The Boutique AI Climb

I owe you a warning before I get into any of this: I'm not a lawyer. I've been trying to figure out the operational part of AI-native transformation for law firms for a while now, coming from a background of building and deploying AI in other regulated industries, and I know that's something a lot of lawyers have learned to read as the opening line of a car crash.

Half the reason I'm writing this is that after a relatively short stint of working with law firms helping them pick, implement, and adopt the technology, I've watched enough Silicon Valley people walk into a partner's office with a slide deck and tell them that everything, their judgment, experience, the pattern recognition built over decades of mistakes and battles and living through the consequences, can now be codified into statistical algorithms and replicated and replaced. And I think that's kind of rude.

This piece is not that piece. I'm staying in my lane, which is the operational side. The firms that have let me into their workflows and the lawyers who have opened their minds to me have taught me everything I know (so far) about how this profession runs, and I owe the whole piece to their generosity. So think of this as what I've learned from being let into the room.

Over the past year and a half I've been advising boutique law firms across Europe, the UK, the US, Canada, Latin America, and Australia on this operational piece specifically, writing about it weekly, and recording more than thirty-five conversations with the people figuring it out from the inside - partners, champions, ops leads, advisors, and a handful of the founders building AI-native firms from scratch. And the more of those firm engagements and conversations add up, the more I keep coming back to one specific kind of firm and one specific kind of question.

Most of the visible AI-native cases now are about firms that started from scratch, such as Lawhive, Eudia, Garfield AI, the YC W26 batch, or about solos who can turn the ship overnight. What about the boutique with 15 or 20 or 25 lawyers, a history of clients, two partners who've heard the AI-native talk a few times now, and a team that's been using ChatGPT seats for six months without much direction? That's the room I keep finding myself in. And the question I keep hearing, in one shape or another, is “what do we actually do from here?”

Helen's Legal AI Value Stack

A conversation I recorded recently for my podcast with Helen Fan , the mastermind behind the Legal AI Value Stack and Open Claw LLP, is what got me to finally put all this on paper. Her stack maps where defensibility lives in legal AI today, and it gives us the vocabulary we need to actually point at things in this market and say what they are.

The really short version goes like this.

There are five levels in total, climbing from raw AI capability up to a fundamentally redesigned firm.

Level 1 is the wrapper products on top of foundation models, chat interfaces with maybe a RAG layer, the ones already commoditized.

Level 2 is AI plus workflow, the legal specific tools, structured outputs, customizable playbooks, the wave of Word add-ins for redlining, which is the layer commoditizing right now.

Level 3 is proprietary data, where a platform itself gets smarter from how its clients use it at scale, in ways no single client's usage could produce.

Level 4 is system of record (Clio, Filevine), where the moat is operational gravity, because the firm runs its whole practice inside the tool.

And Level 5 is the hybrid model - the firms we just talked about (full-stack legal AI "companies"), the ones that stopped selling tools to law firms and became one themselves with AI as the operational backbone. Helen's argument is that this is the only level where value compounds rather than erodes.

If you haven't read it yet, Helen goes much deeper into each level in her original article here.

Workflows First, Data Later

In a traditional sense, you would start by tackling your data. We've all heard "garbage in, garbage out", which essentially means the more high quality data you feed any AI system, the better what comes out. The standard advice is - get your data in order first, then put AI on top of it.

The caveat is that for most boutiques, getting your data in order is an enormous project. Years of information scattered across your DMS, your CLM, inboxes, SharePoint, personal laptops, and sticky notes on people's desks, plus all the process knowledge sitting only in people's heads. The simple test for whether a firm is “data ready” is this: if you ask an associate to pull comprehensive information about a client, do they come back in minutes or hours?

Exactly. And trying to fix all of that before deploying any AI is what firms try and give up on at month four.

Now, Helen's order of operations in her Legal AI Value stack inverts the conventional view. Instead of building the data infrastructure first and putting AI on top of it, the data infrastructure emerges from the workflows themselves, accumulating one workflow at a time. And I think this is the way to go for most boutiques.

It works for two reasons. The data work happens surgically, one workflow at a time, which takes a fraction of the time a firm-wide data project would take. And the savings compound because each workflow you set up builds on the slice of data infrastructure you laid down for the previous one. By workflow three, you're moving faster than you were on workflow one, because the data layer underneath is already half-built.

For boutique firms in the 15 to 30 lawyer range, this is what makes the climb realistic. Three workflows over a few months, with the data layer building itself as you go, is something a real firm can actually pull off without dedicated AI headcount. And by the time those first three are running, you've bought yourself room to start looking at the more complex parts of legal work.

Stitch, Buy, or Build?

I've been part of plenty of conversations where the move gets picked before the problem has been clearly named, and a few months later there's a tool sitting inside the firm without a clear answer to which problem it was supposed to solve. The version that tends to work better in my experience is the boring one. You start from what's actually breaking in the current process, and the right move usually falls out of that on its own.

Stitch is for when the capability the firm needs already exists somewhere in their current setup, and the only real gap is that the pieces aren't talking to each other - wiring Clio into the workflow you've already built in Claude through an MCP server, routing Claude through your DMS, plumbing the intake form through to the practice management system. Cheap, fast, and reversible; and for the boutique firms it's often the move that actually sticks and is the easiest to adjust.

Buy is for when someone has already solved the exact problem commercially and the vendor has done the compliance and engineering work. Heavy contract review is one of the clear examples I keep seeing, where a good point solution will outperform what you can put together on your own. The trade-off is the vendor's roadmap and pace, which can hurt in a market moving as quickly as this one.

Build custom is for when the firm has a workflow specific to how their practice runs, that won't change next quarter, and where existing tools genuinely can't get them to the reliability or specificity they need. Comes with a maintenance commitment from day one because somebody has to keep the thing relevant and running.

In practice most firms end up running some blend of all three across their workflows, and the diagnosis is what tells you which one fits where. I wrote about both my recent newsletter and LinkedIn post. You will find many interesting takes in the comment thread.

The Three Workflows

To put the inversion into practice, divide the firm into three functions first. Start with intake, matter management, and billing. These three are where most of the non-billable capacity at boutiques is being absorbed today, and they're where you don't need AI doing anything close to legal judgement. Get these running, and you've bought yourself room to look at the more complex parts of legal work where AI starts getting interesting. But if your firm hasn't tackled these 3 core functions yet, this is where the climb starts.

Two assumptions I am making about your firm here. The first is that you have a sense of where you want the firm to go, its purpose as a practice, what it wants to be in eighteen months. The second is that the workflows you're about to tackle have been mapped, at least at the level of who does what, when, and how the handoffs work. If either isn't in place, the gap is smaller than it might sound: a strategy conversation for the first, and a workshop plus a couple of days of mapping for the second. Without them, the climb could still work, but it'll be slower, messier, and more expensive than it needs to be.

There's also a purpose to the order of things here: intake captures client structure, matter management adds matter structure on top of it, billing adds financial structure on top of that. Each workflow you finish leaves the next one with less work to do, because the data context it needs is already in place. Your base data infrastructure is being assembled one workflow at a time.

The right HOW for each of these workflows depends on the firm's situation: how technical the team wants to be, what talent is on hand, where they sit on the climb. For most boutiques, Stitch tends to be the common landing because they already have the systems and already have partial records in them, and stitching gives the highest value for the lowest investment. The walkthroughs below show how to think about that.

Intake. Intake usually runs through some combination of a basic form on (possibly) the website, leads landing in a shared inbox, AI used ad hoc to draft follow-ups, and the practice management system where matters eventually land. When you push that setup to its limits, what you'll find most of the time is that the AI layer CAN do the work (drafting, summarizing, pre-screening, categorizing), it just doesn't see the leads automatically, and the form isn't capturing enough structured information to let the AI do anything before manual handoff. So the gap is data access, and the move is Stitch - connect the form, the CRM if there's one, and the practice management system, and capture the right structured fields on the form itself so they flow through. This is the kind of stitching that low-code tools handle well. For example (please don't stone me), Power Automate, if the firm is on Microsoft. The structured client info you capture here, the entities, contacts, matter type, conflict-relevant details, is what matter management runs on next.

Matter management. Matter management usually has the practice management system as its central spine, with documents in that system or a separate DMS, client communications in email and chat tools, and day to day status scattered across all of them plus whatever's living in individual lawyers' notes. When you push that setup to its limits, what you find is that the AI layer CAN summarize and triage when fed the data, but can't see across all those systems on its own. So the gap there is data access too, same shape as intake, just spanning more systems. The move is Stitch - read access from the AI layer into the practice management system, the DMS, and matter-tagged email threads, surfaced into a unified view per matter. The matter structure built here, the status, deadlines, work-product references, is what billing pulls from next.

Billing. Billing usually runs through a billing tool that does the math fine, with the friction sitting everywhere around it. Time entries get captured late or inconsistently upstream, invoice descriptions need partner rewrites downstream, and the loop between the billing tool and the practice management system needs manual reconciliation. So the gaps there are mixed, data access on the systems side, capability on the descriptions. The move is Stitch - connect the billing tool to the practice management system so time entries flow automatically, and use the AI layer to draft invoice descriptions from matter notes for partner review. By the time billing is running, the client, matter, and financial layers span all three workflows. The data infrastructure is real, built as a consequence of doing the work.

To Helen's Point

Once intake, matter management, and billing are connected, you'll start seeing things change. You'll have one consistent picture of every client, every matter, and every dollar that's moved through the firm. The data infrastructure Helen describes gets its shape and starts paying off.

Picture what intake looks like at that point. A new matter lands in the form on Monday morning. Before anyone has touched it, the AI has cross-referenced everything the firm has ever done with that client, run conflict checks, and pulled the patterns from similar clients on similar matters: the angles the firm tends to win on, the risks that have hurt prior matters, the moves that have worked, the associate or partner with the closest prior experience, and the scoping approach that fit similar matters before. The lawyer who picks the matter up walks in with all of that already laid out. Their job is to weigh it.

That's where AI's role changes from saving time to enhancing what the lawyer can do with the (now stitched together) systems. Every workflow you add on top of that makes the firm's own data work harder for the firm.

And the firms that do that are really going to have a massive advantage. Remember how fast this past year went by? Imagine where some of your competitors are going to be next year, because this is what they are already doing.

The Boutique Law Firm Advantage

In eighteen months, the operational difference between a firm that's gone through this climb and one that hasn't is going to be visible - in capacity, in client conversations, in how partners spend their time. Some boutiques are already most of the way through it.

Boutiques are also better positioned to pull this off than larger firms. Fewer lawyers, fewer decision layers. And so one person, whether a partner, an ops lead, or an external advisor, can hold the whole picture and push it through. A 2,000-lawyer firm cannot produce that person by design. A boutique can. That's one more argument for why boutiques are in a much better place to punch above their weight with the help of AI.

What I Haven't Figured Out Yet

There's a couple of things I haven't cracked yet:

When one partner is in and the other isn't. A lot of boutiques don't have partner alignment on AI. One might be curious and engaged, the other skeptical or short on bandwidth. The framework above can run with one partner driving, but what that one partner usually can't do alone is change the firm; the work stays a side project rather than becoming how the practice runs. And before you ask, AI cannot fix your partner alignment.

Past thirty lawyers. This article is scoped to firms in the 15 to 30 range because that's where the framework fits cleanly. Past thirty, the dynamics start to shift. There are more partners, more practice areas, more dotted line interests in any technology decision, and the "one person holding the picture" argument gets harder to make. I don't know exactly where the line is. Somewhere between thirty and sixty lawyers, the framework probably needs to be reshaped.

Where stitching breaks down. I've recommended Stitch as the common move throughout this article, but stitching has costs over time that increase with how complex the stitch is. Vendors change their APIs, the connective tissue between systems needs maintenance that someone has to own, and at some point a firm running on five stitched together tools will probably need to consolidate or rebuild. I haven't seen many firms hit that wall yet, so where the limits actually show up is partly speculation on my part.

Climbing Without Starting Over

The most obvious version of AI-native right now is the one that requires starting over, which is not the path most 15 to 30 lawyer boutiques are on. They can climb without starting over - three workflows shipped one at a time over a handful of weeks each, a data layer building underneath as a consequence of doing the work, and deeper AI applications opening up after that.

If you're a partner, AI champion, lawyer, ops lead at a boutique making decisions about this, I would love to hear from you.