ClerkAI v2 Is Live: An Answer That Knows Which One You Mean

A firm with three matters named after the same defendant and two clients who share a surname has a specific problem: an assistant that guesses which one you mean is worse than no assistant. ClerkAI v2, live as of August 31, 2026, is built around fixing exactly that.

ClerkAI is LeanLaw’s embedded billing assistant. It’s read-only: it answers questions about your billing data, it doesn’t take actions on your behalf, and that hasn’t changed in this release. What’s changed is how confidently it can tell your records apart, and how it behaves when it can’t.

Why “which one did you mean” was the real problem

The value ClerkAI offers isn’t answering easy questions; a standard report filter already handles those. Its strength is complex queries that don’t map cleanly onto a filter menu: cross-matter comparisons, questions that span time periods, questions phrased the way a person actually asks them rather than the way a report is structured. That strength turns into a liability the moment two records look alike, because a query engine that can answer a subtle question can also confidently answer the wrong one.

Say the hard thing first: an assistant that returns a fast, wrong answer costs more than one that’s slow. Before v2, a query about “the Anderson matter” against a matter list with two Andersons had to rely on context or ask you to clarify in your own words. Neither is reliable at volume, and a wrong answer treated as right doesn’t announce itself. It just sits in someone’s understanding of the matter until a reconciliation or a client conversation surfaces the mismatch, usually later than you’d want.

There’s a real cost hiding in the old behavior even when it worked, too. A timekeeper or biller who has to stop, rephrase a question, and re-check which record an answer actually covered is spending time on disambiguation rather than on the matter itself. None of that time gets billed to anyone; it’s overhead the assistant was supposed to remove, not relocate.

@-mention entity reference: naming the exact record

The headline change in v2 is @-mention entity reference. Instead of describing a matter, client, timekeeper, or invoice in prose and hoping the assistant resolves it correctly, you can name the exact record directly, and the assistant treats that as unambiguous. Two clients named Smith and three matters with near-identical names stop being a source of guesswork, because you’re not asking ClerkAI to infer which one from context: you’re telling it.

That single change does most of the work the other v2 additions support. A worked example: a hypothetical mid-size firm runs two matters for different clients, both captioned around the same commercial dispute, plus a third, unrelated matter that happens to share a client’s last name with one of the disputed parties. Before entity reference, a written-time question touching “the dispute matter” was a coin flip between three records. With entity reference, the question names the specific matter, and the answer is scoped to it from the start.

The same logic extends to clients, timekeepers, and invoices. A question about “the invoice we sent last month” is ambiguous the moment a client has more than one open matter; a question that names the invoice directly isn’t. The pattern repeats across every entity type ClerkAI touches, which is why the team built it as one mechanism rather than four separate fixes.

What else changed, and why each piece exists

The other v2 changes aren’t new capabilities so much as v2 taking the disambiguation problem seriously in every direction:

• Clarification instead of guessing. When a request is genuinely ambiguous even with entity reference available, ClerkAI asks a clarifying question rather than picking an interpretation and running with it. That’s the same instinct as entity reference, applied to the cases entity reference alone doesn’t fully resolve.
• Persistent chat context. A conversation with ClerkAI survives closing and reopening, with a history you can return to, so a multi-step question doesn’t have to be re-established from scratch every session.
• A feedback loop. Marking an answer helpful or not gives the team a signal on where the assistant is still getting things wrong, which matters more for an assistant whose job is judgment calls on real financial records than it would for a simple lookup tool.
• CSV export of a result table. When the answer to a complex query is a table rather than a sentence, you can take it with you rather than re-running the same question in a report builder to get exportable output.

Deliberately incremental rather than a new AI paradigm, this is ClerkAI getting more careful about the one failure mode that actually matters for a billing assistant: confidently answering about the wrong record.

Guided example prompts also shipped with v2, aimed at the opposite end of the same problem: a new user who doesn’t yet know what a complex query looks like can start from a prompt built for the tool rather than guessing at phrasing. And when a question is really about how the product itself works rather than about your data, ClerkAI now points you to support instead of attempting an answer it isn’t positioned to give.

What ClerkAI still doesn’t do

ClerkAI doesn’t take actions. It doesn’t file anything, adjust anything, or send anything on your behalf; it answers the question you asked, about the records you named, and stops there. That boundary is deliberate, not a gap on a roadmap waiting to close. Read-only, entity-aware, and increasingly good at telling your firm’s own similarly named records apart is the whole scope of what shipped, and being direct about that scope is part of what makes the answers you do get worth trusting.

What this depends on

• How much of your matter and client naming already overlaps — a firm with distinct client names gets less benefit from entity reference than one with repeat surnames or templated matter titles.
• How consistently your team enters matter and client names in the first place, since entity reference resolves to the records you have, not the records you meant to have.
• Which questions your firm actually asks a billing assistant versus a fixed report — the value scales with how often your real questions don’t fit a report filter.
• How your firm handles the answers ClerkAI gives: as a starting point for a decision, not as a substitute for the judgment a timekeeper or biller still has to apply.

The question worth asking goes past whether an assistant can answer a billing question to whether it knows, with certainty, which record it just answered about — and whether it says so when it doesn’t.

Related reading: what answer engines get wrong about legal billing, AI that touches the firm’s money vs. AI that does legal work, and what LeanLaw’s WIP, realization, and velocity reports show.