AI model selection for lawyers is not a technical question. It is an assignment question, the same one you run every morning when you decide whether something goes to a paralegal, an associate, or you. Most complaints I hear about AI for law firms are assignment errors wearing a costume. Wrong model, wrong surface, wrong task. The output came back bad because the request went to the wrong worker.
Product names change every few months, so I teach this by capability tier instead. The most capable model, the mid tier, and the fastest and cheapest. Whatever the vendors are calling those this quarter, the logic underneath holds.
The most capable tier gets anything with numbers or authority in it
Building systems and skills. Deep income analysis. Careful discovery review. Child support math. Anything touching citations. Those go to the most capable model and I pay the token cost on purpose. Because this is about case citing, I am going to use the most capable model, because it takes more tokens or credits, but it is more in depth. And anytime I am doing research stuff, I use that one. Depth is the point on those tasks. No attorney has ever been grateful that a support number came back fast and wrong.
Citations are the sharpest case for the top tier, and also the case where the model matters less than the fence around it. The architecture that keeps a research tool from inventing authority is a separate discipline, and I laid it out in the research library that refuses to do research.
The mid tier does essentially all of the drafting
Pleadings, retainers, correspondence. All mid tier, all day. That surprises attorneys who assume the most expensive model must write the best petition, and it is where AI model selection for lawyers gets counterintuitive. The top tier does not write a better pleading, because a legal drafting AI setup is not being asked to think hard. It is being asked to apply a template, a format standard, and the firm’s own language to a fact pattern a human already framed. That is execution, not analysis, and the mid tier executes well at a fraction of the cost.
What does eat time in drafting is format, not substance. The formatting is the hardest part. It really is not ever perfect. So I stopped fighting it case by case and pushed it down into a standing instruction that runs on every session. One of my ambient rules is that anytime you are drafting any kind of a legal document, it has to be all black twelve point Times New Roman. That just runs on everything I do, all the time. An ambient rule is not a prompt you remember to type. It is a condition of the environment, which means a rushed associate at four thirty gets the same output standard you would.
The fastest tier reads. It does not draft.
The cheapest tier is genuinely useful and has one hard boundary.
You do not ever want to ask the fastest model to draft any kind of legal pleadings for you because it looks bananas. It does not understand formatting at all.
What it is good for is comprehension work. Reading a long transcript. Producing a memo from a pile of notes. Answering what is going on in this case when you have not touched the file in six weeks. Fast, cheap, and adequate for telling you what a document says. Never for producing one that gets filed.
Surface selection matters as much as AI model selection for lawyers
The second decision is where the work runs. AI model selection for lawyers is only half the answer if the session itself is wrong. Chat is for simple tasks and for anything you need to download and hand to somebody. The folder-connected working mode is for reorganizing files, for discovery work, and for anything that has to read and write to the drive, because output there saves back into the matter folder automatically instead of landing in a browser tab. Anytime I am doing deep financial digs, I typically do it as a folder-connected session. I just find that it does a little bit better with deep financial stuff. Almost everything else I do in chat.
Two mechanics decide this more than preference does. In chat you can stop a response mid-stream, switch models, and re-run, which makes a wrong choice cost you thirty seconds. In the folder-connected mode the model is locked for the life of the conversation, so a wrong choice means starting the whole thing over. That asymmetry is the reason to think for a moment before you open a session rather than after.
Then there is the evidence. A deep financial analysis run in chat mode, on a deadline.
We waited seven minutes, and then it just shut down and didn’t save any of its progress.
Seven minutes of compute, gone, with nothing written anywhere. That is not a preference about which mode feels nicer. It is a workload that did not match its surface. The rule that falls out of it is the living-document rule. Trackers, command centers, running chronologies, anything that updates across the life of a case belongs in the folder-connected mode, because saving writes state back into the matter file instead of producing version eleven on somebody’s desktop. Which only works if the matter file is built correctly in the first place, and most AI failures in a law firm are file structure failures.
Law firm AI costs are a training problem, not a pricing problem
The bill nobody plans for comes from staff, not strategy. If you have ten staff members that are not trained well, and they keep using the most expensive model for dumb things or simple things, you are going to have a very, very big bill. Ten people summarizing two-page letters on the top tier will outspend all your real work, and none of them will know they did it.
Two fixes, both structural. The first is an ambient rule that auto-selects the model by task type, so the choice is made by the system rather than by whoever is at the keyboard. The second is an ambient skill that interrupts before the work runs and recommends a better model or a better surface, which is the one I prefer for a firm in its first ninety days, because it teaches while it saves money. Either way, AI model selection for lawyers becomes a management discipline enforced by the system rather than a personal habit, and you control law firm AI costs at the configuration layer instead of by sending a reminder email nobody reads.
One more filter, and it saves the most money. If you can do it faster yourself, then do not use AI on that task. The thing about it is it can do this while you are doing something else. You cannot do something else while you are doing this. That is the test. Not whether the model can do it, but whether handing it off buys you back time you can bill or time you can leave early with. Firms that get this right tend to have already worked through the sequence I describe in the five levels of AI in a law firm, because model selection stops being confusing once the levels underneath it are stable.
Every suite in the product catalog ships with these assignments already made. Model tier and surface set per task, tested inside an active Florida family law firm on real cases and real deadlines before anybody else got them. You own it, it is yours, and nobody on your staff has to guess which model to open on a Tuesday afternoon.