Handling Data Privacy and Retention in AI Projects for Law Firms

When a law firm starts talking seriously about AI, one question comes up before almost anything else: What happens to our client data?

That's not paranoia. That's professional responsibility. And it's exactly the right question to ask.

At Pivot180, we've worked with law firms across practice areas — mass arbitration, personal injury, immigration, and general litigation support. The firms we work with handle sensitive client information every day. So before we write a single line of automation or connect a single tool, we sit down and talk through how data moves, where it lives, how long it stays there, and who can see it.

This post explains how we actually approach that — not in theory, but in practice.

Why Data Privacy Is Different for Law Firms

Most businesses have privacy concerns. Law firms have ethical obligations on top of those concerns. Depending on your state bar's rules and the nature of your cases, you may have duties around confidentiality, informed consent for third-party tools, and data security that go well beyond what a general business faces.

That's why we don't treat law firm AI projects like a generic SaaS rollout. The architecture we build has to account for:

Who can access what data, and under what conditions
How long prompts, outputs, and logs are retained — and what actually gets stored along the way
Whether any client data reaches a third-party model in a way that could create confidentiality exposure
Whether we can produce records in discovery and honor a legal hold when a matter is open

None of this is hypothetical. These are decisions we make at the start of every law firm engagement.

How We Build AI Solutions Without Exposing Client Data

We Default to Your Infrastructure, Not Ours

The first question we ask is: where does the data need to live? For most law firms, the answer is inside your existing systems, behind your existing access controls.

Wherever possible, we build solutions that process data inside environments you already control — your Microsoft 365 tenant, your case management system, your secure document storage. We're not pulling client files into a shared consumer cloud or running them through a consumer AI tool.

When we do connect an AI model to your data, we use private or enterprise-tier API access with a data processing agreement in place, configured so that your data is not used to train the underlying model and is not retained by the model provider beyond the processing window.

We Design for Minimum Necessary Data

This is a principle borrowed from HIPAA, and it applies just as well to legal work. Our workflows are built so that only the data actually needed for a specific task is ever touched by automation.

If we're building a tool to draft demand letters, it doesn't need access to your entire case management database. It needs the specific fields relevant to that letter. We scope the data access tightly from the start, and we document exactly what's in scope.

We're Honest About What Gets Logged — and for How Long

Here's the part a lot of AI vendors gloss over, and where we'd rather be straight with you.

Any serious automation keeps an execution record. The platform that runs your workflow logs each step — which means the exact prompt sent to the AI model and the exact response that came back are both stored for a period of time. That's not a bug; it's what lets us troubleshoot a bad output, prove what happened, and produce records if you ever need them. But it means "the AI didn't keep anything" is almost never the whole truth, and we won't tell you it is.

So we make retention an explicit decision, not a default you inherit by accident:

On our standard hosted setup, execution records are retained for a bounded window — on the order of the most recent runs over roughly the last 30 days, whichever limit is reached first. On a high-volume firm, the real window can be shorter than that, because newer runs push out older ones.
If your policy requires something different, we can self-host the automation layer. At that point retention stops being a plan limit and becomes yours to set — from short windows to keeping everything indefinitely to match your records policy. The honest tradeoff: self-hosting means we (and you) own the server, the backups, and the security posture that comes with holding a law firm's data. That's a real decision, and we'll walk through it with you rather than flip a switch.
Intermediate files created during document processing are deleted when the task completes, not left lying around.

We put the retention model in writing before the project starts. You shouldn't have to guess what's stored or for how long.

Access Is Tightly Held

Access to the workflow layer is limited to us and anyone who holds credentials to your automation instance — nobody else. And we design the workflows themselves so that data access mirrors your existing permissions.

If a paralegal doesn't have access to a certain client folder in your case management system, they shouldn't be able to reach that information through an AI tool either. This matters especially in firms where different teams handle different matters — a mass arbitration team shouldn't be able to pull data from an unrelated personal injury case just because both use the same AI assistant.

We Can Produce Records in Discovery

Because prompts and outputs are logged, we can produce them. Today we can export execution data by workflow and date range, so if you need to show exactly what was sent to a model and what came back, it's retrievable.

Two things we're upfront about: if you need records pulled per matter, or need specific records exempted from deletion while a case is open (a litigation hold), those aren't sitting on a shelf ready to go — they're capabilities we build for your firm specifically. If you anticipate needing them, tell us early and we'll design for them from the start rather than retrofitting under pressure.

We Don't Use Client Data for Testing

This one surprises some firms: during development and testing, we use synthetic or anonymized data. We don't run real client files through a workflow to see if it works. We build it, test it with safe data, validate the output, and only then connect it to live data in a controlled environment.

What We Tell Firms About Vendor Selection

Some firms come to us already using tools they found on their own — AI research assistants, contract review platforms, drafting tools. Our job isn't to tear those out. It's to help you understand what they're doing with your data.

Questions we walk through together:

Does this vendor offer a data processing agreement, and have you signed it?
Is the tool using your inputs to train its models? (Many consumer-tier tools do this by default.)
Where are the servers located, and does that matter for your clients or your regulatory environment?
What happens to your data — and to the vendor's stored logs — if you cancel your subscription?

For firms just getting started, we help evaluate tools before you commit — our guide on how to choose AI tools for your law firm walks through that comparison in more depth. For firms already using tools, we help you understand what you've already agreed to.

What This Looks Like in a Real Engagement

One firm we worked with handles a high volume of demand letters across hundreds of open matters at any given time. The intake team was spending hours each week pulling the same set of facts from intake forms and case notes to draft the same letter structure — exactly the kind of administrative bottleneck AI is well suited to clear.

We built a workflow that generates a draft by pulling only the relevant fields from their case management system — no full file access, no document-level retrieval. The draft is assembled inside their Microsoft 365 environment. The AI model processes the request through an enterprise API under a data processing agreement, so the provider doesn't train on the firm's data or keep it after the run.

We were also straight with them about what is kept: each run leaves an execution record — the prompt and the output — under a defined, bounded retention window they signed off on, so they can troubleshoot a bad draft or produce a record if a matter ever calls for it.

The result: significantly faster drafting, no new confidentiality exposure the firm hadn't accounted for, and a workflow their managing partner was comfortable signing off on because they understood exactly how the data moved and what was retained.

That's the goal. Not just a faster process, but one the firm can actually stand behind.

Frequently Asked Questions

Does using AI for legal work violate attorney-client privilege?

Using AI tools doesn't automatically waive privilege, but it depends on how the tools are configured and what agreements are in place. If client data is sent to a third-party tool without a proper data processing agreement, or the vendor trains on your inputs, there's potential for exposure. The key is choosing tools with appropriate enterprise agreements and configuring them so client data isn't used for training or retained beyond what you've agreed to.

Can a law firm use ChatGPT or similar tools without compromising client confidentiality?

Consumer-tier tools like ChatGPT can pose confidentiality risks because they may retain conversation history and use inputs for training by default. Enterprise versions — ChatGPT Enterprise, or an API with a signed data processing agreement — offer stronger protections. Law firms should use business or enterprise tiers and confirm in writing that data won't be used for training or retained beyond the agreed window.

How long does Pivot180 retain data from a project?

We're specific about this rather than claiming we keep nothing. The AI model providers we connect to are configured not to train on your data or retain it after processing. The workflow layer, however, keeps an execution record of each run — including the prompt and the output — so we can troubleshoot and produce records. On our standard hosted setup that record is retained for a bounded window (roughly the most recent runs over about 30 days, whichever limit comes first). If your policy needs something different, we can self-host and set retention to match — including keeping everything indefinitely. We document the retention model in writing before the project starts.

Can you produce records in discovery, and honor a legal hold?

Yes. Because prompts and outputs are logged, we can export execution data by workflow and date range. If you need records pulled per matter, or need specific records exempted from deletion while a case is open, those are capabilities we build for your firm — so flag them early and we'll design for them up front.

What should a law firm ask an AI vendor before signing up?

Four questions cover most of the ground: Does the vendor offer a data processing agreement? Is client data used to train their models? Where is data stored, and for how long are logs kept? What happens to your data if you cancel? These apply whether you're evaluating a document review tool, a legal research assistant, or a general AI writing tool.

Is Pivot180's approach to data privacy specific to law firms, or do you do this for all clients?

The core principles — minimum necessary data, tight access controls, explicit retention, and enterprise-only vendor configurations — apply across every industry we work in. But law firms carry additional ethical obligations under state bar rules that shape how we design and document AI workflows, and they're more likely to need discovery and legal-hold capabilities. For legal clients we build in more explicit documentation and structure the engagement around those professional responsibility requirements.

Find out where your firm stands on AI privacy and readiness.

If you're thinking about AI for your practice, the first step is understanding what's already in place and where the gaps are — before anything gets built. Take the free 2-minute AI Readiness Assessment, designed for professional services firms like yours.