A law firm can use AI safely by controlling three things: what data goes into which tool, who is allowed to use which tool for what purpose, and the order in which the firm builds its technology foundation before layering AI on top. The short version: public, consumer-grade AI tools should never see privileged or confidential client information, full stop, and any AI tool a firm does trust with client data needs to sit inside infrastructure the firm actually controls, not a free consumer account. Getting this right is less about picking the perfect AI tool and more about doing the unglamorous setup work first.
This is not legal advice, and a firm’s ethical obligations around confidentiality and privilege are between that firm and its bar association. What follows is how we think about the technology side of that problem.
Why is this the objection that actually stops legal buyers?
Every other industry worrying about AI adoption is thinking about accuracy or cost. Legal is different, because a single accidental disclosure of privileged information is not just embarrassing, it can put privilege on the underlying matter at risk entirely. That is a much higher stakes mistake than a bad AI-generated first draft, and it is why so many firms have quietly told staff “don’t use AI” instead of working out where it is actually safe to use.
That instinct is understandable. It also usually means the firm is losing the time savings on the safe, low-risk uses too, because nobody drew the line between what is fine and what is not.
What should never go into a public AI tool?
If it is client-identifying, case-specific, or covered by attorney-client privilege, it does not belong in a free or consumer-tier AI chatbot. That includes client names paired with case facts, draft settlement positions, internal strategy discussions, anything under a protective order, and documents produced in discovery. Consumer AI tools generally are not built with the data handling guarantees a law firm needs, and once information leaves the firm’s control into a tool without a business-grade data agreement, the firm has lost the ability to say with confidence where that information went.
A useful test for staff: if you would not paste it into a public forum or send it to a personal email account, it should not go into a consumer AI tool either.
What does a safe AI setup actually look like?
A safe setup has a few concrete features. The AI tool sits inside infrastructure the firm controls or has a real business agreement with, not a free personal account someone signed up for on their own. Access is limited to the people and purposes it is actually needed for, not open to the whole firm by default. There is a written policy telling staff exactly what is and is not allowed, so nobody is guessing. And someone at the firm, even if that is an outside partner, is responsible for reviewing how the tools are actually being used, not just setting them up and hoping.
None of that requires an enterprise-scale budget. It requires deciding these things on purpose instead of letting each staff member make their own call about which tool to trust with client information.
This is not hypothetical for us. In a case study published by Pax8, one of our Microsoft partners, Cory Ruthardt, President of Simpatico Systems, described exactly this setup in a real legal deployment: “Copilot deploys within a customer’s environment where data is already protected, giving the law firm confidence.” That is the direct answer to the privilege objection, from an actual firm, not a hypothetical one. The firm is not named in the published case study. Full case study: https://www.pax8.com/en-us/case-study/copilot-agent-helps-lawfirm-transform-case-management/
Why does the order of operations matter?
This is the part firms skip, and it is usually why AI adoption goes sideways. The order that actually works is stable infrastructure first, then structured data, then automation, then AI.
Stable infrastructure means your systems are patched, your access controls are set up correctly, and you are not running case management software on machines nobody has updated in two years. Structured data means your client and matter information lives in a system that is organized and searchable, not scattered across shared drives with inconsistent file names. Automation means the repeatable steps, the ones covered in our companion post on practical AI and automation use cases for law firms, run through defined processes instead of whoever remembers to do them. Only after those three are in reasonable shape does AI have a stable, organized foundation to sit on top of.
Skip straight to AI without the first three steps, and the AI tool inherits every existing problem: messy data, inconsistent access, and no clear process for who is responsible for what. That is how a firm ends up with an AI tool that has access to more than it should, because nobody structured the data or the access controls before turning it on.
That sequencing is no longer just our recommendation, it is what actually happened in the legal deployment referenced above. Per the Pax8 case study, Simpatico ran a year-long data cleanup and cloud security initiative for the firm before building the Copilot case management agent. Infrastructure and structured data came first. The AI came after, on a foundation that was already in shape.
Doing it in that order did not slow the results down. The same deployment cut client onboarding from 3-5 days to under one business day, dropped paralegal support needs by roughly three to one, and moved case-acceptance decisions from about five days to the same day. Getting the security and data foundation right first was not a tradeoff against speed. It was the reason the speed was possible.
Think of it as a permissions problem before it is ever an AI problem. If a paralegal’s login can already reach every client file in the firm regardless of which matter they are staffed on, that exposure already exists. Adding an AI tool on top just makes it faster to find. The fix is the same either way: get access scoped to what each person actually needs before you add a tool that can search across everything in seconds.
What does this look like for a firm still running on shared drives and email?
It looks like the same three steps, just slower to reach. A firm with client files scattered across a shared drive, personal inboxes, and whatever software each attorney happens to prefer is not ready for AI yet, and that is fine to say out loud. The honest starting point is consolidating where client and matter data actually lives, then cleaning up who has access to what, before any AI tool enters the picture. Skipping straight to the tool just moves the same disorganization somewhere faster and harder to audit.
Why does Simpatico’s security posture matter here?
Legal is not the only regulated environment we work in. A meaningful part of our client base is defense industrial base contractors working toward CMMC, one of the more demanding compliance and cybersecurity frameworks a small business can be asked to meet. That work means our default posture toward client data, access control, and vendor risk already assumes an environment with less room for error than most firms operate in day to day.
What can a law firm borrow from Simpatico’s own AI policy?
A starting template, and now a real deployment to point to as well. We did not just tell our own staff to be careful with AI and leave it there. Every Simpatico employee went through required AI training and signed a written company AI use policy before using AI tools on the job. If your firm does not have a written AI policy yet, the fact that we built one, trained on it, and required sign off before anyone touched an AI tool is a reasonable starting template for the shape yours could take: what tools are approved, what data can and cannot go into them, and who signs off before a new tool gets added.
Beyond our own policy, there is now a published, third-party case study showing what this looks like in an actual legal deployment, the same Pax8 case study referenced earlier in this post. It is a stronger reference point than our internal policy alone, because it is a real client environment described by a third-party publisher rather than a company describing its own rules.
You can see how this fits into our broader legal services approach here: https://simpatico.com/industries/legal-services/
What should a firm actually do first?
Start smaller than feels ambitious. Pick one low-risk, high-volume task, something like drafting a routine internal memo or summarizing a public, non-privileged document, and run it through an approved tool with a real data agreement behind it. Write down the rule in one page: what is approved, what is not, and who to ask if unsure. Then expand once the infrastructure and data work underneath it is actually solid, not before.
Is this legal advice?
No. We are not your firm’s ethics counsel, and this post is not legal advice about your specific privilege or confidentiality obligations. Talk to your bar association or ethics counsel about what your jurisdiction requires. What we can help with is the technology side: making sure the tools, access, and data handling around AI adoption are built the way a firm handling privileged information actually needs them built.
If your firm is trying to figure out where to draw that line, that is worth a real conversation. [INPUT NEEDED: CTA link or booking URL for a legal-specific consultation]
Where has your firm drawn the line on AI so far, and where are you still unsure? Tell us in the comments.
