AI Tools for Law Firms: What Actually Helps and What Is Noise

The AI conversation inside small and mid-size firms has gone somewhere odd. Partners are being pitched tools that promise to draft pleadings, predict outcomes, answer client questions at two in the morning and reduce headcount, and the pitch almost never starts from the thing that actually consumes a firm's week. Nobody in the demo asks how many hours the associate spent last month reading a medical chronology, or how many new enquiries went three days without a reply because the person who handles intake was in court. Those are the real losses, and they are boring, which is exactly why they get skipped.

I want to be specific rather than atmospheric about this. There are three places where AI genuinely pays for itself in a firm of five to forty people, and there is a much longer list of places where it is a solution wandering around looking for a problem, generating enthusiasm at the demo and quiet abandonment by week six. The dividing line is not how clever the model is. The dividing line is whether the task has a cheap verification step, because a lawyer has to check the output either way, and if checking takes as long as doing, the tool has produced nothing.

The other thing worth saying up front is that AI does not touch the operational spine of a firm. It does not make your trust ledger correct, it does not stop a walled user from opening a matter they should not see, and it does not keep a limitation date from slipping. Those problems are solved by structure, by systems that make the wrong thing impossible rather than merely discouraged, and no amount of language modelling substitutes for that. Keep the two categories separate in your head and the buying decisions get much easier.

Start with the task, not the tool

The firms that get value out of AI did not start by choosing a vendor. They started by writing down, honestly, what ate the week. A partner who tracks this for a month usually finds the same shape: a large block of time reading things in order to compress them, a smaller block producing first drafts of documents that follow a familiar structure, and a scattered block of administrative response that is urgent but not skilled. Those three blocks are where AI has something to offer, and none of them are what the marketing leads with.

Once you have the task list, judge each candidate task by one question. If the output is wrong, how quickly will you know? A summary of a deposition can be checked against the transcript in minutes because you can search for the passage it cites. A predicted settlement value cannot be checked at all until the matter resolves, by which point the prediction has already influenced your advice. The first is a usable tool. The second is a confident stranger in the room, and the fact that it sounds authoritative makes it worse, not better.

Drafting first passes, where the gain is real but narrower than advertised

Drafting is the strongest and most misunderstood use case. AI is good at producing the scaffolding of a document that follows a known shape, a client letter setting out options, a standard set of interrogatories, the recitals and definitions of a routine agreement, a chronology laid out from facts you supply. It is good at this because the structure is predictable and the failure modes are visible. When it invents a clause that does not belong in your jurisdiction's version of a document, an experienced lawyer spots it on the first read, and the correction takes seconds.

Where it stops being useful is the part firms hope will be automated, which is the judgement. The second draft, the one where you decide what this particular client needs and what risk you are willing to carry on their behalf, is not a text-generation problem. Realistically the gain is that you skip the blank page and the mechanical assembly, which is worth perhaps thirty to fifty percent of the drafting time on routine documents and close to nothing on the bespoke ones. That is a genuine gain, and firms that describe it honestly to their team get adoption. Firms that promise it will write the whole thing get a team that tried it twice, got burned, and went back to the precedent bank.

Summarising long records is the quiet winner

If you only adopt one thing, adopt this. Personal injury firms working through medical records, employment firms reading years of email, family firms with financial disclosure running to hundreds of pages, all of them spend enormous unbillable or heavily discounted time doing something that is closer to reading than to lawyering. A model that produces a first chronology, flags the dates and the actors and the gaps, and points you at the twelve pages that matter out of four hundred, changes the economics of that work in a way that clients actually feel in the fee.

The discipline that makes this safe is simple. The summary is a map, never the territory, and every fact you rely on gets traced back to the source page before it leaves the office. Set the expectation with your team in exactly those words, because the failure pattern is not the model hallucinating a document, it is a tired associate quoting a summary in a witness statement without opening the underlying record. Firms that keep the source document and the summary side by side in the same matter file, with a note recording who checked what, avoid the whole category of problem. In Casely, every document carries a comment field recording what changed and why, which is the natural place for that verification note to live so it survives staff turnover.

Intake triage, where AI works on the front end and then stops

New enquiries are the other honest use case, and the value is in speed rather than intelligence. Most firms lose work not because they assessed a matter badly but because nobody replied for two days. A model that reads an inbound enquiry, extracts the parties and the rough issue, tags the practice area, flags obvious urgency such as an approaching deadline mentioned in the message, and routes it to the right person, does something a human would do identically but eight hours later. That is a real commercial gain and it carries very little risk, because a misrouted enquiry is a mild annoyance rather than a professional problem.

Where firms overreach is letting the front end give answers. The moment an automated system tells a stranger whether they have a case, quotes a fee, or discusses the merits of their situation, you have created questions about what was said, what the person reasonably understood, and whether a relationship was formed, and the answers to those questions vary by jurisdiction. Rules on unauthorised practice, advertising and the formation of a lawyer client relationship differ substantially between US states, England and Wales, the Canadian provinces and the Australian states, so confirm your own regulator's position before you put anything conversational in front of the public. Triage and route, yes. Advise, no. The extracted enquiry should land as a real record, and Casely's conflict checking then searches the full contact and matter history, every role a party has played, including closed matters, before anyone opens a file.

The verification step is the whole job

Every serious discussion about AI in practice comes back to one thing. The lawyer signing the document is responsible for it, whatever produced the first version, and courts in several common law jurisdictions have already sanctioned lawyers who filed material containing citations that did not exist. That is not a reason to avoid the technology. It is a reason to design the workflow so that verification is structurally unavoidable rather than a matter of individual diligence on a bad Friday.

In practice this means two rules that a firm can actually enforce. First, nothing produced by a model goes to a client, a court or an opposing party without a named person having read it against the source, and the file records who that was. Second, citations and figures get checked in the original, always, without exception, including when the draft looks perfect. The second rule matters more than people expect, because polished output triggers less scrutiny than rough output, which is precisely backwards. Build the check into the matter workflow so it is a step someone completes, not a habit you hope survives a busy month.

What AI does not fix, and never will

Here is the part vendors skip. The problems that actually end careers in this profession are structural, not cognitive. A trust account that goes overdrawn because a disbursement was made against money that had not cleared is not a knowledge problem, it is a systems problem, and the fix is a system that refuses the transaction rather than one that offers advice about it. Casely blocks any disbursement exceeding a matter's actual trust balance at the database transaction level, not as a warning dialog someone clicks past, and corrections are voided and stay visible rather than being deleted. No model produces that guarantee, because the guarantee is about enforcement, not intelligence.

The same holds for conflicts and confidentiality. An ethical wall that exists in the interface is not a wall, it is a suggestion, and the test is whether a walled user can reach a restricted matter by any path at all, including a search result, a calendar entry or a link a colleague forwards without thinking. Enforcement at the server and data access layer answers that question. A smart assistant sitting on top of a permissive system answers nothing, and arguably makes things worse by putting a fast retrieval tool in front of data that was never properly partitioned in the first place. Get the structure right first. Then add the assistant.

Confidentiality is a product question, not a policy question

Every firm evaluating AI eventually writes a policy telling staff not to paste client information into public tools. That policy is necessary and it is also insufficient, because it relies on a busy person remembering a rule at the exact moment the rule is inconvenient. The more useful move is to answer the product questions concretely. Where does the text go, who at the vendor can read it, is it retained, is it used to train anything, and what happens to it if you stop paying. If the vendor cannot answer those in writing, the evaluation is over regardless of how good the demo was.

Then check that the rest of your stack holds up its end. Document storage should be encrypted at rest with a per firm key rather than a single shared key across every customer, so that a compromise of one tenant does not expose yours, which is the standard Casely holds to with AES-256 encryption on every document. Confidentiality obligations themselves vary in their detail across jurisdictions, and some regulators have issued specific guidance on generative tools while others have not, so read your own rules rather than a summary written for a different country. The general shape is consistent everywhere, which is that you remain responsible for client information no matter which vendor is holding it.

The long list of things that are mostly noise

Outcome prediction is the clearest example. Tools that claim to forecast how a matter will resolve are selling certainty into a domain that does not have any, and the output cannot be verified before you have already relied on it. Automated legal research that returns confident answers without traceable sources belongs in the same category, not because research assistance is worthless but because an answer you cannot trace is an answer you have to redo. Client facing chatbots that discuss merits, AI generated marketing content nobody reads, and dashboards that score your lawyers on invented productivity metrics all share the same defect, which is that they generate activity rather than removing it.

There is a subtler category of noise too, which is AI bolted onto a task that was never the bottleneck. Automatic billing narratives are the classic case. Firms do not lose money because narratives are badly written, they lose money because hours never got recorded at all, and a model that rewrites the entries you already captured has improved the wrong number. The same goes for AI that summarises documents nobody was reading anyway, or drafts emails that took ninety seconds to write. Ask what the task cost you last quarter in hours or in fees. If you cannot name a figure, the tool is not solving a problem you have.

Where the operational gain is actually larger

This is uncomfortable for a post about AI, but most small firms have more money sitting in workflow than in intelligence. Time that never made it onto a matter, invoices that sat unissued for six weeks, deadlines tracked in a personal calendar, clients calling for updates because they have no other way to get one. Those are large, measurable, recurring losses, and they are fixed by structure rather than by cleverness. One click invoicing that turns every unbilled hour into a single itemised draft removes a delay that costs real cash flow, and a deadline diary that attaches dates to the matter with next date tracking removes a risk that costs considerably more than cash.

The client side is the same story. A privilege filtered portal where a client can see status on their phone and sign a document inside the same login, with no separate account to create, eliminates a large share of the update calls that interrupt fee earning work all week. A matter stage tracker that matches how your firm actually runs, configured per practice area rather than imposed by a vendor, tells everyone where a file stands without anyone asking. None of that is AI, all of it is measurable within a month, and a firm that fixes it first will get far more out of AI later because the underlying data will finally be clean enough for a model to be useful on.

How to run a trial that tells you the truth

Two weeks, one task, one person, and a number written down before you start. That is the whole method. Pick the single task that costs you the most hours, name the person who owns the trial, record what that task took last month, and then measure it again with the tool in place including the verification time. Verification time is the number firms forget, and it is the number that decides whether the tool is worth anything, because a draft produced in nine seconds and checked for forty minutes has cost you more than the precedent you already had.

Kill it decisively if the number does not move. The sunk cost pattern in legal tech is brutal, a firm signs an annual contract, three people use it in month one, nobody uses it by month four, and it renews anyway because cancelling requires someone to admit the decision was wrong. Build the exit into the trial by agreeing in advance what result means you stop. And run the trial on real matters with real records, not on the vendor's sample data, because sample data is chosen to make the tool look good and your files are not.

What your supervision policy needs to say

A one page policy beats a twenty page one nobody reads. It should name which tools are approved and which are explicitly not, state that no client identifying information goes into an unapproved tool, require that output leaving the firm has been checked against the source by a named person, and say where the record of that check lives. Add a line making clear that the responsibility for the work sits with the supervising lawyer regardless of what produced the draft, because that is the point people forget when they are under pressure.

Then check it against your own regulator, because this is genuinely a moving area. Professional conduct rules on competence, supervision of non lawyer assistance, confidentiality and client communication all touch AI use, and the guidance issued so far differs meaningfully across US state bars, the regulators in England and Wales, the Canadian law societies and the Australian state bodies. Some have addressed disclosure to clients directly, some have not addressed it at all, and a few courts have imposed their own filing requirements independent of the bar rules. Confirm what applies where you practise before you write the policy, and revisit it, because what was accurate a year ago may not be accurate now.

The honest scorecard

Strip away the noise and the picture is unglamorous but useful. AI will save your firm real time on summarising long records, will give you a running start on routine drafts, and will stop enquiries sitting in an inbox over a weekend. It will not make your trust accounting compliant, will not enforce an ethical wall, will not catch a limitation date, and will not fix a firm whose problem is that time never gets recorded. Buy it for the first list, and never let it be the answer to the second.

The sequencing matters more than the shopping. Firms that put the operational structure in place first, matters, deadlines, trust, conflicts, billing, documents, and then layer targeted AI on top of that, get compounding value because the model is finally working on organised information rather than on a pile of email attachments. Firms that buy the clever tool first usually discover that its output has nowhere to live and no one accountable for checking it. If you want a concrete starting point, look at what document management and automation actually does to your drafting week before you evaluate anything that calls itself intelligent, because the ordering of those two decisions determines whether the second one pays for itself.

Casely is cloud native with nothing to install, and there is a free plan, so testing this properly costs you a fortnight of attention rather than a budget line. Pick the one task that hurts most, measure it, and let the number decide. That is a less exciting answer than the one the market is selling, and it is the one that leaves your firm better off at the end of the year.