Retrieval, and deliberately nothing past it

Ask a practitioner in a four person firm where their contracts are and the honest answer is usually four places at once. Some sit in a document management system with reliable metadata. Some sit in a folder tree that made sense to whoever built it in 2014. Some exist only as a scanned counterpart attached to an email, executed, binding and effectively invisible. A few live in a filing cabinet.

When a client asks whether their standard indemnity survived the last round of negotiation, the practitioner knows exactly what the answer depends on. They need the executed version of one particular agreement and the paragraph in it that deals with indemnities. That is a retrieval problem wearing the costume of a legal problem, and the profession pays for it in six minute units.

Full text search is not the answer either, at least not on its own. Searching for the word indemnity in a folder of three thousand PDFs returns three thousand results, most of them the definitions clause. Searching for the concept, phrased the way a person would say it out loud, is the part that current retrieval methods are genuinely good at.

The word gets used loosely, so here is the intended behaviour in the plainest terms available.

• A firm points the software at a set of documents it already owns and has the right to hold.
• Each document is split into passages. Each passage is indexed, in place, with a reference back to the document it came from and the position within it.
• A question typed in ordinary language is matched against that index.
• The result is a ranked list of passages, quoted verbatim, each one labelled with the document name, the date on the face of the document and where in the document the passage sits.
• Clicking a result opens the source. That is the end of the interaction.

Notice what is absent from that list. Nothing is written. Nothing is condensed into a sentence that the practitioner then has to trust. There is no chat window that will happily keep talking after the useful part has been said. The software has no view about the clause and no capacity to form one.

A language model that drafts a clause produces something that reads as though a lawyer wrote it. That is precisely the difficulty. The failure mode of generated legal text is not gibberish, it is plausible text with a defect that only careful reading finds, arriving in a format that discourages careful reading.

Courts in several jurisdictions have now dealt with filings containing citations that did not exist. The pattern is consistent and it is not really about the technology. A tool that produces confident prose shifts the burden of verification onto a busy person, at exactly the moment they have been given a reason to relax.

Retrieval has the opposite shape. The worst case for a retrieval failure is that the software returns the wrong paragraph, which a practitioner notices immediately, because the whole interaction is built around reading the paragraph. A tool that fails visibly is a better tool for this profession than one that fails invisibly.

This is a product decision rather than a permanent statement about what is possible. If it ever changes, it will change on this page first, in writing, with the reasoning set out and before any such feature is available, rather than quietly in a release note.

Nothing here exists yet, so treat this section as a statement of intent that can be held against us later rather than as a description of something running today.

The documents in a legal practice are the most sensitive material most firms hold. Client contracts carry commercial terms, personal information about counterparties and, very often, the subject matter of advice. Any sane design starts from the position that the vendor should see as little of that as the architecture allows.

The intended arrangement is that indexing happens on infrastructure the firm controls, and that AI-LAW PTY LTD does not receive copies of a firm's documents as a matter of course. Where a firm chooses a hosted arrangement instead, that becomes a written agreement with a named location for the data, a named list of any subprocessor and an obligation on us to act only on the firm's instructions.

What we would never want is the arrangement that has become normal elsewhere, in which a vendor takes a copy of everything for the stated purpose of improving the product. Client documents will not be used to train anything. That commitment is repeated in the privacy policy so that it sits somewhere with contractual weight rather than only in marketing copy.

A solicitor owes duties of confidentiality to a client under the applicable conduct rules, and legal professional privilege belongs to the client rather than to the practitioner. Handing documents to a third party supplier engages both, and it is the firm rather than the supplier that carries the professional consequence of getting it wrong.

We are not in a position to advise any firm on whether a particular arrangement is consistent with its obligations, and we will not pretend to be. What we can do is make the technical facts easy to put in front of whoever does advise the firm, which means writing down where data would sit, who could reach it, what is logged and what is retained, in language that survives being read by somebody who is not an engineer.

A practice considering any tool of this kind, ours included, should satisfy itself independently before a single client document is copied anywhere.

Legal practice in Western Australia is regulated under the Legal Profession Uniform Law, applied in this State by the Legal Profession Uniform Law Application Act 2022 (WA). Section 10 of the Uniform Law prohibits a person from engaging in legal practice unless they are entitled to do so. Section 11 prohibits representing or advertising an entitlement that the person does not have.

• AI-LAW PTY LTD is not an incorporated legal practice.
• It does not hold, and has never held, an Australian practising certificate.
• It has no principal who is an Australian legal practitioner.
• It does not engage in legal practice and does not represent that it is entitled to.
• It does not hold professional indemnity insurance of the kind required of a law practice, because it is not one.
• It holds no trust account and receives no client money.

Section 18 of the Australian Consumer Law prohibits conduct that is misleading or deceptive or likely to mislead or deceive. A software company that lets a buyer form the impression it is a law firm, or that its output is advice, is doing precisely the thing that section addresses. The four denials on the home page exist for that reason as much as any other.

A registered Australian company, a domain name, this website and a design brief. That is the complete list, and the temptation to make it sound larger is the exact temptation this page exists to resist.

There is no released software, no repository open to the public, no application programming interface, no trial and no waiting list. Nobody has paid the company anything. No firm has evaluated a build, because no build has been offered to anyone to evaluate. If you find a page anywhere claiming otherwise on our behalf, it did not come from us and we would like to know about it.

If any of those slip, the honest thing is to say so here rather than to ship and hope. This page will be updated when the list changes, including when something is dropped.