Legal AI

Grounded assistance, not a chatbot that practises law.

Not a chatbot that practises law

Legal AI, as we build it, is a set of tools on top of a matter file: a draft the model proposes from documents the user can already open, a PDF extract a clerk checks, a precedent search that returns a passage and a citation. It is not a chatbot that practises law. It is not legal advice. It is not a substitute for a qualified practitioner. LavisTech is a software shop in Chengannur, Kerala. We ship systems. We do not appear, we do not advise clients, and we do not file.

If a vendor shows you a chat box and a gavel and says the junior is now optional, you are not looking at a tool. You are looking at a liability with a subscription. The model predicts text. It does not hold a vakalat. A person still owns every sentence that leaves the firm.

This page assumes you want assistance grounded in your matters and your library — the pattern described as retrieval-augmented generation — not a public model that invents a section number because the prompt felt legal. If you only wanted reminders and a bill, you want practice management, and you may not want a model at all. That distinction is the whole of AI versus automation.

Grounded drafts from the matter in front of you

The useful button is not “ask anything.” It is “draft from this matter.” The action gathers the note, the last orders, the client’s letter, and any passage retrieved from a library you chose. It returns a draft in the box the advocate already edits. They change it. They send it — or they do not. The model never sends.

Grounded means the prompt was handed those materials and told to stay inside them. It does not mean we connected a shared drive from 2016 and hoped. A dump of every scan since the office opened is how you cite the wrong year. Curation — current matters, current templates, current permissions — is the work. If the set is small, we paste structured extracts. If the set is large, we retrieve. Either way, a person sees what was used.

We log the draft, the final text, and whether they changed it. After a hundred uses you know if the button saves time or creates cleanup. “People like chatting” is not a result. Change-rate is a result.

PDF extraction a person still checks

Notices, orders, agreements, and a phone-scanned vakalat arrive as PDFs. Extraction means suggested fields: parties, dates, identifiers, a hearing if one is printed clearly. The clerk accepts or corrects. Only then does the matter update. A wrong date written straight into the limitation list is worse than no extraction.

Read-only first applies here. The first weeks show the extract beside the PDF. A person copies or confirms. When the change-rate is boring, we write back through the same screen they already trust, with the same roles. We do not give the model a silent write to the calendar.

Bad scans stay bad. A model will guess. Guesses are labelled as guesses. If we cannot see a date, the field stays empty. Empty is honest. A confident fiction is how a practice misses a hearing and blames “the AI.”

Precedent search that shows the citation

Search over your own opinions, your own templates, and a library you have a right to use. Each hit shows the passage and a citation a person can open — file, paragraph, page. If we cannot show where a sentence came from, we do not show the sentence as a source. Hallucinated citations are the failure mode this feature exists to refuse.

We will not scrape a paid reporter you did not license and call it innovation. We will not fine-tune a model on judgments you cannot share. If your library is thin, say so. A search box over twenty templates is still useful. A theatre of “all Indian law” is not something a shop our size should sell, and we will not.

The advocate decides whether a passage applies. The software retrieves. Retrieval is not advice. Applying a precedent to a live client is legal work. The button does not apply. It points.

Read-only first

The sequence we use on any existing application is the one in how to add AI to an existing application: pick one task, see the data, ground the model, show a draft, measure, then maybe write back. Legal work does not get a special exemption from that sequence. It gets a stricter one. The first production feature does not email a client, does not file a form, and does not update a limitation date without a click.

Read-only is not timidity. It is how you learn the change-rate before the model can harm a file. Teams that skip it spend month three arguing about a letter that went out with the wrong annexure. The log would have shown the draft. There is no log if you started by wiring send.

If the practice system cannot accept a button, you are not ready for Legal AI. You are ready to put the matter in a system a program can see. Bolt-on chat beside a cupboard produces a screenshot for a seminar and a process that still lives in WhatsApp.

Person in the loop, always

A person with the right role accepts the draft, the extract, or the search hit before anything leaves the firm or becomes a date the firm trusts. The model does not send to a client. The model does not file a document. The model does not serve, does not upload to a court portal, and does not post a voucher. Those actions stay on a human click, through the same permissions the file already has.

Person-in-the-loop is not a slogan we print and then bypass for “low risk” mail. There is no low-risk mail once a client name is in it. If you want automation without a model — a reminder that always goes at 9 a.m. for dates a person entered — that is a workflow, and it still should not invent a date.

We will walk away from a brief that says “just let it send the routine ones.” Routine is where the wrong attachment hides. A junior clicking send is cheaper than a commission of enquiry.

The same permissions as the file

The model call inherits the signed-in user. If that user cannot open the family matter, retrieval must not return it. If they can open the commercial file, the draft may use it. We test that with two accounts and two folders, not with a slide about “enterprise security.” There is no second permission model the model “kind of understands.”

This is the same rule as the practice system. Legal AI is not a back door for the curious intern. Logs show which passages were retrieved for which user on which matter. If we cannot say that on Thursday, we did not finish the feature.

Client data does not go to a public model under a personal account. If a hosted model is used, it is under a contract you can read, with a retention you accepted. If that contract does not exist, we do not flip the feature on to look modern.

The model does not send or file

We will write this in the statement of work until it is dull. Drafts sit in the editor. Extracts sit as suggestions. Search sits as hits with citations. Send, file, share, and export-to-client are buttons a person presses. The model is not behind those buttons. If a future workflow looks like “prepare a bundle,” a person still confirms the index.

Inbox classification — “this looks like a notice, this looks like a bill” — can be a suggestion on the mail the clerk already reads. Filing the mail onto a matter is still their click. Classification is not advice about what the notice means.

Audit you can open on Thursday

Log the user, the matter, the passages retrieved, the raw model output, and the text they actually used. That log is how you answer “why did it say that?” It is also how you fix a prompt or a folder in an afternoon. Without it you will argue about feelings and then ban the button.

Keep the log under the same retention conversation you have for the matter. Do not invent a shadow archive the partners have never seen. Do not send the log to a vendor we cannot name. This is still the client’s file, with a model-shaped appendix.

What this is not

• A substitute for a qualified advocate or a responsible clerk.
• A system that computes limitation and files it as truth.
• A chatbot trained on “all law” we do not have a right to use.
• An agent that emails clients or uploads to a portal alone.
• Legal advice from LavisTech. We write software. Your practitioners advise.

It is also not magic on a cupboard. If the PDFs are unnamed and the matter is a nickname in someone’s head, extraction will fail in public. Do the file hygiene, or start with practice management and come back.

Cost and calendar

One grounded draft button on a matter you already store, with logging and read-only send, is a feature measured in weeks — the same shape as other AI attachments, not a transformation programme. Extraction of one document type is another feature. Precedent search over a curated library is longer, because curation is the work. A “legal AI platform” with no task is not something we will quote as a single number.

Model bills are yours to keep paying. They are usually small next to the integration. They are not zero. We will not invent a rupee build total before we have seen whether the matter system can accept a button and whether the library is a pile or a set. The wider menu of business AI, with honest time bands, is on AI for business.

If the existing application has no API and no way to add a control, the first sprint is access, not the model. That may be practice-management work. Calling it AI does not make the cupboard searchable.

How it sits on practice software

The matter is the system of record. The model is a reader and a drafter. Time, bills, and Tally stay where they are. WhatsApp stays a channel people use; the model does not become a silent participant in the group. If you need the file to exist first, start with the practice system. If the file already exists in software we can extend, start with one button.

We usually build this on the same stack as the practice application — often .NET — so permissions and audit are not a second product. The integration discipline matches our other AI integration work: schema, validation, no write without a person, measure the first hundred.

How to measure whether it helped

Pick a number before you start. Time to a first draft of the routine reply. Time to key a notice into the matter. How often someone still asks “where is that order.” After a hundred uses, look at change-rate and that number. If people rewrite every draft, fix retrieval or stop. Do not add a second button to rescue a bad first task.

A seminar demo that answers three planted questions is not a measure. Production is ugly PDFs, incomplete notes, and a senior who will not send anything they have not read. Design for that Tuesday. If the feature only works on clean text, say so in the scope.

How to start

Write one sentence. “On the matter screen, draft a reply from this record and the last three notes. Show sources. Do not send. Log the change.” That paragraph is a project. “Add ChatGPT to the firm” is not. Send the paragraph. If the reply is a platform diagram and a gavel, keep looking. If the reply is a list of assumptions about your matters, your roles, and your model contract, you have a partner.

Name the document type if you want extraction. Name the library if you want search. Name who may never see family files. Those constraints are the design. We will refuse a send-without-click. We will refuse a silent write to limitation. We will refuse to call the output advice.

If you want to walk through a matter screen and one task, write to us with the system you already run and the button you can describe in a sentence. We will tell you whether you need AI, automation, or a practice file that exists. We will not sell you a chatbot that practises law.