Lia · RIA
Product vision
Jev in Lia: what changes in a law firm's day to day
Jev is an artificial intelligence model that does only one thing: it takes a text and a closed question (yes or no?, which of these?, what score?) and returns the answer along with how confident it is in it. It does not chat and it does not write. Underneath Lia, this settles the small decisions nobody in the firm makes today: which practice area this lead belongs to, who was left without a reply, whether the card should already have moved to the next stage, whether the message about to go out asserts something nobody checked.
Vision document. None of the functions described here is live in Lia today. The numbers come from RIA internal tests run on September 20 and 21, 2026, with real conversations from a client law firm, and they are early results on a small sample. They are not performance guarantees.
The problem: the decisions nobody makes #
A law firm is made of thousands of small decisions, and most of them never get made. Is this lead in my practice area? Is this client still waiting for a reply? Should this card have moved to the next stage by now? Nobody has time to decide this message by message, so it goes undecided: the CRM nobody updates, the lead that went cold without anyone noticing, the "I'll get back to you" that turned into a week, the dashboard with only one column filled in.
What Jev is #
If Lia is the teammate who talks to the client and drafts the filing, Jev is the intern who glances at a stack of five hundred papers, sorts them into three piles in two minutes and honestly points out the ones it was unsure about.
Three things that make it useful:
- Fast: it answers in under a second, even with several questions at once.
- Cheap: classifying a firm's entire conversation history costs less than one Brazilian real (under 20 US cents).
- Knows when it does not know: the confidence it reports is real. When it says "I'm not sure", the decision goes to Lia or to a person.
Three things it does not do:
- It does not draft filings, contracts or messages.
- It does not give legal opinions or decide what belongs to the lawyer.
- It does not replace Lia: it decides whether and which; Lia does the what.
This also shows up on the bill: today every small decision Lia makes draws on the firm's own AI account, which is the login subscription under the client's company tax ID. Jev takes the intern work off her plate, at intern prices, and leaves Lia free for what only she does well: talking, understanding the case and writing.
A day at the firm: five moments #
| Moment | Today | With Jev underneath Lia |
|---|---|---|
| Morning, before coffee | The lawyer scrolls the screen to find out who was left without a reply; it depends on the memory of whoever handled things yesterday | A short list of who is waiting on something from the firm, sorted and with the reason |
| A lead comes in | The card is created with no practice area, no stage and "normal" priority, like every other one | The lead arrives already classified: practice area, pipeline, whether it is a lead, a former client or a wrong number |
| The conversation moves on | The client says "go ahead and send the contract" and the card stays where it started, because moving it means remembering to move it | With every reply, a silent question: has the case moved to another stage? Lia is notified and moves it, within the firm's rules |
| Before sending a reply | An answer given off the top of someone's head can state that a person has no claim and turn away a client who did have one | The message goes through one question: does this give a verdict or promise something without a source? If so, Lia checks the record before sending |
| End of the day, the dashboard | The CRM views exist, but they show everything in a single column | The views truly separate things: how many leads await a decision, how many contracts are unsigned, who went quiet |
And the house rules #
Every firm has its rulebook: "don't discuss deadlines with a client without going through me", "billing only after checking the amount in the system", "a lead who already receives the benefit is never dismissed". Lia reads all of it, but a rulebook with hundreds of rules does not fire at the right moment. With Jev, before each action Lia asks the whole rulebook at once: which of these rules apply right now? Two or three come back. On a greeting, none. It is the checklist for the moment, assembled on its own.
Across Lia's five roles #
Items marked tested were measured with real conversations from a client law firm in September 2026; the others are planned applications.
- Intake: screening at the door (lead, former client, vendor or wrong number), practice area and the right pipeline (tested). Also: "does this message still need a reply?" and "does this reply give a verdict without a source?" (tested).
- Administrative assistant: practice area, status and "client waiting for a reply" on every card (tested). Daily sentinel: does the card's stage still match the conversation? She flags it, she does not move it. Initial load of an empty CRM: the entire history classified in minutes.
- Billing assistant: what the client meant in the billing conversation (sent a receipt, will pay later, disputes it, asks to pay in installments). Before sending: does this message mention an amount, an installment, a date or an agreement?
- Controller: auditing against what is written down. Did this reply go against a recorded rule? Repeat issues and promises with no follow-up (tested).
- Drafter: Jev does not write a single line. Around the writing: which document arrived and whether it is complete, whether the filing meets the firm's checklist, which of the firm's templates is closest to the new case.
The pattern is always the same #
Nothing irreversible (billing, dismissing a lead, marking a case as lost) happens without going through this number.
What we measured #
In September 2026 we ran Jev over 30 real WhatsApp client conversations from a client law firm, with eight questions per conversation, without touching anything: just measuring.
- Under 1 second per conversation, with all eight questions answered at once.
- US$0.01 total cost for the test with 30 conversations; the firm's entire history would come to less than one Brazilian real.
- 25 of 30 legal practice areas classified with high confidence; on the 5 cards that already had the subject matter noted, it got all 5 right.
- 9 of 30 conversations flagged as "client waiting for a reply". We checked two by hand: in one, Lia had written "I'll take it to the lawyer and get back to you" five days earlier, with no follow-up; in the other, a question about case activity was left hanging. Both were right.
- 26 of 26 on "does this message still need a reply?", including the classic "perfect, thank you so much, I'll wait to hear from you", which is not a sign-off: the person expects something to be delivered.
- 16 sentences rephrased in the "verdict without a source" test: the current keyword filter let 9 through, all in the dangerous direction (the message goes out). Jev caught all 9.
- 36 rules from a firm's rulebook evaluated in a single question, in one second: it lit up the right rule in four different situations and none on a greeting.
What did not work well #
Priority (urgent, high, normal, low) came out with low confidence in almost half the cases: the concept is too vague for an automatic score. It works as a reading order with a person checking, not as an action. And the pipeline stage only works well when the stages are described in plain language; where the stage depends on something that does not show up in the conversation (a payment in the system, a click by the lawyer), the system knows, not the text.
What still belongs to the lawyer #
- Marking a case as won or lost.
- Dismissing a lead, billing, negotiating fees.
- Giving the verdict on the specific case.
- Setting the house rules and writing them so they can be looked up one by one.
And what stays with Lia: talking to the client with the full context, drafting messages, filings and contracts, carrying out the action after Jev has flagged it and deciding the cases where Jev reported doubt.
Two conditions for it to work #
- Text only. Jev reads text. Audio and photos of documents are transcribed first, something Lia already does.
- Rules written as units. A rulebook with blocks dozens of pages long does not light up at the right moment; short, named rules do. In the tests, the same rule went from "not found" to "first on the list" just by being split out of the block it was in. Organizing the rulebook is worth it with or without Jev.
Where to start #
- CRM first. Practice area and "client waiting for a reply" are the two most reliable classifications we measured, and the two most missing from the dashboard today. Start there, at a pilot firm, in suggestion mode: Jev proposes, Lia or the lawyer confirms.
- Then, the verdict brake. The question "does this assert something without a source?" before every send, in place of the keyword filter.
- Finally, the rulebook. Once the rulebook is written as units, the checklist for the moment starts assembling itself.
Frequently asked questions #
Does Jev replace Lia? #
No. It decides whether and which; Lia does the what. Jev does not talk to the client, does not draft filings, contracts or messages, and does not give legal opinions. When it reports doubt, Lia or a person makes the decision.
Is this already working at my firm? #
No. This is a vision document: none of the functions described is live in Lia as of the publication date. What exists are internal tests with real conversations from a client law firm.
How much does it cost to run Jev? #
In the test with 30 conversations and eight questions each, the total cost was US$0.01. By the same math, classifying a firm's complete history would cost less than one Brazilian real. Jev also lightens the firm's AI bill, because it takes the classification work off Lia.
What happens when Jev gets it wrong or is unsure? #
It reports its degree of confidence along with the answer. High confidence proceeds on its own; medium confidence goes to Lia, with the full context; low confidence goes to a person. Nothing irreversible (billing, dismissing a lead, marking a case as lost) happens without that number.
Are my data and my clients' data exposed? #
The client law firm in the tests is not identified and the conversation quotes were anonymized. Jev reads text to classify it and returns a closed answer; it does not write or send anything to the client.
Data sources #
All numbers come from RIA internal tests run on September 20 and 21, 2026, and from the model's public documentation.
- Test with 30 real client conversations (RIA, 09/21/2026). Thirty WhatsApp conversations from a client law firm, eight questions per conversation: average time, total cost, 25 of 30 practice areas with high confidence, 9 of 30 "waiting for a reply" and the two manual checks. The firm's complete history (568 conversations) was sized from that cost.
- "Does this message still need a reply?" test (RIA, 09/20/2026). 26 tricky messages taken from cases documented in the system itself; 26 of 26 with the existing rule + Jev combined.
- "Verdict without a source" test (RIA, 09/20/2026). 16 sentences rephrased from field cases; the keyword filter let 9 through, Jev none.
- Firm rules test (RIA, 09/20/2026). 36 rules from a client firm's rulebook evaluated in a single question against five situations.
- TypeSafe AI: model and pricing documentation. Jev model; billing only for the text sent, with no charge for the answers; basis for the per-conversation cost calculation. https://docs.typesafe.ai/models
Early results, measured on a small sample; they are not performance guarantees. None of the functions described is live in Lia as of the publication date.
Reinaldo OliveiraFounder, RIAPublished September 22, 2026 by Reinaldo Oliveira, RIA. Markdown version for AI: jev-in-lia-law-firm-decisions.md · All articles in English · Blog em português