Many law firms are understandably reluctant to adopt agentic AI because of its well-documented security risks, including prompt injection. Running open-weight AI models (usually free) on computers you own—but that are not connected to your business network—lets lawyers who want to experiment do so far more safely.
Isolation doesn’t guarantee there will never be a prompt injection. It removes the adverse consequence of a security breach. There is no exfiltration channel, no tools to hijack, no network for anything to spread across. The result is much less risk of leaking confidential information.
For lawyers, that is not merely a comfort; it is a cleaner answer to a Rule 1.6 confidentiality analysis than any cloud vendor’s contractual assurances. When client information never leaves a machine you physically control, the question of who else can touch that data largely answers itself.
There is an ancillary benefit: cost control. The twenty-dollar-a-month plans are still with us, but they were never really all-you-can-eat, and the caps are tightening. Free and budget tiers now carry advertising, the most capable features keep migrating to new $100-and-up tiers, and industry analysts generally agree that current consumer pricing is promotional rather than sustainable. Given the industry’s economics, it’s pretty clear they can’t continue.
Analysts expect even standard consumer AI subscriptions to double or triple over the coming years. Industry projections suggest that leading app developers are considering pricing structures that could push individual power-user subscriptions toward $100-$200 per month to better reflect true operating costs. Of course, enterprise plans are expected to increase even more.
There’s another advantage: a lighter footprint. Running a model locally for your own queries means you’re not adding marginal load to strained grids and data-center buildouts. AI models are trained in power-hungry data centers, but the day-to-day querying you do afterward need not be. For a desktop machine sipping a few watts, the day-to-day querying you do after a model is trained barely registers on your power bill—or anyone else’s.
The Isolated-Machine Approach
The idea is to run models on standalone hardware entirely disconnected from the firm’s network. Running an open-weight LLM (like Meta’s Llama, Google’s Gemma, or OpenAI’s gpt-oss) on a completely isolated local machine allows a firm to poke, prod, and intentionally attempt to break a system without risking a data breach or a prompt injection attack spreading to client files. It is the digital equivalent of evaluating a suspicious package in an explosion-proof container, rather than opening it in the partner’s lounge.
One practical wrinkle: you will need a connection to download the model in the first place. The simplest approach is to set everything up on a connected machine, then disconnect it—or move the downloaded model files over to a machine that never touches the network at all.
How Is This Possible?
A frontier model is enormous, but the version you run locally is a compressed edition—shrunk down through a process called quantization so the essential capability fits in the memory of a high-end laptop or desktop. You give up a little precision and gain the ability to run the whole thing on your own hardware.
The performance gap is narrower than you might expect. A well-chosen local model now handles a large share of everyday legal-adjacent tasks—summarizing, drafting, reformatting—well enough that the difference from a frontier model is invisible for much of the work, even if it persists for the hardest reasoning.
Both Windows PCs and Apple Macs can do this. As a rule of thumb, 16GB of memory comfortably runs the quantized small-to-mid-size models that handle the everyday tasks above; 32GB or more opens up larger models where quality noticeably improves. Macs running Apple’s M-series chips are a particularly attractive option because the chips integrate graphics processing and share a single pool of memory with the rest of the system—an arrangement that is unusually well-suited to running these models efficiently. A base Mac mini currently runs about $600, which reframes “dedicated isolated hardware” from an exotic security measure into a modest line item—less than a year of one power-user cloud subscription.
You will get better results with more memory. Memory chip shortages mean a tricked-out Mac can create sticker shock. The consolation is that after the initial investment, you save money every month.
Local Model Limitations
To be sure, running AI apps locally is not nirvana.
Plain-English Supplemental Sources
A Beginner’s Guide to Running AI Models on Your PC
- This guide explains what a “local AI” is using zero jargon.
Offline AI Made Easy: A Layman’s Explainer
- This article demystifies how a computer can run a chatbot completely offline. It breaks down how data stays entirely on the hard drive, making it a highly reassuring read for professionals worried about confidentiality and data security.
- Free for personal and business use.
Run LLMs Locally on Mac with LM Studio
- Walks through downloading the app, loading a model, and starting a conversation.
For a visual walkthrough designed for non-technical users, the video “Local LLM Beginner Guide: You Can Actually Run AI For Free” provides a straightforward, step-by-step demonstration that explains how anyone can download and run an AI model entirely on their own machine, without a tech background.
Don’t want to do it yourself? Hire someone to help. Either way, don’t miss out on the advantages of running LLMs locally.