Should your business run its own private AI? What it takes, and the five setup steps

If your client files can't go into ChatGPT, the work AI could take off your desk stays on it: the meeting notes, the first drafts, the sorting of incoming documents. In a Flareworks AI audit in October 2026, the founder of a Singapore field-service group with about 50 staff put the dilemma plainly. She wanted AI, but "our own AI agent", not one running on Claude or ChatGPT, because of privacy.

The caution is reasonable. A clinic's patient notes, a law firm's case files, a financial advisor's client portfolios, HR records: under Singapore's Personal Data Protection Act (PDPA), the business stays responsible for that data, including when it goes to an outside service such as an AI tool. But there is a way to use AI without sending it anywhere. Run the model yourself, on a computer you control, so nothing leaves your network.

Many businesses don't need this. A business plan from a major AI provider usually costs less to run, gives access to stronger models, and the major providers state that business-plan data isn't used for training by default. The security note explains the differences between plans.

This is a practical guide, not legal advice. Check your PDPA and industry obligations with your adviser before you move regulated data anywhere.

You don't need to set it up yourself. The first half of this page covers what you need to decide, in plain words. The second half is written for whoever looks after your IT, so you can hand it straight to them.

First, decide whether you need this

There are options in between, such as your own cloud account or a separate environment a vendor runs just for you. The four setups are compared in Every AI proposal has a second spec.

Open models on one machine handle summarising, drafting, sorting and answering questions from your own documents well enough for many jobs. On hard reasoning they generally trail the best hosted models. Choose the job first, then the tool.

What you need

Memory decides which models fit on a machine, and memory speed decides how fast they answer. On a Mac, the system keeps part of the memory for itself, so in the Flareworks test a 32 GB machine fitted models up to about 20 GB.

In a 13 October 2026 Flareworks test on a 32 GB MacBook, a mid-sized model (9 billion parameters, a rough measure of size) read test messages and proposed customer-record updates correctly in 6.4 seconds each. A smaller model was slower, at 17 seconds, and raised false alarms on small talk. The test didn't cover smaller machines, so treat 32 GB of memory as the starting point. There's no monthly fee for the AI itself: the software is free. The main cost is the computer. Apple's Singapore store lists a Mac mini with 32 GB of memory and 1 TB of storage at S$2,649 (August 2026 pricing), plus the time of whoever sets it up. The numbers, and the cost against a cloud server, are in Private AI for a small business, measured.

What setup involves

Installing the software and a chat window takes an afternoon. Connecting a model to real work takes longer: in the same Flareworks test, a background job reading live messages needed fixes for separate login credentials, garbled text encoding, an undocumented search limit and a timing gap that could skip a message. The steps are at the end of this page, written for whoever looks after your IT.

What it looks like running

Flareworks has run this setup live since October 2026. A Mac in the office checks new messages every minute, and a model on that machine flags customer-record updates and sales signals. No message is sent to an outside AI service, and the model can only suggest changes: plain code, not the AI, decides what gets sent or saved.

Keep it private

Running AI on your own machine only protects your data if that machine is secure. The rule to apply is the principle of least privilege, set out by computer scientists Jerome Saltzer and Michael Schroeder in 1975: every person and program gets only the access it needs. Ask your IT person to confirm three things: the AI can't be reached from the internet, only the people who need it can log in, and no plug-ins quietly send data elsewhere. The same kind of bots that tried to break into a Flareworks server 165,000 times in 90 days scan for any computer left open to the internet, including one running AI.

Where to start

Don't start with "put the whole company on a private AI". Start with one sensitive task that repeats every week, such as summarising client meeting notes, drafting replies to patient enquiries, or sorting incoming documents. Load ten to twenty documents for it, not everything at once, run it on your own machine for two weeks, and compare the results with how it's done today. If it's good enough, expand from there. If it isn't, you've found that out on one laptop, not across the business.

The same principle runs the knowledge base behind the private search engine case study: a 66 MB model on an ordinary laptop indexes more than 13,000 notes, so the notes are never uploaded to an outside embedding service, and the search costs nothing per query.

For whoever looks after your IT

The steps, in order:

1. Install Ollama, free software that runs AI models on your own computer, from its download page.
2. Download and run a model with one command, such as ollama run gemma3, and check on its library page that the licence allows business use.
3. Add specialised models from Hugging Face, the largest public library of open models, if you need a particular language or task.
4. Give your team a chat window in the browser, so staff use it like ChatGPT without touching a terminal.
5. Connect it to the tools you already use, so it can sort enquiries or draft replies, with plain code deciding what is actually sent.

The commands and settings for each step are in a separate setup guide for IT, so you can forward one file.

It ends with a security checklist, and that is the part not to skip. The AI should only be reachable from your own office network, only the people who need it should be able to log in, and any feature that sends data elsewhere should be switched off.