AI belongs in personal injury practice. The firms getting real value from it are not the ones with the best tools. They are the ones that made four decisions everybody else skipped.

AI is the newest employee in your office. At most firms it arrived with no supervisor, no job description, no training, and nobody checking its work.

There is nothing new on that list. It is what you already do with every person you hire. The only difference is that when it is software, nobody thinks to make the decisions at all.

Walk into a personal injury firm today and you will find AI already in use in some form or fashion. Usually not integrated with their case management system. Usually in a browser tab next to it, and usually whichever AI the person has started to like and prefer.

An attorney uploads a bunch of medical records into a Claude project, to get oriented on a treatment timeline before reading four hundred pages himself. Another asks ChatGPT for a first pass at a motion. A paralegal runs medical bills and expenses through Gemini for a quick pre-litigation demand letter, or to see whether anything was missed.

Different people, different preferences, different tools, and not surprisingly, no consistent context and no quality check. And none of it is actually reckless. They are all doing what any smart professional would do when a useful tool such as AI shows up, which is to try it on their day to day work. And you know what, it actually works! That is the part worth sitting with. These one-off experiments are not failing. They are succeeding often enough that the habit forms, and before you know it, it sticks.

Let me be clear that I am not describing a problem here. I am describing an opportunity that arrived without a plan. We build AI, I use it every day, and I would much rather a firm be experimenting than sitting still. The experimenting is the good part.

What is missing is everything around this habit. Ask the firm which tasks these tools are appropriate for and which they are not, and there is no answer. Ask what happens to HIPAA protected data and information covered by attorney client privilege once it goes into public ChatGPT, and there is no consistent answer. Ask how anyone would know, a year from now, that a particular chronology or draft was written with AI help, and there is no answer, because there was no audit trail. Ask who trained the staff on any of this. Nobody did.

This is not a small firm or a big firm problem, and it is not a problem with firms that avoid technology.

Figures here are approximate and reflect what several industry surveys of legal professionals have reported over the past year. Very few firms report any formal training or oversight.

Which is to say: everyone at the firm is now an AI expert. Everyone has a tool they vouch for. Everyone is certain their prompts are the best. And nobody has agreed on anything.

The first instinctive knee jerk reaction would be to draft a policy document. Circulate it. Have everyone sign it. Plenty of firms are being told to do exactly that, and it is not bad advice. It just does not do very much.

Think about why. A policy is a handbook, and a handbook tells people what they ought to do. Now ask yourself what a handbook actually accomplishes on a new paralegal's first Monday. Honestly, not much. More than the new employee handbook and the policy documentation, what determines the success of your new paralegal is the decisions you make: who she reports to, which cases she is allowed to touch, what can go out without your review, and whether anyone sat her down and told her the ways, the norms and the no's.

If your firm is making those decisions with every new employee you onboard, then that AI handbook is almost decorative. And if you are not, then no policy handbook can properly onboard, orient and align anyone with your organization, your process and your technology.

AI is nothing but the newest employee in your office. And at most firms it showed up with no supervisor, no job description, no training, and nobody checking its work. Worse than that, unlike a new paralegal, it never comes to your door and says it is not sure. It just answers, with conviction and authority.

The tool itself is not really the problem. The same tool that produces something careless at one firm produces something solid at another. The second firm does not have better software. It has a better setup.

Which brings me to the claim this whole article rests on.

An AI policy document will not help you scale your firm. Four decisions will.

Who supervises it. Who is allowed to use it, and on what kind of work. What it is actually assigned to do, and where that sits in the case. What it can see, what it works from, and whether you can check what it did.

There is nothing new on that list. It is what you already do with every person you bring into the firm. The only difference is that nobody thought to do it for the software.

And the obligation behind it is not new either. The profession settled this long before any of this showed up. You are responsible for the work done in your office by people who are not lawyers. Software that drafts, summarizes, and organizes is doing exactly that kind of work. The duty is already there. What is missing is the setup that would let you meet it.

Ask the person you trust most with a file for a detailed summary of the case facts and the timeline. Then a detailed medical chronology. Then a motion. Now ask the one whose work you always check twice for the same three things on the same file. Three years ago you could tell them apart in thirty seconds. Today both come back organized, complete and professional looking. They read the same. Quite often they are the same.

That sounds like good news, and in one sense it is. The weakest work in your firm got noticeably better. People who were turning in thin work are turning in first drafts they could not have produced three years ago.

But two things came with it, and neither of them is good news. Look at what moved with it. Individual work product was how you knew who was good. It was how you decided who to put on the difficult case, who was ready for more responsibility, and who needed help. That signal is much weaker than it was, and most firms have not yet replaced it with anything.

The difference between those two people has not gone anywhere, of course. The one you trust knows which provider always undercodes, which gap in treatment defense counsel is going to attack, and which adjuster moves before mediation. None of that shows up in any of those three documents. It shows up in the case, usually much later, and usually at a bad moment.

That is the first effect. The second one is different, and it lands on the people you can least afford to lose, the newest ones who should become your best in ten years.

Your newest people cannot catch what the AI got wrong, because catching it requires exactly the expertise they have not built yet. And they may never build it, because the work that used to build it, sitting with the file until the case starts to make sense, is the work they just skipped. The inexperienced do not become experienced by watching good output arrive.

None of this is an argument against AI. It is an argument about how you put it in. The firms that can see the work, and not only the output, keep the signal. The firms that cannot, lose it without ever noticing it went.

For a firm trying to grow, this is the whole problem in one sentence. You scale by knowing who can do what. If everyone's work looks identical, you cannot staff the difficult case with any confidence, and you cannot tell who is ready for more responsibility.

And there is a bigger version of the same problem waiting outside your office. We compete on attention to detail and on our ability to recover a superior result through a strategy that belongs to us. What actually separates you is how the case gets built and who builds it. Now every firm has the same tools, and everybody's paper is starting to look alike. If you cannot see the difference inside your own firm, your client certainly cannot see it from the outside.

Now the reasonable response to all of this is to write a policy. So let me be specific about why that does not reach it. Here are three lines that appear in almost every AI policy. Each one is sensible. Each one falls apart the moment somebody tries to follow it.

First. An attorney reviews anything AI produces before it is used. Now try it. Your paralegal uploads a stack of records and gets back that chronology. Most tools will show her where each entry came from, and that is a real improvement over where we were two years ago. What no tool shows her is what it left out. Which records it never opened. Which visit it passed over. A missing entry does not announce itself, and she cannot review something that is not on the page. To catch it she would have to read all four hundred pages, which is the work she was trying to avoid. So review becomes a careful look at what is there and no look at all at what is not. Nobody broke the rule. The rule could not be followed.

If a new paralegal handed you that chronology and could not tell you what she had not gotten to, you would send her back to her desk.

We accept it from software because it arrives looking finished.

Second. The firm is accountable for the work product. Everyone agrees with that. But accountability means you can reconstruct what happened. Who produced this, from what, reviewed by whom, and when. Without an audit trail you do not have accountability. You have blame, available after the fact. You also have no way to get better at any of it, because you cannot see what is working and what is quietly creating rework.

Third. AI does not know where its own competence ends. Ask it something it cannot reliably do and it answers anyway, with the same conviction it uses when it is right. A new paralegal who is unsure comes and asks you. The software never will. A policy cannot fix that, because the policy is addressed to your people and the behavior belongs to the tool.

Three different rules, one failure. Every one of them asks a person to make up for something nobody set up.

Which is why the well run firms are not exempt. We have seen careful, thoughtful policies at serious firms with exactly the same gaps sitting underneath. Discipline was never the constraint. It was never a rules problem.

Much of what is written about AI and legal ethics is written from inside a commercial practice. Billable hours, corporate clients, document review at scale. That is a different world and the advice does not transfer as cleanly as it looks.

Start with our client. Something happened to them. They did not plan for this, they did not shop around for it, and most of them have never hired a lawyer before in their life. They are injured, often out of work, in pain, sometimes on medication that makes it hard to hold a conversation, worried about money in a way that has nothing to do with legal fees, and in the worst cases grieving.

Now compare that to a corporate client, who hires a law firm as a business decision, compares three of them, negotiates the rate, and moves on if it does not work out. Advice written for that relationship does not fit ours.

Then the economics. We do not sell hours. Time still matters, and it matters a great deal in the right places. Case data collected and processed as it comes in, rather than reconstructed in a scramble the week the demand is due, moves that demand out weeks earlier. A statute that gets calendared correctly is the whole case. A treatment gap caught in month three is a very different problem from the same gap discovered in month eleven. But speed is not the goal on its own, and in negotiation it is often the enemy. In a litigation that runs three years, it is perfectly fine if a well crafted motion takes a week. And there are plenty of cases where you need the other side to have real skin in the game before they will move seriously, which means letting them spend on experts and depositions rather than reaching for the first number they put in front of you. Sometimes it comes down to who blinks first.

The fastest route to a settlement is very often the worst one for your client.

Knowing when to move quickly and when to hold is strategy. And strategy is a call your people make, not one the software makes for them.

So the right question about any of this is not how much time it saved. It is whether the case is in better shape.

There is a second thing, and it is the part I think about most. Some moments in a personal injury case have to be human. Telling a client the offer came in lower than they were hoping. The call after a diagnosis got worse. The conversation with a family member who is calling because the client cannot. Those moments are the job. A good setup does not merely avoid them by accident. It knows which moments those are, it hands them to a person, and it keeps doing the work underneath so that person actually has the room to be present for them.

That is not something you add later. It gets decided when you define the job, which is one of the four decisions.

There is one more thing specific to our work. Personal injury is a records business, and the record is far bigger than the medical file. What we are really building is a case timeline, and it is the highest level view of the case. It carries the medical history and the treatment timelines, and it also carries the loss of employment, the interrupted education, the disabilities that came out of the injury, the property that was lost, the loss of consortium, and the way pain and suffering built over months and years. None of that happens in a single event. It develops, and the timeline is where you can actually see it develop.

This is where AI genuinely helps, and it helps more here than almost anywhere else in law. Thousands of pages, a dozen providers, arriving out of order over months. No human reads that faster or more consistently than a machine does, and pretending otherwise helps nobody. It is also exactly where it matters most what the AI actually read. A chronology built from your file, with every entry traceable back to the source it came from, is a different object from one built on a general sense of how these cases usually go. Both read well. Only one of them survives a defense expert.

And underneath all of it there is a story. Most of these cases never reach a verdict. They settle, before trial or during it. What actually moves the other side is the threat of what a jury would do with that narrative if it ever got in front of one, and a story that is coherent, credible and fully documented is exactly that threat. It is the difference between a case the defense prices to make go away and a case they price seriously. Compress it into a competent summary and something real is lost, even when nothing in the summary is wrong. The goal was never the shortest accurate version of somebody's life.

We have spent years listening to personal injury lawyers tell us how they got here, through the Trial Lawyer's Journal and the Celebrating Justice podcast.

The stories are not what you normally hear from other professions. A crime victim who became a lawyer because of what happened to them. A Marine who left the service and found this work at par with serving the nation. Someone with a deep passion for carpentry and building things with his handcraft finds the same artistry in seeking justice. And so many more. I encourage our readers to visit triallawyersjournal.com for more motivating examples and for these unsung heroes. What runs through all of them is that they are here on purpose, driven by something specific, and every one of them runs a case their own way.

That is not incidental to the work. It is the work. The vision, the drive and the strategy an individual lawyer brings is what makes one firm different from another, and it is why two firms holding the same facts reach different results.

So let me be clear about what we are not trying to do. We are not building AI to standardize that. We are not trying to democratize creativity or turn expertise into a template. That would take the one thing that genuinely differentiates this profession and average it out.

The democratization I described earlier happened by accident. Nobody chose it and it quietly cost you something. What I am describing now is the deliberate kind, and done properly it is exactly what lets a firm grow. Everything depends on where you draw the line between the two.

Your expertise is not one thing. It is the cases you have tried, the adjusters you know, the providers you do not trust, the instinct about when a case is ready and when it is not. That is the part that must never be standardized.

Here is the line, and it is the most important sentence in this article.

Standardization of the case groundwork is good. Standardization of your expertise is the harm.

The case groundwork is case data collection, information processing, analysis, and the day to day operations of the firm. Intake. Provider lists. Wage and employment records. Bills. Chronology assembly. Deadline tracking. Correspondence. Flagging a treatment gap, or a provider nobody has requested records from. None of that is where the human touch lives. All of it should be industrialized, and being slow at it helps nobody.

Your expertise is what the analysis means and what to do about it. Case valuation. Whether the damages have finished developing. Whether the adjuster in this venue will move. When to hold. Every moment where a client is on the other end of the conversation. That is what must not be industrialized, and standardizing it is where the harm shows up.

Everything worth building gets decided by which side of that line the work falls on.

Which is why AI should enhance the individual lawyer in exactly two ways.

It takes the mundane administrative and tactical work off the desk, so that creativity has more bandwidth to work with.

And it supports the lawyer with analysis, research and trends, so that the decisions they are already making are made against better information.

That is the whole of it. It does not make the decision, and it does not supply the vision.

And it is worth saying what we are building, not only where we draw the line. Intelligence that works from your own file, inside your own process, so that what comes out of it is more yours rather than less. That is a harder thing to build than a template, and we think it is the version worth building.

I keep calling AI it, rather than her or him. That is deliberate. It does an employee's work, so you make employee decisions about it. But it is not an employee, and confusing the two is how firms end up trusting output that nobody supervised.

Somebody has to own this by name. Not a committee. Not the managing partner in theory. A person with the authority to say what the firm does and does not do with AI, and who expects to be asked about it six months from now.

This sounds procedural. It is the decision everything else hangs on. Remember the two numbers from the beginning. Seven in ten people using these tools. One in three firms having decided anything about it. That gap is not a technology gap. It is an ownership gap. Until somebody owns it, every decision gets made by default, one browser tab at a time.

Then the same question you ask about any new hire. Who uses it, on which matters, and who has been trained.

Most firms are providing no training at all. And notice what training usually means when it does happen. Somebody demonstrates what the tool does well. That is the least useful thing you can teach. What your people need to know is how it fails, what it tends to get wrong in our kind of case, and what it looks like when it is confidently wrong. That is the only training worth the hour.

Train your veterans too, not just the new people. Your veterans are the only ones who can reliably tell when the output is wrong, and right now most of them are using these tools with less guidance than the juniors are.

Give it a job description. A defined input, a defined output, and a named reviewer.

Think about what you would actually say to a new paralegal on her first day if you put her on chronologies. It would take you about thirty seconds. Now write the same thing for the AI.

It builds the first draft of the case chronology. It works only from the documents in that case file. Anything it cannot source, it flags rather than fills. Every draft goes to the assigned paralegal before it moves downstream. It does not dictate a valuation or push toward a settlement. And it knows the difference between a routine client message and a conversation that needs a lawyer.

That is a job description. Most firms have never written one, which is why the software ends up doing whatever the person in front of it happens to ask, with no idea where the edges are.

A defined job can be measured, improved, audited and trusted. An errand somebody runs on the side in a browser tab can be none of those. The corollary is that every use has to earn its place by solving a real problem in the case. Not because it is impressive in a demo. Legal software is full of capability nobody in a firm ever asked for, and every piece of it costs your people attention.

This one has four parts, and these are the ones decided by engineers long before you ever see the product.

What it can see. Your data stays inside your environment. Nothing crosses to another firm and nothing from your matters trains anything for anyone else. Get that in writing. A vendor either commits to it or does not.

What it works from. This is the one that matters most and gets discussed least. A chronology built from your file, with every entry traceable back to the source, is a completely different object from one built on a general sense of how these cases usually go. Both read well. You cannot tell them apart in a demo. You find out much later.

How narrow the job is. A bounded job gets done well. An unbounded one gets attempted, confidently, including the parts it cannot do. Scope is not a limitation to apologize for. It is the control.

Whether you can check it. What ran, on what, producing what, reviewed by whom and when. This is the audit trail, and it is what gives you back the signal you lost. When you can see the work and not just the output, you can tell one person's thinking from another's again, and you can teach from it.

It does not force feed a case valuation and it does not push toward a settlement. Those are calls, not outputs, and a system that nudges toward either is quietly making the most consequential decision in the case on your behalf.

On client communication the boundary is not silence. It is discernment. A well built system can tell a routine communication from one that carries weight. A scheduling confirmation, a records status, an acknowledgment that something arrived, those are routine and there is no reason a client should wait on them. A conversation about a disappointing offer, a worsening diagnosis, or a family member calling on the client's behalf is not routine, and handling it in a brisk, automated, overly eager manner does real damage. Those conversations need empathy and they need a lawyer leading them.

Routine it can carry. The rest it hands over. That distinction has to be designed in, not hoped for.

None of these four are novel. You make all of them every time somebody new joins your firm. The only difference is that when it is software, nobody thinks to make them at all.

This is the work we have spent the last several years on at CloudLex, building AI into the system of record rather than beside it, working from your own case file, with experienced paralegal support sitting in the same place rather than in a separate vendor relationship. But the four decisions do not belong to us or to any vendor. You can make them well with tools we did not build, and badly with tools we did.

And this is what turns three years of individual cleverness into something the firm actually owns.

We are not going to tell you how to run your intake or how to build your demands. Every personal injury firm does those differently and there is no standard operating procedure in this practice to appeal to. Anything prescriptive we wrote would be wrong at most firms and stale within a year, because the tools are moving faster than that.

But there is one place where a short list does travel, and that is the conversation you have before you buy anything. None of this is a reason to be suspicious of vendors. It is a reason to be specific with them. Here are the six questions I would want answered before signing anything, including if the vendor were us.

1. What does your system work from. Our case file, or its general knowledge of how these usually read. Make them be specific. This is the single most consequential thing about any legal AI product and you cannot see it in a demo, because the output looks the same either way right up until somebody pushes on it.
2. Does anything from our matters leave our environment, and does any of it train anything, for us or for anyone else. Get it in writing, not in a conversation.
3. What is this designed not to attempt. A vendor who answers that quickly and without hedging has drawn real boundaries. A vendor who cannot answer it yet has probably not drawn the line, which means it gets drawn later, in one of your case files.
4. What record does it leave. What ran, on what, producing what, reviewed by whom, and when. If that answer is vague, then whatever supervision you owe, you will be owing it with no means to deliver it.
5. Where does it stop and hand the work back to a person, and how was that boundary decided.
6. When it is wrong, how will we find out. Not whether it will be wrong. When.

The signal is usually not the content of the answer. It is whether an answer exists at all, and how long it takes to arrive.

A vendor who answers the first question with a technical explanation you can follow is in a very different position from one who answers it by describing the benefits. Ask the same question twice, a week apart, to two different people at that company, and see whether you get the same answer. The delay, and the drift, is the finding.

None of this requires a policy document. It requires four decisions, made deliberately and written down somewhere your people can find them. If any of them has no answer at your firm today, that is not a failure. That is simply your next decision, sitting there waiting to be made.

Let me be clear about what is settled here and what is not.

We have spent over a decade building software for plaintiff personal injury firms, and the last three years building AI into it and watching how firms actually use it. The pattern does not vary much. Firms that decided who owns this work, what it is assigned to do, and how it gets checked are getting real value from it, and it compounds. Firms that decided none of that have a lot of individual activity and very little the firm can actually rely on. I am confident about that, and I do not think it is a close call.

What is genuinely still moving is the tooling. The models improve every few months, and what they can do reliably keeps expanding. Anyone who tells you they know exactly where that lands in five years is guessing, and you should discount them accordingly.

There is a great deal of gold rush energy in this market at the moment. I am enthusiastic about the technology and cautious about the rush, and I do not think those two positions are in any conflict at all.

Which is precisely why I would not build your approach around the tools. Anything written against what is available today expires with what is available today. The four decisions do not, because they are about your firm and not about the software. Supervision, training, assignment and verification hold no matter what shows up next year, and a firm that has answered them can adopt whatever comes without starting over.

I would rather be held to a standard than describe one. So here is what I think any firm should be able to require in writing from anyone building in this space, including from us.

• That your data stays in your environment, that it never crosses to another firm, and that nothing from your matters trains anything for anyone else.
• That the work is produced from your case file, and that you can see the source behind any part of it.
• That the limits are stated plainly and in writing, including what the system will not attempt.
• That every piece of work leaves a record you can retrieve yourself, without asking the vendor for it.
• And that the boundary between what the software handles and what a person handles is documented rather than left to the software's discretion.

Notice that every one of those is verifiable. A vendor either puts it in the agreement or does not. That is the difference between a commitment and a marketing claim.

That last one matters most to us. The person on the other end of a personal injury case did not choose to be there. They handed your firm the worst year of their life and asked you to handle it properly. Everything else in this article is engineering. That is not.

Here is what I keep coming back to. In five years every firm in your market will have access to the same tools. The tools are not going to be what separates you. What will separate you is whether your firm made those four decisions or never got around to them. That is the difference between a handful of people who are personally good at AI and a firm that is genuinely better at handling cases.

If your client could see exactly how their case was handled, start to finish, would you be glad to show them?

Make the four decisions and the answer takes care of itself.