It’s nearly nine and a first-year solicitor is building an indemnity clause from a blank page, one careful sentence at a time. Three years ago, a partner in the same firm spent the better part of a fortnight getting that exact clause right for a client in the same industry, arguing it through two rounds of markup until it finally held.

That version exists. It is sitting in the document system right now, attached to a closed matter, under a filename nobody remembers. The junior can’t find it, so she writes it again, slower and a little worse, and the client pays for the firm to relearn something it already paid to learn once.

The firm that can’t find what it knows

This is the part of legal practice nobody puts in the pitch to clients: a great deal of expensive expertise gets reinvented every week because it can’t be found.

Look at where the hours actually go. On Clio’s 2024 Legal Trends numbers, lawyers record a 37% utilisation rate, which is just under three billable hours out of an eight-hour day. And of the hourly work that does get done, about two-thirds (66%) is documenting, retrieving information and analysing it, with roughly 74% of hourly-billed work being the kind of thing generative AI could take on. Read those two figures together and the shape of the problem appears. The scarce, expensive thing, a qualified lawyer’s day, is being spent mostly on finding and reassembling knowledge, not on judgement.

The systems don’t help. Actionstep’s 2025 report on midsize firms found professionals use an average of 6.6 different tools to run a single client matter, 45% keep duplicate client and matter records because those tools don’t talk to each other, and 73% say their current tools don’t support how they actually work. The fee-earner becomes the integration layer, carrying context by hand from one system to the next.

And it shows up in the money. On Clio’s mid-sized data, firms realised 83% and collected 84% of their billable work, which is another way of saying roughly a sixth went uninvoiced and a sixth uncollected, with a median 97 days of lockup between doing the work and seeing the cash. Some of that leak is a billing problem. A lot of it is a memory problem: work done well, then lost.

The bottleneck isn’t drafting

Here is the thing the market keeps getting wrong. The bottleneck in a law firm isn’t writing the clause. Any competent solicitor can write the clause. The bottleneck is that the firm cannot find the version it already perfected, so the writing happens again from scratch.

The knowledge is genuinely there. Every matter your firm has ever run is a précis of hard-won judgement: the clause that survived the other side’s markup, the letter of advice that held up, the argument that worked. That is your precedent bank and your matter archive, and for most firms it is effectively write-only. Things go in. Almost nothing comes back out, because there is no way to ask the archive a question in plain English and get the exact clause, from the exact matter, with the exact wording that was signed.

The AI they’re selling you is the one you can’t reach

So the sensible move looks obvious. Buy AI that finds and drafts law. Except the AI the industry advertises hardest is the tier that helps you least, for two separate reasons.

The first is reach. When RMAI ran an adversarial check across the legal stack in 2026, the marketed “AI lawyer” products, the ones you see at every conference, cluster in a tier that an external system can barely touch: read-only, sales-gated, or closed, and in several cases competing for the very workflow you are trying to own. That’s the least reachable layer in your business. The reachable layer is the unglamorous one, where your matter data already lives: the cloud practice-management systems that own the matter and billing record (Clio, Actionstep, LEAP, Smokeball), the document store (NetDocuments, iManage, Microsoft 365), the ledger (Xero, MYOB), e-signature, and, distinctively for Australian firms, the search and e-conveyancing rails (InfoTrack, GlobalX, PEXA). Those expose real read-and-write interfaces an agent can use today, inside your own tenancy and under your own access controls. InfoTrack has even shipped an agent-native interface built for exactly this. Reachability, not model cleverness, decides which AI actually pays back, and it points away from the shiny product and toward your own filing cabinet.

The second reason is sharper, because it’s about safety. The research-and-draft-the-law tools are precisely the ones that invent law. Stanford’s RegLab benchmarked the leading legal-research AI tools and found they hallucinated on somewhere between 17% and 33% of queries, more than one in six. That is not an abstract risk in Australia. In the Dayal matter, the Victorian Legal Services Board and Commissioner varied a practitioner’s certificate on 19 August 2025 after he filed AI-generated authorities he had not verified. He can no longer practise as a principal, cannot hold trust money, may practise only as an employee solicitor, and must be supervised for two years. No wonder that in Thomson Reuters’ 2024 survey, 50% named a lack of demonstrable accuracy and 42% a lack of demonstrable security as the barriers holding AI back, while 95% of midsize firms are using it anyway and only 46% feel confident they can govern it (Actionstep, 2026).

Retrieval of your own knowledge has the opposite risk profile. It doesn’t reach out into the whole invented universe of case law. It reads what your firm has already written, and it can show you where it got every word.

What actually changes

Put the AI lawyer to one side and picture the boring, valuable version instead.

The solicitor stuck on the indemnity clause asks, in plain English, whether the firm has done this before. An assistant sitting on the document store and the practice-management system searches the firm’s own corpus and surfaces the partner’s clause, from three years ago, quoting it back with a citation to the source matter and the exact version that was executed. It shows its working. And when the answer genuinely isn’t in the archive, it says so and refuses, rather than filling the gap with a confident fiction.

The lawyer stays in charge of every step. The tool retrieves and surfaces precedent with its source attached; an admitted practitioner reads it, checks it against the current matter, and verifies the source before a word of it is used. Nothing is auto-filed. Nothing is auto-advised. This is not a compromise the technology forces on us; it is what the rules already require. The NSW Supreme Court’s Practice Note SC Gen 23, in force since 3 February 2025, permits generative AI for exactly this kind of work, chronologies, indexes, document review and draft submissions, bars it from the content of sworn evidence, and requires the practitioner to verify every citation, verification that cannot itself be done by an AI. A grounded, cited, refuse-when-unsure retrieval tool with a lawyer signing off sits squarely in that permitted zone.

What changes is not who exercises judgement. It’s whether the firm’s best thinking is available the next time someone needs it, instead of being rebuilt from a blank page.

The honest ledger

Two things I won’t dress up. Legal is a new sector for RMAI, and there is no independently-audited Australian small-firm case with hard financials to point to yet. The strongest available numbers are vendor-reported, so treat them as illustrative and never as our result. For a sense of magnitude only: at Australian commercial firm Holding Redlich, LexisNexis reports one research task falling from 4.5 hours to 35 minutes, with the firm estimating around 11 hours a week of potential saving. That is their story with their tool, not a promise about your firm.

Second, retrieval only pays back if it plugs into where your data actually lives, and some of your systems are more reachable than others. A build typically ships in three to six weeks in the $10k–$60k band, and it starts with a free 30-minute discovery call where we map your stack before anyone promises anything; the build itself is quoted fixed-scope after that, on your numbers. If a core system is being retired, and some are, FilePro is sunsetting by 31 December 2026, that is exactly the kind of thing the diagnostic surfaces first.

Where to start

You don’t need a transformation programme to find out whether this is real for you. You need one honest look at what your firm already knows and can’t currently find, and which of your systems can actually be reached.

Two ways in, both low-risk:

• Read the map first. We’ve written a plain-English brief on the Australian legal software landscape and the knowledge-retrieval gap, which of your systems can be reached, by what mechanism, and what “closed” really costs you. → https://realmindsai.com.au/guides/legal/
• Book a free 30-minute discovery call. We’ll walk one real path, a precedent search or a matter intake, name the reachable layer underneath it, and show you the highest-value place to close the gap, before anyone promises a build. → https://outlook.office.com/book/Consultation@realmindsai.au/?ismsaljsauthenabled

There is a clause being written from scratch somewhere in your firm tonight that a colleague already perfected years ago. The only question worth asking is whether your systems will ever let the next lawyer find it.