The 9-Step Framework for Deploying Agentic AI in a Law Firm

Here is the part nobody says to the founder of a law firm out loud. Somewhere between your first handful of matters and a full caseload, you stopped being the lawyer and became the operating system.

Every new instruction, every court date, every "have we billed that time", every client chasing an update on where their case has got to waits on your attention, which means the whole practice runs exactly as fast as one senior partner can read.

This is the framework we use to take that job off the partner and give it to a system, without the risk that comes with letting software near client matters, court deadlines and money held on account. It is nine steps in three phases.

Four before you build anything, three to build it, and two to do the thing the whole exercise is for, which is to remove yourself from the middle of every matter.

It is worth saying what this is not. It is not a new case management package, and it is not "let AI practise law".

The judgement stays with your qualified lawyers. What changes is that the reading, the filing, the chasing, the diary management and the first-draft work stop landing on one desk, so a limitation date never slips because somebody was buried in disclosure the week it fell due.

You quietly became the operating system

It happens slowly. In the early years you took the instructions, drafted the documents, ran the hearings and raised the bills yourself because there was nobody else, and you were good at it, so the practice grew.

Then it kept growing and the routing never left your head, because it was faster to just do it than to write it down. Now you are the one who remembers a defence is due, who knows a client still has not sent their witness statement, who spots that a full interim hearing somehow never made it onto a bill.

None of that is written down anywhere. It lives in your attention, and in a law practice attention is the scarcest thing there is.

Three things follow, and every one of them costs the firm real money and, worse, real exposure.

• The practice runs at reading speed. A partner spending two or three hours a day reading correspondence, chasing documents and checking who is on top of which matter is normal. That is not fee-earning work. It is triage, done by your most expensive fee-earner.
• Deadlines get met because somebody remembered. Which is fine until the one busy week nobody did. A missed filing deadline is a strike-out risk, an application for relief, an awkward call to the client, and in the worst case a limitation date gone that turns a live claim into a negligence exposure against your own firm.
• Nothing survives your holiday. A practice where one partner is the router does not pause when that partner is away. Documents pile up unchased, time goes unbilled, clients go quiet, and you find out when you get back.

The instinct is to fix this by hiring another fee-earner. That works, and it also adds salary, supervision, and one more person who has to learn all the unwritten rules you were already the only keeper of, from the conflict-check habits to the way this firm always prepares a bundle.

You have not removed the bottleneck, you have given it a second head to depend on. The alternative is to write the routing down and let a system run the parts that never needed a qualified brain in the first place.

That is what the next nine steps do.

For a law firm, the client experience is the marketing

One belief before the framework, because it decides how you read the rest of it. In a professional practice, the client experience is the marketing.

Not the website, not the crest on the letterhead. The document served comfortably before the deadline, the question answered the same day, the update that landed before the client had to ask where their matter had got to.

That is what gets a firm referred, and it is produced almost entirely by the nine jobs we are about to automate.

This is why it belongs on a growth marketing site rather than in an IT catalogue. A missed deadline is not an admin slip, it is a complaint, a fee write-off and churn you are about to pay for.

Time worked on a matter and never billed is not a finance oversight, it is revenue you earned and then gave away. A client who cannot tell whether anything is happening on their case does not wait patiently.

They call a friend, ask who they use, and drift to a firm that simply answers. When the operating system is one tired partner, the client experience frays in exactly the places a client notices and a competitor asks about.

Which brings up the trap most growing firms fall into. They try to grow by taking on more matters, into a practice whose delivery still runs through the partners.

That does not produce profit. It produces missed deadlines, later nights and slipping quality, and in this profession slipping quality has a regulator attached to it.

This is the capacity problem, and it is one of exactly three things almost every stuck firm is stuck on. The other two are getting the right clients in and converting them, and I have written the full diagnostic in the 3A Machine.

So the goal of deploying agentic AI in a law firm is not "use AI to do the lawyering". It is to build the capacity that lets you grow without the quality falling over and without a deadline going unwatched.

Get the nine steps running and the same partners and the same fee-earners hold two or three times the caseload without a single court date slipping. Fix capacity first, then go and win the clients.

The nine steps, in three phases

Here is the whole map on one page. Your firm already runs all nine of these jobs today, whether or not anyone has ever named them.

The framework does not add work. It names the jobs, then decides one at a time whether a person does each one or a system does it.

Notice the shape. Four steps happen before anybody builds anything, and skipping them is the single most common reason a firm's "let us try some AI" project quietly dies.

People buy the clever part first, point it at nothing in particular, and end up with a fast tool that does not know how the practice actually runs, how this firm checks conflicts, how it prepares a bundle, when it will not file without a second pair of eyes. The order below is designed to stop that.

Step 1. Diagnose: find what is actually bleeding

You cannot fix what you have not counted, and most firms have never counted this. So the first step is a leak audit, and its only job is to put a number on the money and the risk already walking out of the firm.

Not a survey of how everyone feels. A count of losses.

In a law practice the leaks are always in the same few places. Deadlines that were met by the skin of the diary, or nearly missed, and the near-misses nobody logs because "we caught it in time".

Time that was worked and never billed, and in a firm this is almost always larger than the partners guess, because interim hearings, completion calls, quick advice notes and "just a five-minute question" from a client rarely make it onto a bill. The hours the team burns every week chasing clients for a signed statement of truth, a witness statement, an exhibit, an estate document, and re-typing details from a PDF into the matter file.

And the quietest leak of all, clients who drifted away not because the advice was wrong, but because they could never tell what was happening on their matter and a slicker firm made them feel looked after.

Add it up honestly and the total is almost always bigger than expected, and concentrated in two or three places rather than spread evenly. In most firms the two biggest numbers are the unbilled work in progress sitting in the system right now and the fee-earner hours lost to chasing documents.

That concentration is the gift. It tells you exactly where to point the build first, and it becomes the before number you use, at the very end, to prove the system paid for itself.

Step 2. Shadow: follow one live matter from instruction to filing

Now you watch. Pick one live matter, a commercial dispute, a probate, a lease, an employment claim, and follow it end to end, writing down every single thing that happens to it and every decision anybody makes.

Not the tidy version in the office manual nobody opens. The real one, from the day the instruction lands to the day the document is filed and the bill goes out.

This is where you discover the practice does not run on the written process. It runs on a hundred unwritten rules that live in your senior fee-earners' heads and, more than anywhere else, in the email threads and the case-file notes.

Which documents have to be in before a defence can be served. How the conflict check actually gets done, and by whom.

The way this firm always formats a bundle, and the practice note that changed the filing window this term. The client type you never file for without a partner review.

None of it is written down, all of it is load-bearing, and all of it is the difference between a matter that goes in clean and one that gets refused at the door.

The excavation is the real work of this step. You read back through the threads and pull out the rules the firm actually operates on, the ones that were never a decision, just a habit that turned out to be right.

That messy, lived-in reality is what you are about to encode. Skip it and you will automate the fantasy version of the practice, ship it, and watch the seniors quietly go back to doing it their own way because the system does not know what they know.

And be honest about the fragility this surfaces: in most firms the real operating manual is one long-serving litigator's memory, and it walks out of the door the day they retire.

Step 3. SOP: write the unwritten procedure down

Everything you excavated now becomes one written procedure. Plain language, step by step, in the order it really happens, with every rule and every exception stated.

This is the least glamorous step in the framework and it is the one that makes an agentic system possible, because a matters controller can only work a procedure that has actually been written.

For serving a defence it looks like a hard checklist: the signed statement of truth, the client witness statement, the exhibit bundle, the disclosure list, counsel sign-off, and a recorded conflict check, and the matter does not advance to filing until every one of those six is present. Written as a procedure, the messy reality becomes a gate a machine can enforce perfectly, every time, without a tired fee-earner waving something through on a deadline day because the diary was on fire.

The same goes for opening a new matter, taking probate instructions, or preparing a tribunal bundle. Each becomes a written sequence rather than a thing your best people simply "know".

Write it for a smart trainee on their first week, not for a machine. If a person could follow it without asking a single question, a system can run it.

If it still needs somebody to "just know" which exhibits go where, the procedure is not finished, and you are not ready to build.

Step 4. Baseline: agree the numbers before you touch anything

The last step before the build, and it is thirty minutes that saves you a year of arguments. Write down the current numbers and get the partners to agree them.

How long it takes to file a document into the right matter today. What share of court deadlines actually get met comfortably rather than at the wire.

How much work is sitting delivered but unbilled right now. How fast the first reply to a client goes out.

Real numbers, honestly rounded, on the record.

You do this for one reason. In three months, when the system is running, memory rewrites history.

People forget how bad it was, decide it was "always basically fine", and start to wonder what they are paying for. The baseline is the receipt.

It is what lets the weekly proof report at the end say "filing a document to the right matter went from about twenty minutes by hand to about two" and have that land as a fact both partners agreed up front, rather than a vendor's claim.

It is also the moment to decide what "better" means for this firm, so the build aims at a number and not a vibe. A practice bleeding on unbilled WIP is not chasing the same win as one whose fear is a missed limitation date.

Name the number now. It is what everything you build next is pointed at.

Step 5. The command center: the whole practice on one screen

Now you build, and the first thing you build is the place the partner looks. One screen that shows the whole practice at a glance, in the order that costs money and risk, so nobody reconstructs the state of the firm from six systems, a shared diary and a group chat every morning.

Build it on one spine, not twelve tools. This matters more than it sounds.

Every job in the audit is tempting to solve with its own separate app, a diary system here, a document tool there, a billing package, a client-messaging add-on, and a year later you have a dozen subscriptions that do not talk to each other and a partner who is now the integration layer between them. Build the whole thing on one foundation, one place the data and the context live, and the next piece is nearly free because everything it needs already exists.

The centre of the screen is a matter radar: every live matter, ranked not by the date it was opened but by which deadline bites first. On a full caseload that is dozens of matters running at once, and the radar puts the defence that is 48 hours from a filing deadline at the top, in red, whether or not anybody asked.

Around it, the numbers a partner actually needs, the matters that need action today, how many deadlines are at genuine risk, how much work is delivered but unbilled, how many limitation dates fall in the next month.

Then the limitation clocks. A practice does not have one deadline, it has dozens running at once across every matter, and a shared diary flattens them into a comforting list that hides the one about to end a case.

The command center pulls them apart and shows each on its own, so the filing deadline for the defence, the limitation date on a claim that is weeks from being time-barred, and the disclosure due before a bundle must be served are three separate clocks, and the dangerous one is the one you see. This is the screen that matters most, because in law a missed date is not a penalty you can pay off.

It can be a claim you can never bring, and a negligence exposure against your own firm for letting it lapse.

Beside the clocks, the money on the floor: every piece of work delivered, done, advised, attended, and never billed. In most firms this is a genuinely uncomfortable number the first time it appears on a screen, tens of thousands of dollars of work in progress sitting across a dozen or more matters, because the work was already paid for in fee-earner hours and simply never got converted into a bill.

It was not strategy. It was that raising the invoice was a small annoying job nobody owned while the next hearing loomed, which is a fair description of most of what leaks in a practice.

And the document wall. An established practice is sitting on an enormous pile of matter files, hundreds of thousands of documents across folders nobody has fully mapped in years, and a single heavy litigation matter can hold well over a thousand files on its own.

The command center reads that existing storage where it already lives, with no migration and no "please move every matter into our new system", and makes it navigable and countable, so any document is found in seconds instead of a frightened hunt through a bundle while a client, or a judge, waits.

One design choice worth naming, because clients always ask. The command center is owner-locked, and the team gets read-only logins scoped to what they need.

A partner sees the fees and the whole board. A fee-earner sees the matters assigned to them.

Nobody can quietly change a deadline or a figure they should only be reading, and the partner never loses the single honest view of the practice the whole build exists to give them. In a profession with client confidentiality and conflict rules to honour, that scoping is not a nice-to-have, it is a condition of the thing being allowed to exist at all.

Step 6. The procedure as data: the firm's mind

The procedure you wrote in step 3 is a document, and a document just sits there. In step 6 you turn it into data the system can reason over: every rule, every court deadline, every checklist, every practice note, and every good precedent the firm has ever produced, connected by meaning rather than filed in folders.

This is the firm's mind, and it is what makes the matters controller in the next step sound like your practice instead of a generic model that half-remembers the law.

In practice it means anybody can ask a plain question and get the firm's own answer, not the internet's. What must be in the matter file before we can serve a defence.

What changed in the latest court practice note and does it move any of our filing windows. What is our conflict-check procedure for a new corporate instruction.

The answer comes back grounded in the firm's own written procedure, with the source step it came from, so it is checkable rather than a confident guess, which in this profession is the whole difference between a tool you can rely on and one that gets a fee-earner into trouble.

The reason this matters more than it looks is drift. A general model, asked the same procedural question twice, will happily give two confident and slightly different answers, and in a regulated profession that is not a quirk, it is a liability with your name on the file.

Grounding every answer in the firm's own written procedure kills it. The system is reading your rules and quoting them with the source, and when a court practice note changes the bundle format or the filing window, you change it in one place and every answer changes with it, and the limitation clock updates to match.

It is also where the compliance checklist becomes enforceable rather than advisory, because the matters controller in the next step reads its gates from exactly here.

Step 7. The matters controller: reading and filing that shows up on time

Now the part people picture when they hear "AI". An employee, not a chatbot, and the distinction is the whole thing.

A chatbot waits to be asked. An employee wakes on a timer, reads the live practice, does its round, files the work, and flags what needs a human, whether or not anybody prompted it.

I have written the longer version in AI employees, not chatbots.

In a firm it is a small crew of them, each with one job, all reading from the firm's mind. A matters controller that reads an incoming document, a signed client care letter, a witness statement, an exhibit, understands what it is, runs the conflict check, renames it to the firm format and files it to the right matter.

A refusal fence that will not let a matter move to filing until every required item is present, the statement of truth, the witness statement, the exhibit bundle, the disclosure list, counsel sign-off and a recorded conflict check. An invoice filler that drafts the bill for delivered work straight from the recorded time on the matter file.

And a talking agent that answers a client's status question, in their own language, without pulling a fee-earner off a hearing.

There is a quieter agent in this crew that partners never ask for and always end up valuing most: a critic. Before any piece of work reaches a human, a second agent grades it against the procedure.

Is the document filed to the right matter, is the bundle in the firm format, does the draft actually contain the six items the SOP demands, did it do what was asked. Anything that fails goes back to be redone before you ever see it.

This is the line between AI you can run and AI you can trust in a practice where a mistake has a court, a regulator and a client on the other side of it.

An employee that can act is useful and it is also dangerous, so this step ships with four safety rules and they are not optional. They are the reason you can hand a system this much and sleep.

• It refuses out-of-procedure work. If a document or a request does not match the written procedure, the matters controller declines and escalates rather than improvising. It will not serve a defence with an exhibit missing, no matter how close the deadline. The refusal fence is a feature, not a failure.
• Every file operation stays inside set boundaries, and deletes go to a recycle bin. It cannot reach outside the matters and folders it was given, and nothing it removes is ever gone. A document filed to the wrong matter is always recoverable, and confidentiality between matters is never crossed.
• Every client-facing message waits for a human yes. It drafts the update, the chase, the bill, the retainer top-up request, then stops. A person reads it and presses send. Anything that touches a client, their money on account or the firm's name gets a qualified human on it.
• Ambiguous dates come back empty, never guessed. If it is not sure when a limitation period runs out or which court window applies, it says so and asks. A blank is safe. A confident wrong date on a matter is exactly how a system does real and irreversible damage in this profession.

Step 8. Hand it to the team

The build is running. Now it stops being the partner's private tool and becomes how the team works, and the shift is subtle but total: people move from performing the work to approving it.

The system does the first pass of everything, the filing, the chase, the draft bill, the status update, and a fee-earner says yes, changes a line, or sends it back. Same team, far more matters handled, and your qualified people are left doing the advocacy and the judgement that was always the actual job.

That handover is made of a few specific things. A daily status drafter that writes the update a client is owed, ready for a fee-earner to approve.

An activity wall so a partner can see what the system and the team did on every matter without asking. A change watcher that notices when a matter stalls at the same stage for days, or a court date moves.

And the piece everyone feels first: the morning brief.

At eight in the morning, before anyone opens anything, one message lands in order of what costs money and risk: here are the two matters hours from a deadline, the defence due in 48 hours and a limitation date eight days out, here are the three retainers that need topping up inside 45 days with the requests already drafted, here is the work in progress across fourteen matters waiting to be billed, and here is the probate stuck four days because the estate documents never arrived. The team walks in already knowing the day instead of spending the first hour discovering it.

A word on the human side, because it decides whether any of this sticks. The word removal frightens a team, and if they think it means removing them, they will quietly starve the system of the knowledge it needs, and in a firm that knowledge is the whole asset.

It does not mean that. It means removing the filing, the chasing, the document hunts and the diary-watching, and leaving people with the advocacy, the advice and the client relationships that were always the fee-earning work.

Say it out loud, early and often. A team that believes the system is on their side will feed it.

A team that fears it will fight it, and win.

Step 9. Removal is the whole point

Here is the step everyone gets wrong, and it is the reason the framework exists. Step 9 is not training.

It is removal. The goal was never to teach the partners to use a clever new tool.

It was to take the operating-system job out of the partner's head, so the practice stops running at the speed of one person's attention.

You know you have reached it by a specific test: a partner can take a genuine week off and nothing leaks. Documents still get filed to the right matter.

Court deadlines still get met. Limitation dates still get watched.

Bills still get raised on delivered work. Clients still get answered in their own language, the same day.

Nothing waits for one person to come back and look, because that person is no longer the thing the routing runs through. They approve the exceptions from their phone, or they do not, and it holds either way.

And this is where the baseline earns its keep. Every Monday a proof report goes to the partners and says, in the numbers they agreed at the start, what the system did this week.

Documents read, renamed and filed to the right matter in minutes instead of the afternoon. Every court deadline met on time.

Zero limitation dates hit unwatched, every clock caught in time. Dozens of status questions answered without a fee-earner touching them, and the hours the team used to spend chasing files and typing by hand down to nothing.

It is not a dashboard somebody has to go and check. It is the system reporting to the owner, unprompted, on whether it is still earning its keep.

And then the question that is really the point: what does the partner do with the attention they just got back. In every firm the honest answer is the same, and it is why this belongs on a growth marketing site.

They go and do the work only a partner can do. Winning the right clients, the referral relationships that carry the real margin, the higher-value matters, the next hire, the next office.

Removal is not the end of the partner's involvement. It is the first time the partner gets to actually be the owner of the firm rather than its busiest fee-earner.

What it costs to run

This is the question every partner asks within about ninety seconds, and the honest answer surprises people in the right direction. It is far less than they expect, and far more transparent, because the running cost is metered and shown like a utility bill rather than hidden inside a flat fee.

For a working system on a real firm's caseload, the AI usage runs on the order of a few dollars a day, and the hosting and database are a few tens of dollars a month, so the whole thing comes to roughly a hundred dollars a month to run. Your volume will differ, which is exactly why the meter is built in from day one and shown to the partners on day one.

No hidden meter, ever.

Now put that next to what it replaces, with round numbers you can redo with your own. Say a partner and the team lose two and a half hours a day to chasing documents, re-typing details into matter files and building status updates by hand, across twenty two working days.

That is fifty five hours a month. Value that time at even a modest fifty dollars an hour, low for qualified fee-earners, and it is over two and a half thousand dollars a month of attention, against roughly a hundred to run the system that gives most of it back.

And that ignores the two biggest wins, because neither shows up as time saved: the limitation date that stopped being one bad week away from a negligence claim, and the work in progress that finally got billed. Those show up as revenue and risk that did not leave.

Where to start on Monday

Nobody builds all nine steps at once, and you should not try. This is a map, not a project plan, and the map is useful the moment you have it, because it tells you where you are and what the next single step is.

So do not start with the screen that impressed you most. Start with the step the audit said is bleeding worst.

If work in progress keeps getting delivered and never billed, the money-on-the-floor wall and the invoice filler pay for the whole build in the first month. If deadlines and limitation dates are the fear, start with the matter radar and the limitation clocks.

If your fee-earners are drowning in document filing and chasing, start with the matters controller. The full logic for choosing the first thing is in what to automate first in a service business, and the honest way to measure whether it worked is in true ROI versus reported ROAS.

One thing at a time. Diagnose, build the one piece that hurts most, put a human gate on anything irreversible, then the next piece, which will be cheaper than the first because it sits on the same spine.

That is the whole method. This same framework runs in an accounting firm, a shipping firm, a clinic and a construction consultancy, the labels on the screens change and the nine jobs do not, and the full walk-through of the general version is in the 9-step framework for deploying agentic AI.

The firms that end up with a real system instead of a browser full of half-used tools all started the same way: one step too small to fail, aimed at a number they had already agreed was bleeding.