty of room and is always available. Controlled opens after three clean runs on your computer, and Experimental needs your approval. On battery or in low power mode, the Safe profile is used.

If the computer is under strain, the app holds new model loads, and under heavy strain it stops one model and saves the job instead of freezing the Mac.

A checklist before the first client file#

  1. Confirm in the model list that the verdict for your Mac says fits, not borderline. A borderline model may slow other apps down.
  2. Run the three trial jobs on a document that contains nothing confidential and read the results yourself.
  3. Assign the local model to the jobs that touch client files, and check that no cloud credential is assigned to those jobs.
  4. Keep the approval rules on, so every outgoing message waits for you.
  5. Write down, for your own file, which model processed which client's documents. The app records approvals; your professional rules may ask for more.

What it costs#

A local model has no per-use bill; it uses your Mac. Growth OX is in development and not on sale yet, so this article states no price for the app. For the full data picture read Privacy.

Questions#

Does a local model send my client documents anywhere?
No. A local model runs on your Mac and sends no text over the internet. Downloading the model and updates use the internet, but your documents are not part of that.
How much memory do I need?
A 16 GB Mac can run Qwen3 8B, which the app recommends. Larger models such as gpt-oss 20B need more memory. Run the measurement in the app: it shows a verdict for your Mac.
Is a local model accurate enough for legal or tax work?
It can draft and summarise, but it can be wrong. Treat its output as a first draft and check it yourself.
Can I use a cloud model for some jobs and a local one for client files?
Yes. You choose which model does which job, so client files can stay with the local model.
Does this make me compliant with my professional rules?
No tool can promise that. Check your professional rules and your client's terms.

Sources#