Law firm AI adoption fails most often for one reason, according to LeanLaw’s own experience adopting AI: the firm starts with the tool instead of the job. In LeanLaw’s AI and the Firm’s Money webinar, CEO Jonathon Fishman and Head of Engineering Steve Owings walked through what LeanLaw’s AI adoption got wrong in early 2025, what it eventually got right, and the decision framework — three adoption paths, a four-question vendor test, and an adopt/pilot/wait rule — any firm can borrow. The full replay is below; the expanded [AI Decision Checklist] (link when live) is the takeaway built to be handed to whoever runs your billing and operations.
[Embed: full webinar replay — YouTube]
What did LeanLaw get wrong first?
Jonathon opened with the confession most vendors skip. In early 2025, LeanLaw built its first AI tooling into the product — not because customers asked, and not to solve a defined problem, but to be able to say it had AI. His words: “We did it for the sake of the feature. And it fell on its face.”
The diagnosis, looking back, was a checklist that didn’t exist yet. There was no clear job to be done. There was no metric that would define success. And the team’s understanding of security and data privacy wasn’t yet deep enough to release the feature commercially. So the company went backwards to go forward: it stopped building AI into the product entirely and started over with a harder question — how should LeanLaw itself use AI before it asks a customer to?
That reset had three phases. First came the fumble. Second came culture: a company-wide message that everyone would learn to work with AI — tied to LeanLaw’s core values, and not optional. Engineers, it turns out, resist new ways of working exactly the way experienced lawyers do; deep expertise hardens workflows, whatever the profession. Third came organized adoption with explicit directives: pick the metrics, measure the change. Only after all three did AI go back into the product.
The result of that discipline is the number Jonathon shared on the webinar: engineering output up roughly 60–70% — with no added headcount and no increase in defects.
The three paths any firm can take with AI
Steve laid out the frame the rest of the hour hung on. A firm’s AI strategy takes one (or more) of three shapes:
Buy software with AI already embedded in it — a tool you use anyway, with AI doing a defined job inside it. The fastest start, with a constraint worth naming: you inherit that vendor’s opinion of what the job to be done is.
Use standalone AI tools — Claude, ChatGPT — fed with structured data and stitched into your workflows. Powerful and flexible, but the onus shifts to your people: prompting, refining, learning the tools’ vocabulary. Most teams plateau here without a small internal group pushing the advanced capabilities.
Build your own data architecture with AI layered on top. The biggest potential reward and the biggest risk — and the wrong place to start. If your firm hasn’t developed competency in the first two paths, the build path mostly transfers money to consultants against a problem you haven’t defined yet.
No path is wrong. The failure mode is choosing one without knowing which job it’s doing — which is precisely the mistake LeanLaw opened by confessing.
| Path | What it looks like | Best for | The tradeoff |
|---|---|---|---|
| Buy | Software with AI embedded (LeanLaw’s ClerkAI, document tools with AI layered in) | The fastest start; a defined job inside a tool you already run | You inherit the vendor’s view of the job to be done |
| Use | Standalone AI tools (Claude, ChatGPT) fed with structured data | Flexible workflows; individual productivity | The learning burden shifts to your people; most teams plateau without internal champions |
| Build | Your own data architecture with AI layered on top | Firms with defined jobs, mature data, and budget | Biggest reward, biggest risk — don’t start here |
The data ladder: where does your firm actually live?
The webinar’s most useful self-diagnostic started with a rule Jonathon insists on: workflow creates data. Nothing above the first rung works without the discipline underneath it — how the firm onboards clients, captures time, and runs billing.
From there, the ladder climbs four rungs. Reports — static tables, usually exported to Excel, describing what already happened. Most firms live here. Dashboards — interactive views that surface the number that matters, like a realization rate. Better, but as Jonathon put it: “Cool, my realization is 88%. Now what?” Embedded AI — insight plus direction inside the tools you own, including asking questions in plain language that standard reports can’t easily answer. Agentic experience — your data moving securely out of one system and into the AI tool your firm actually works in, where it can be combined with data from your other systems.
That last rung is where LeanLaw planted its strategic flag, and it’s the clearest statement of intent from the hour: LeanLaw is not trying to become an AI engine. The long-term bet is structured data conduits — an API today, an MCP layer in private beta — so a firm can bring its financial data to whatever AI tool fits, alongside data from the rest of its stack. One connected experience, now including the firm’s own AI.
Steve’s plain-English translation of the technical part: an MCP is “just the way Claude or ChatGPT knows how to talk to LeanLaw” — a secure tunnel between two applications. One solo contingency attorney, he shared, already runs his legal workflow in Claude and uses LeanLaw as the revenue layer underneath it — the billing, trust accounting, and QuickBooks Online connection that a general AI tool doesn’t have.
The checklist we wish we’d had in early 2025
The framework the webinar promised — expanded in the downloadable version — starts before any vendor demo. Name the job: what specific problem or opportunity is AI being pointed at? Name the number: which metric you already track should move, and by how much? Moving a collections rate from 83% to 88% is six-to-seven-figure money at most firms — that’s the scale worth defining before you spend anything.
Then the four questions for any AI pitch: What number does this change? What exactly does the AI do — read, draft, decide, or act — and who reviews it? What data does it see, and where does it go? And show me the before-and-after in a firm like mine. The third question deserves its own list — our data security questions for legal AI vendors covers what a good answer sounds like, and ABA Opinion 512’s requirements explain why the answer is an ethics matter, not just an IT preference. LeanLaw’s own answer, stated plainly on the webinar: we don’t train AI on your data, and we hold the same guarantee from every vendor we use.
Finally, the decision frame. Adopt now where AI drafts or surfaces, a human reviews, the data stays in your systems, and it moves a number you track — background time capture like Billables AI is the canonical example. Pilot carefully where AI touches client-sensitive work or acts on its own — small group, secure setup, people with appetite for learning. Wait where a vendor can’t pass the four questions, where it sounds too good to be true, or — the subtlest wait signal — where your firm doesn’t yet know what it’s trying to solve.
Where software ends and culture begins
The webinar’s sharpest moment of honesty came from a software CEO arguing against software. No tool will magically fix a collections problem, Jonathon told the group — collections starts with the engagement terms you set, the payment methods you require, and whether partners actually follow up. Tooling gives you the visibility and the instrumentation; the firm supplies the expectations. The same applies to AI adoption itself: the tools have never been more capable, and they still can’t install the culture that makes them useful.
Steve’s Monday-morning advice compresses the whole hour: find one job you think AI could do better, take a real stab at it, and — most importantly — share what happened with a colleague, whether it worked or not. That loop, repeated across a firm, is what adoption actually looks like.
Frequently asked questions
What are the three ways a law firm can adopt AI? Buy software with AI embedded in it, use standalone AI tools like Claude or ChatGPT fed with structured data, or build your own data architecture with AI layered on top. Most firms should develop competency in the first two before attempting the third.
What should a law firm ask an AI vendor before buying? Four questions: what metric the tool changes, what the AI actually does and who reviews its output, what data it sees and where that data goes, and what before-and-after results a similar firm achieved.
Does LeanLaw train AI on customer data? No. LeanLaw uses third-party AI providers that don’t train on customer data without permission — and doesn’t grant it. LeanLaw holds the same guarantee from every vendor it uses.
What is an MCP, in plain terms? A standardized, secure way for AI tools like Claude or ChatGPT to communicate with another system — a tunnel between applications. LeanLaw’s MCP layer is in private beta with a small number of firms.
Written by
Rachel Bondurant
Head of Brand and Content
Rachel Bondurant leads brand and content at LeanLaw, where she writes about legal billing, trust accounting, and the financial operations of modern law firms. Her work translates the realities of law-firm finance — billing workflows, IOLTA and trust compliance, and revenue leakage — into practical guidance for attorneys, firm administrators, and the accountants who support them.
Related articles