Private AI for Lawyers: Confidentiality Without the Cloud

Every lawyer we talk to has the same two reactions to AI, usually in the same breath. First: "This would save my team hours every week." Second: "I cannot paste client documents into someone else's server."

Both reactions are correct, and the second one is not paranoia. It is the job. Lawyers hold information under duties of confidentiality that most businesses never face, and several state bar opinions on generative AI say some version of the same thing: understand where client data goes before you send it anywhere. For a solo practitioner in Portland or a four-attorney firm in Manhattan, that homework usually ends the conversation, because nobody has time to parse a cloud vendor's data processing addendum between filings.

There is a version of AI where the homework gets dramatically shorter: the model runs on a Mac in your office, and client documents never leave the building. This post explains how that works, what it honestly does and does not solve, and what private AI for lawyers looks like in the two workflows firms actually care about, document review and contract analysis.

Why confidentiality is the blocker, not the technology

The technology question ("can AI summarize a deposition?") was answered years ago. The blocker is a data-handling question, and it is worth stating precisely.

When an attorney evaluates any tool that touches client information, the analysis usually turns on a few concrete facts. Who receives the data? Where is it stored, and for how long? Who at the vendor can see it? Could it be used to train future models? Could it be produced in response to a subpoena served on the vendor rather than the firm?

Cloud AI vendors have real answers to these questions, and the reputable ones have good answers. Enterprise API tiers commonly offer zero-retention options and contractual commitments against training on your inputs. But the firm still has to find those answers, verify which pricing tier they apply to, confirm they cover the specific product being used (consumer chat apps and business APIs often have different policies from the same vendor), and re-verify when terms change. That is genuine diligence work, repeated per vendor, forever.

A local AI setup on Mac collapses most of that analysis, because the answer to "who receives the data" becomes: nobody. The document is processed on your own hardware, by a model file sitting on your own disk, with the network cable metaphorically unplugged.

What actually happens to your data with a cloud API

To see the difference clearly, follow one contract through each pipeline.

With a cloud API, your 40-page services agreement is transmitted to the vendor's servers, tokenized, processed on their GPUs, and a response is transmitted back. Along the way it may be held in transit logs, retained for abuse monitoring (30 days is a common default window), and handled under whatever jurisdiction the data center sits in. Each of those steps is governed by a contract you must read, on terms the vendor can update. None of this makes cloud AI reckless, and we say so plainly in our breakdown of the hidden costs of cloud AI. It makes cloud AI a third-party disclosure question that a lawyer must analyze, document, and often disclose in engagement letters.

With local processing, the same agreement is read from your SSD into your Mac's memory, the model generates its analysis on the Apple Silicon chip, and the output lands in a file on the same machine. You can verify this yourself in about ten seconds: turn off Wi-Fi and run the query again. It works identically, because there is no server on the other end. That is the practical meaning of AI without cloud, and it is a fact you can demonstrate to a managing partner rather than a policy you must trust.

One honest caveat: local means local. If your firm then syncs that output folder to a consumer cloud drive, you have reintroduced a third party. A good setup includes deciding where outputs live, which is part of what we cover in our use case walkthroughs.

Real workflow one: document review on your own hardware

Document review is the highest-volume, most confidentiality-sensitive work in most small firms, which makes it the natural first workflow for on-device AI for business.

A concrete setup we configure often: a Mac Studio with an M4 Max chip and 64 GB of unified memory ($2,499 at Apple retail) running a 70B-class open model at roughly 18 to 22 tokens per second. At that speed the machine reads and summarizes a 25-page document in a couple of minutes, unattended. Point it at a folder of discovery documents overnight and a three-attorney firm in Chicago arrives to a first-pass summary of each file: parties mentioned, dates, key assertions, documents flagged for human review.

The important design point is that the model is a first-pass reader, not a reviewer of record. It surfaces and organizes; attorneys judge. Open models in the 32B to 70B class are genuinely good at "read this and tell me what is in it," and their occasional mistakes are exactly why the workflow keeps a lawyer in the loop. We are direct about this limitation because firms that expect a robot associate return the hardware, and firms that expect a tireless summarizer keep it running every night.

Smaller firms do not need the flagship machine for this. A Mac Mini with an M4 Pro chip and 48 GB of RAM (about $1,799) runs 32B-class models at interactive speeds, which handles routine review comfortably. Our benchmark comparison of the Mac Mini and Mac Studio has the full tok/s tables by model size.

Real workflow two: contract analysis and drafting support

Contract work is lower volume but higher stakes per document, and it benefits from configuration more than raw horsepower.

A configured local agent for contract analysis typically does four things: extracts defined terms and cross-references into a checklist, flags clauses that deviate from the firm's preferred positions (loaded from your own precedent bank as knowledge files), summarizes obligations by party with dates, and drafts redline suggestions in your house style for attorney review. Because the precedent bank and the client paper both stay on the machine, the firm's accumulated negotiating knowledge, which is itself confidential work product, never becomes someone else's training data or retention-policy question.

This is agent configuration rather than out-of-the-box behavior, and it is the step most firms should not attempt alone. We walked through the mechanics in our guide to custom AI agents with Open WebUI, and it is one of the four services in our setup packages: the agent arrives configured for your document types, with guardrails written in, so the first week is useful instead of experimental.

The numbers case is straightforward too. A contract-heavy pipeline running through a cloud API commonly lands between $200 and $500 a month in metered fees, roughly $2,400 to $6,000 a year, and the bill grows with your caseload. The one-time local machine costs $1,799 to $2,499. We ran the full breakeven math, including a Manhattan firm scenario, in our ROI analysis for four industries; the fee-only breakeven landed near month 10, before counting roughly $10,400 a year in paralegal hours recovered.

What local AI does not do, said plainly

This section matters more for lawyers than for any other audience we write for, so here it is without hedging.

Local processing is designed for privacy-sensitive workflows. It is not a compliance certification, and buying a Mac does not resolve your professional responsibility analysis. Whether a given use is consistent with your duties of confidentiality and competence, whether client consent or engagement-letter disclosure is warranted, and how your bar's AI guidance applies to your practice are questions for you and, where appropriate, your ethics counsel. What local processing changes are the underlying facts that analysis runs on: no third-party recipient, no vendor retention policy, no terms-of-service drift. Simpler facts, same duty of care.

Two more honest limits. First, frontier cloud models still outperform local open models on the hardest reasoning tasks; a firm doing novel appellate argument analysis may reasonably keep a cloud tool in the mix for non-confidential work, and local for everything client-related. Second, the machine needs the same physical and access security as the file server it sits next to. On-device AI inherits your office's security posture, good or bad.

What a private setup costs, in one-time terms

Because everything above runs on hardware you own and open models that are free to download, the cost structure is a purchase, not a plan.

• Mac Mini M4 Pro, 48 GB RAM ($1,799): 32B-class models at interactive speed. Right for solo practitioners and two-to-four attorney firms doing review and drafting support.
• Mac Studio M4 Max, 64 GB RAM ($2,499): 70B-class models at 18 to 22 tokens per second. Right for firms with heavy contract analysis or overnight batch review.
• Electricity: a few dollars a month. Model updates: free, and the open models we recommend improve every quarter.

If you want the selection, installation, and configuration handled, a professional local LLM setup service is a one-time line item; our pricing is structured that way deliberately, with hardware sourced at Apple retail prices and no markup. The same logic that makes lawyers cautious about cloud data handling tends to make them appreciate a vendor relationship that ends cleanly, with optional ongoing support only if you want it.

Key Takeaways

• The AI blocker for law firms is data handling, not capability. Cloud APIs are a third-party disclosure question requiring per-vendor diligence; local processing removes the third party.
• You can verify local privacy yourself: turn off Wi-Fi and the model still runs, because client documents are processed entirely on your own Mac.
• Document review and contract analysis are the proven workflows, with the model as first-pass reader and attorneys as the judgment layer.
• Hardware runs $1,799 to $2,499 one time at Apple retail prices, versus roughly $2,400 to $6,000 a year for a metered cloud pipeline at typical small-firm volume.
• Local processing supports a privacy posture; it is not a certification. Your confidentiality and ethics analysis still belongs to you and your bar's guidance.

If your firm has been stuck at "we would use AI, but the confidentiality question," that question is exactly where we start. Maai Machines handles hardware recommendation and sourcing at no markup, the complete local AI setup on your Mac, custom agent configuration for document review and contract workflows, and ongoing support after the install. Visit maaimachines.com to book a conversation about running AI your clients never have to worry about.