A personal injury attorney uploads multiple pages of medical records to an AI tool, asks for a summary, and gets a clean chronology back in two minutes. It’s organized by date. The language is precise. The treatment timeline is coherent.
It’s also missing four months of physical therapy, two surgical consultations, and the imaging studies that established causation. Those records were in the case file. They weren’t in the upload.
The summary is accurate about what it received. It’s incomplete about what it didn’t. And the attorney, looking at a well-formatted document, has no reason to suspect the gap unless they already knew the full record, which would have made the summary unnecessary.
This is the problem with AI that operates outside the case file. It doesn’t fail by getting things wrong. It fails on what it wasn’t given. And it gives no sign that anything is absent.
Contents
How most AI reaches law firms today
The typical AI workflow in a PI firm looks something like this. A lawyer or paralegal opens a general-purpose AI tool. They copy a section of a document, paste it in, and ask a question or request a draft. The tool responds based on the text it received. The user reviews the output, edits it, and pastes it back into whatever system holds the case file.
Every step in that process requires a human to decide what the AI sees. That’s not supervision. That’s curation. And it introduces the same variability that the AI was supposed to eliminate. If the user selects the right documents, the output is useful. If they miss something, the output reflects the miss. The AI doesn’t flag it because it has no way to know.
A Harvard Business School study conducted with Boston Consulting Group gave consultants access to a leading generative AI model and measured what happened. On tasks that fell within the model’s competence, the AI group completed more work, finished faster, and produced higher-quality output. On tasks that fell outside it, the same tool made performance worse, and the people who leaned on it hardest were the ones who got it wrong. The researchers called this the jagged frontier: AI is reliable in some places and quietly unreliable in others, and the output looks the same either way. For a PI firm, the boundary that matters most isn’t the type of task. It’s the inputs. A tool fed a partial record produces confident, polished, incomplete work, and nothing in the result tells you which kind you’re looking at.
Where built-in AI changes the outcome
Five moments in a PI case illustrate the difference between AI that sits alongside the work and AI that operates inside it.
Intake. When a lead comes in after hours or through a web form, an external AI tool can help draft a follow-up email. An AI built into the case management software does something different. It captures the intake data, creates the case record, and routes the matter based on criteria the firm has already defined. The information entered once becomes the foundation every other stage builds on. Nothing gets re-keyed. Nothing gets lost in a handoff.
Medical record review. It’s the one where the gap is widest. A single PI case can generate hundreds of pages of medical documentation across multiple providers, spanning months or years of treatment. An external tool can summarize whatever a user uploads. An AI inside the case file works from the full record set in the system, not a hand-picked excerpt of it. It can surface gaps in the treatment timeline, identify possible pre-existing conditions, and organize by provider, date, and relevance, all subject to attorney review. And it does this without someone first having to decide which pages to include. It still can’t know about records the firm never obtained, but it isn’t blind to the ones already in the file.
Demand drafting. A demand package depends on treatment records, billing data, lien information, liability documentation, and the specific facts of the case. When an AI tool operates outside the case, a lawyer has to assemble all of that, feed it in, and hope nothing material was left out. When the AI drafts from inside the case, it pulls from what’s already there. The draft reflects the actual record, not an excerpt of it.
Case Q&A. Attorneys and paralegals ask dozens of small questions in the course of working a case. When did the client first see the specialist? What’s the outstanding balance from the surgical group? Has the adjuster responded? An AI with access to the full case file answers these immediately. An external tool can’t answer them at all without being fed each piece of relevant data first.
Staff review. When AI-generated output is connected to its source documents inside the same system, a supervising attorney can verify a summary against the records it drew from in the same view. When the output was generated externally and pasted in, verification means going back to the source material separately, reconstructing the path the AI took, and hoping nothing was altered in transit.
Why adoption hasn’t matched the hype
Most PI firms are aware that AI tools exist. Fewer have integrated them into daily operations in any consistent way. The reasons are practical, not philosophical.
The first is conflicting signals. Depending on the source, firms hear that AI is non-negotiable, that AI is risky, that clients want it, that clients don’t trust it, that it saves hours, that it requires more oversight than it saves. When the message is contradictory, the rational response is to wait. And waiting, in this context, usually means individual attorneys experimenting with general-purpose AI tools on their own while the firm’s official position remains undefined.
The second is the ROI problem. Most firms that have tried AI tools can point to individual moments where the technology helped. Fewer can quantify the aggregate impact. That’s partly because AI used as a side tool creates scattered value, a few minutes saved here, a draft that came together faster there. The gains are real but diffuse. Without a system that tracks what AI contributed to a case outcome, the value stays anecdotal.
The ABA’s ongoing guidance on technology competence points to an obligation that’s growing clearer each year. Attorneys have a professional duty to understand the tools they use and to supervise how those tools contribute to client work. In practice, that duty is easier to meet when AI operates inside a governed system with clear inputs and auditable outputs. It’s harder to meet when AI is a browser tab someone uses at their discretion.
What belongs inside the workflow, and what doesn’t
Not every AI task needs to live inside the case management software. General legal research, broad writing assistance, and questions that don’t depend on case-specific data work fine as standalone tools. Those tasks are about the law, not about the case.
But the tasks that define PI practice are about the case. Medical chronology generation. Demand assembly. Intake qualification. Treatment gap identification. Lien calculations. These tasks depend entirely on data inside the case file, and they produce output that becomes part of the case file. Running them through an external tool and copying results back in adds a step that introduces error, loses context, and creates a version control problem that compounds across a caseload.
The dividing line is straightforward. If the AI needs case data to do the task, it should be inside the system that holds the case data. If it doesn’t, it can live anywhere.
CloudLex built Lexee AI on this principle. It operates inside the case file, with access to the full record. It generates medical summaries from documents already in the system. It drafts demands from actual case data. It answers case-specific questions by querying the file. And it keeps a human attorney in the review seat at every step, with source documents visible alongside AI-generated output.
For PI firms evaluating where AI fits into their practice, the first question is whether the tool can see the whole case. If it can’t, it’s working from fragments. And fragments, no matter how well summarized, produce incomplete work.
Ready to see it in action?
