AI That Touches the Firm's Money vs. AI That Does Legal Work

Two different AI conversations are happening in legal software right now: one about AI that practices law, and one about AI on the business that turns legal work into collected revenue. LeanLaw lives in the second conversation, and it isn’t the bigger one.

AI that practices law is the bigger story, and it isn’t ours to tell

Research, drafting, and review are being reshaped by AI in ways that matter to how lawyers do the work itself: how a memo gets researched, how a first draft gets produced, how a contract gets reviewed before anyone bills a client for reading it. That’s a real shift in how legal work gets done, and it’s happening independently of anything a billing or trust accounting platform touches. This post is about a narrower conversation. LeanLaw doesn’t practice law, review documents, or draft anything, and we won’t pretend our lane is the more consequential of the two. A firm can, and probably should, be having both conversations at once, with different vendors, for different reasons.

Our lane: AI on the money the work produces

Once work is performed, it still has to become an accurate time entry, a correct invoice, a matched payment, and a number a partner can trust in a report. That sequence, from work performed through billing, collections, and the reporting that comes out of it, is Legal Revenue Operations, and it’s where LeanLaw’s own AI work sits: on the operation behind the legal work, not on the legal analysis itself.

Four numbers describe how healthy that operation is: utilization, realization, collection rate, and billing velocity. An hourly firm typically leaks through realization, a flat-fee firm through margin, a contingent firm through how quickly cash moves after a case resolves. None of that changes because AI enters the picture. What changes is whether a firm can see those numbers clearly enough, and quickly enough, to act on them, which is a data and reporting problem before it’s an AI problem.

Embedded AI: an assistant that answers, and stops there

ClerkAI is live, early, and read-only: it answers questions your standard report filters can’t easily handle, and it takes no actions on your data. Ask it about a matter, a client, or an invoice, and if two records could plausibly be what you meant, say two matters named after the same defendant, it asks which one instead of guessing. That’s the whole shape of embedded AI, as we define it: a feature inside the product that does something useful for you right there, without ever reaching for a lever on your behalf.

The distinction between answering and acting is doing real work in that description, and it’s worth being precise about it rather than letting “AI-powered” stand in for whatever a firm imagines. An assistant that answers a complex query about your book of work is a different thing, categorically, from one that would go make a change to an invoice or a client record on your say-so. LeanLaw’s own embedded AI is squarely the first kind today, and any move toward the second kind is not something we’re describing here with a date attached, because it hasn’t happened yet.

Agent Experience: designing for the day something acts on a firm’s behalf

A separate, newer discipline is forming around a different question. Rather than what an AI feature does inside a product, it asks whether the product itself is built so an agent acting for a user can work with it at all. Gartner’s 2026 Hype Cycle for Agentic AI names this as an emerging category. “Agent-ready” is our own plain-language term for what that discipline produces.

Two concrete facts, not a roadmap: LeanLaw’s public API is something an agent can already use today, and a private-beta layer built specifically for agent access exists but isn’t generally available. Where that layer goes, and when, isn’t something we’re putting a date on here, or anywhere else.

The standard behind that private-beta layer, the Model Context Protocol, was donated to the Linux Foundation’s Agentic AI Foundation at the end of 2025, co-founded by Anthropic, Block, and OpenAI. That donation is why building toward it counts as a bet on shared infrastructure multiple vendors are converging on, rather than a bet on any single vendor’s approach.

Any conversation about an agent touching a firm’s data has to answer a governance question before it answers a features question. Ours: third-party AI providers don’t train on customer data without our permission, and we don’t grant it. We don’t sell client data. That sentence has to hold regardless of which specific agent capability a firm is evaluating, or how far along it is.

Two conversations, one firm

Nothing about this requires picking a side. A firm can adopt AI for research and drafting on one track and expect its billing platform to be agent-ready on a completely separate track, at its own pace, governed by its own rules. The two conversations don’t compete for the same budget line or the same decision, and treating them as one conversation is usually how a firm ends up either over-cautious about both or unquestioning about both, instead of applying the right level of scrutiny to each.

Picture, hypothetically, an agent working on behalf of a managing partner who wants a weekly answer to “which matters are furthest behind on realization.” Whether that’s worth building toward depends less on whether an agent can technically reach the data and more on whether the firm is comfortable with an agent asking the question unattended in the first place. Every firm gets to make that governance call for itself, on its own timeline, no matter what any single vendor ships.

What this depends on

• Whether a given AI feature answers a question or is being asked to take an action on your firm’s data
• How your firm’s own data governance policies treat third-party AI access, regardless of what any vendor allows
• Whether the workflow you’re picturing depends on a capability that’s still in private beta or purely exploratory
• How your billing and practice management systems expose data in the first place, since an agent can only act on what it can see

The question worth asking about any AI feature

Every AI feature you evaluate this year, in either conversation, reduces to one question: does it answer, or does it act? A vendor that can’t say plainly which one it’s offering, and what happens to your firm’s data along the way, hasn’t decided that yet either, and that’s worth knowing before the feature ever touches a live matter.