When an attorney pastes a client's medical records or HR file into a consumer chatbot to get a quick summary, the action feels harmless — even efficient. But the question worth pausing on is not whether the summary was useful. The question is: where did that data go, and who controls it now?
For personal injury and employment law firms, this is not a theoretical concern. Client files contain Social Security numbers, medical diagnoses, wage histories, protected class information, and confidential communications. The gap between how a consumer large language model (LLM) handles that data and how a purpose-built legal AI platform handles it is significant — and understanding it matters for professional responsibility as well as IT hygiene.
This post walks through the specific data-handling differences that matter most to small law firms, explains what to look for when evaluating any AI tool, and shows how purpose-built platforms are architected differently from the ground up.
How Consumer Chatbots Actually Handle Your Inputs
Consumer-facing AI products — the kind you access through a browser tab without an enterprise agreement — are generally designed to improve the underlying model over time. That means your inputs, including the documents you paste or upload, may be used as training data unless you have explicitly opted out or signed a separate data processing agreement.
The default data retention and usage policies of many consumer AI products allow the provider to store conversation history, use it for model training, and share it with affiliated services. Even when a "no training" option exists, it is often buried in settings, not the default, and not retroactive.
For law firms, the practical implications are serious:
- No confidentiality guarantee at the infrastructure level. A consumer chatbot's terms of service are not a Business Associate Agreement. If you upload a client's medical records without a BAA in place, you may be exposing protected health information to a third party without the contractual protections that HIPAA is designed to require.
- No audit trail. Consumer tools typically do not provide the kind of access logging that allows a firm to demonstrate, after the fact, who accessed what data and when.
- No data residency assurance. Many consumer AI products route data through global infrastructure. Where your client's file sits geographically — and under what legal jurisdiction — may be entirely opaque.
- Model outputs may surface training data. Some researchers have documented instances in which LLMs reproduced text resembling prior training inputs, though the frequency and conditions are not fully characterized. If your client's records were to become training data, that risk, however uncertain, would exist.
None of this means consumer AI tools are useless for lawyers. They can be valuable for drafting general correspondence, brainstorming arguments on hypothetical facts, or researching publicly available information. The problem arises when attorneys treat them as a secure environment for live client data.
The Architecture of a Purpose-Built Legal AI Platform
Purpose-built legal AI platforms are designed with a fundamentally different data model. Rather than feeding your inputs into a shared model that improves with use, they operate on a retrieval-augmented generation (RAG) architecture: the AI reads your documents, indexes them within your isolated environment, and generates answers grounded in your specific case file — not in a pooled model trained on other firms' data.
Here is what that means in concrete terms:
Your data stays in your environment. A platform like ProvaLens hosts all case data on Microsoft Azure with AES-256 encryption. Your documents are not shared with other firms, not used to train a general model, and not accessible to the AI provider's staff without your authorization.
A BAA is signed before the first upload. This is the contractual foundation of HIPAA-compliant AI use. ProvaLens executes a Business Associate Agreement before any client data enters the system, establishing clear obligations around how protected health information is handled, stored, and secured.
Full audit logging. Every query, every document access, and every export is logged. If a firm ever needs to demonstrate what happened with a client's records — for a bar complaint, a malpractice inquiry, or a client's own request — the audit trail is there.
Answers are cited to the source. Because the AI is reading your documents rather than generating from general knowledge, every answer it returns is tied to a specific page and paragraph in your case file. This is not just a quality feature; it is a data-integrity feature. You can verify that the AI's output is grounded in your records, not in something it hallucinated from training data.
This architecture also addresses a question that comes up frequently in deposition preparation: can AI help surface contradictions across a case file? A purpose-built platform is designed to cross-reference your entire case file simultaneously — something a human reviewer doing linear document review cannot replicate at scale. For example, ProvaLens's Potential Conflicts feature can flag possible discrepancies between documents, such as a treating physician's deposition testimony about first onset of symptoms compared against intake notes from months earlier, and cite both passages to their exact source pages. Whether any flagged discrepancy is meaningful is always a judgment call for the reviewing attorney; the platform surfaces the issue, it does not resolve it.
What to Actually Verify When Evaluating Any Legal AI Tool
The marketing language around AI security has become so saturated with phrases like "enterprise-grade" and "bank-level encryption" that it can be hard to know what questions actually matter. Here is a practical checklist for evaluating any AI tool before uploading client data:
1. Does the vendor sign a BAA?
If you are handling any medical records, health insurance information, or anything that could constitute protected health information, a BAA is non-negotiable. Ask for it before the demo ends. If the vendor hesitates or says it is only available on certain tiers, that tells you something.
2. Is your data used to train the model?
Ask directly: "Is any data I upload used to improve your AI model?" Get the answer in writing, in the contract. A purpose-built legal platform should answer no.
3. Where is the data hosted, and by whom?
Know the cloud provider (AWS, Azure, Google Cloud), the data region, and whether any subprocessors have access to your data. Reputable platforms will disclose this in their privacy policy or data processing addendum.
4. What is the data retention policy?
How long does the vendor retain your data after you close an account? Can you export and delete everything? What happens to data from archived matters?
5. Is there an audit log?
For professional responsibility purposes, being able to demonstrate what data was accessed, by whom, and when is increasingly important. Consumer tools do not offer this. Purpose-built platforms should.
6. Does the AI cite its sources?
This is both a data integrity question and a professional reliability question. An AI that returns answers without citations is, from a practical standpoint, asking you to trust its output without verification. In a legal context, that is not acceptable. Every answer should point back to a specific document, page, and paragraph.
The Expert Witness Scenario: Where Data Handling Meets Document Review
Consider a practical scenario that illustrates why this matters beyond abstract compliance. An employment attorney is preparing to depose an expert witness retained by the defense. The attorney wants to know whether any AI tool can help cross-reference the expert's prior reports against the current case file to surface possible inconsistencies in the documents.
The instinct might be to paste the expert's prior reports and deposition transcripts into a consumer chatbot. The problem is twofold. First, those documents may contain references to other parties' confidential information — prior plaintiffs, prior cases, protected class data. Second, the consumer chatbot has no way to cross-reference those documents against the specific facts of your client's case file, because it does not have access to your case file in a structured, indexed way.
A purpose-built legal AI platform approaches this differently. When the expert's prior reports, the current case's medical records, and the deposition transcript are all uploaded into the same matter, the platform can cross-reference passages across those documents and flag possible inconsistencies — cited to the exact page — for the attorney to evaluate. This is document cross-referencing, not legal strategy: the AI surfaces what the documents say and where they may conflict; the attorney decides what is significant and how to use it.
The following is a hypothetical illustrative example only. Imagine a small PI firm preparing to cross-examine a defense IME doctor whose report attributes a plaintiff's lumbar injury to pre-existing degenerative changes. In this hypothetical, the firm's AI platform, working across the full case file, surfaces a passage in the doctor's deposition from a prior case — already uploaded into the file — where he testified that the same type of degenerative changes are clinically indistinguishable from acute trauma without imaging taken within 72 hours. The platform flags the two passages as a potential conflict and cites both to their source pages. The attorney then reviews the flagged passages and decides whether and how to use them in cross-examination. This kind of cross-document flagging is the capability a purpose-built legal AI platform is designed to support.
Why the Subscription vs. Per-Case Distinction Matters for Firm Economics
One final dimension worth addressing: the economics of legal AI are sometimes misunderstood in ways that affect how firms think about data handling.
A monthly platform subscription — the kind that gives your firm access to document intelligence, timeline generation, contradiction detection, and AI chat with citations — is firm overhead. It is analogous to your practice management software or your research subscription. It is not something you pass through to individual clients.
However, if you engage a service like ProvaLens's Expert Services — a done-for-you AI case analysis that produces interactive timelines, medical chronologies, and observations drawn from the case documents for a specific matter — that is a genuine, out-of-pocket case cost. Like an expert witness fee or an eDiscovery vendor charge, it may in some circumstances be appropriate to treat as a case expense billed at cost to the client, though attorneys should consult their own bar guidance on disclosure and informed consent requirements before doing so. The distinction matters because conflating the two can create billing practices that draw scrutiny.
From a data handling standpoint, both the subscription platform and the Expert Services operate within the same secure architecture: BAA in place, Azure-hosted, AES-256 encrypted, fully audited. The security posture does not change based on which service tier you use.
Closing Thoughts
The gap between consumer AI and purpose-built legal AI is not primarily about capability — it is about architecture, data governance, and professional accountability. Consumer chatbots are powerful general tools. They are not designed to hold a law firm's confidential client data securely, generate citable answers from a specific case file, or provide the audit trail that careful professional practice calls for.
For small PI and employment firms that handle sensitive medical records, wage data, and protected class information every day, the choice of AI tool is a data decision as much as a productivity decision. Understanding where your client's data goes — and what contractual protections govern it — is the starting point for using AI responsibly in legal practice.
If you want to see what a purpose-built, HIPAA-aware, document-grounded legal AI platform actually looks like in practice, Start your free ProvaLens trial and upload a real case file — the difference in how your data is handled, and what the AI can do with it, will be immediately apparent.