The editorial moat just collapsed.

If you take a JADE Professional or LexisNexis subscription and strip out the editorial layer, what you are left with is raw text retrieval against a case corpus. AustLII gives that away for free. The premium tools earn their price through something more structured: headnotes that summarise the holding, catchwords that index the case, a topic taxonomy, treatment labels that say whether a later case followed, distinguished, or overruled the earlier one, practice points that translate the ratio into an actionable rule, and a citation graph that connects everything.

That editorial layer used to be the product. It was written by people. Its cost structure was, in the main, payroll.

What the editorial layer used to cost.

The exact size of a premium research tool’s editorial team is not usually disclosed. Working backwards from public output volume and turnaround times for editorial updates on cases of the last five years, we estimate the sustained team for a JADE-sized corpus at somewhere between twenty and forty full-time lawyers. At a fully loaded cost of roughly A$120,000 per editor per year, including management overhead, tooling, and workflow supervision, annual editorial payroll lands in the range of A$2.4 to A$4.8 million for the upper bound team.

Take the upper end as a working figure: roughly A$4 million per year for the editorial layer on a 1.2 million case Australian corpus. That is the number that priced A$95 per month for JADE Professional, A$171 and up for individual LexisNexis subscriptions, and A$30,000 to A$60,000 per year for most small and mid-sized firm Lexis licences.

None of this was unfair. Editorial review at this scale is real work by real experts, and the price reflected the cost.

What happens when a model writes the headnote.

Run a capable reasoning model over a judgment with a careful prompt and it returns a three-paragraph headnote, a catchword list that maps to a known taxonomy, a practice point extracted from the ratio, and a topic classification. Per-case compute cost, at current published API rates, sits at roughly A$0.02. The number drifts as models and prices change, but it has been below A$0.10 per case for more than a year.

Multiply that out. For a one-pass editorial run over a 1.2 million case Australian corpus the bill is around A$24,000. The steady-state cost of staying current as roughly 1,000 new judgments are published per day is around A$20 per day, or A$7,000 per year.

Total editorial cost for an AI-driven stack running the modern equivalent of what a traditional research team produces, on the same corpus: less than A$50,000 per year.

The ratio between the old and new editorial cost structures is roughly eighty. It is not a rounding error. It is the collapse of the single largest line item on the cost side of a premium research tool.

The counterargument, taken seriously.

The obvious objection: AI editorial is not equivalent to human editorial. A trained lawyer reading a difficult appellate decision spots nuance that a model does not. Landmark cases deserve the extra care. Edge cases, especially in fast-moving fields like migration and taxation, reward someone with the time and the judgement to sit with them.

We agree. This is precisely why the product has to be designed around “Authority not verified” as a first-class output, why human review is a hard gate before any substantive AI output goes to a client, and why we publish the confidence score for every extracted field rather than pretending otherwise.

The honest summary is that AI editorial is more consistent than human editorial at scale (the model does not vary by reviewer), and less nuanced than human editorial at the edge. For the 99 per cent of cases a working lawyer uses in a given year, consistency at scale is the property that matters. For the 1 per cent where nuance matters, lawyer review is the right answer regardless of who wrote the headnote. AI editorial does not claim to replace that judgement.

What becomes true about pricing.

Under the old cost structure, A$95 per month was a rational price for JADE Professional, and A$30,000 to A$60,000 per year was a rational firm price for the full LexisNexis surface. Both reflected a gross margin on the order of 20 per cent after editorial and infrastructure.

Under the new cost structure, a new entrant can offer the same load-bearing 85 per cent of editorial coverage, on a larger and fresher corpus, at a materially lower price and still run gross margin above 95 per cent on editorial. CaseSharp needs roughly 500 to 1,000 paying student subscribers at A$5 to A$10 per month to cover editorial costs. Matter Desk needs roughly 150 paying firms at typical small-firm pricing.

The same arithmetic explains why CourtAid, Habeas, and CaseNote are able to charge a fraction of the incumbent prices while still running viable businesses. They have each absorbed the LLM cost curve and adjusted their pricing accordingly. The open question is which of them assembles an authority engine, a matter-linked workflow, and a dual-product pipeline into a single product. That is the assemblyMatter Desk and CaseSharp are building together.

Why 2026 and not 2030.

Cost curves rarely move prices overnight. Incumbent research tools do not walk away from A$4 million per year editorial teams within a planning cycle. Career editors, union considerations, institutional customer expectations, and internal risk tolerance all pull toward a slow glide. The economics have changed. The organisational response has not, yet.

That lag is the window. A new product built around an AI-first editorial layer from day one can ship with the load-bearing research coverage, at a different price point, with a different margin profile, aimed at a segment the incumbents cannot reach without disrupting their own pricing. That is the definition of disruptive entry, and the window for the Australian small-to-mid firm segment closes when either the incumbents reprice, the PMS vendors finish bundling, or the global enterprise tools localise. Any of those could land within eighteen months. We are building for the window we have.

A note on what this argument does not claim.

The argument above is an economics claim, not a product claim. It says that the labour cost that priced the old premium research tools is no longer defensible. It does not say that shipping a good AI legal research product is easy, or that editorial quality at scale is free of engineering, or that any team with API access can produce something firms will bet live matters on.

Authority trust, matter-linked workflow, careful commentary discipline, human-review gating, and the hundred small product decisions that make research feel safe are still the hard part. This piece is about why the pricing ceiling has moved, not whether building the product is straightforward. Reading it as a claim that “AI ate legal research” misses the point.

Assumptions and sources.