The legal AI stack in 2026
Legal AI has expanded rapidly. The global legal AI market reached $3 billion in 2025 and is projected to grow at 28% annually to $7.1 billion by 2032. Corporate legal adoption more than doubled in a single year — from 23% to 52% between 2024 and 2025. By early 2026, Thomson Reuters CoCounsel had reached one million professional users. These are not niche tools. They are becoming infrastructure.
But the market has grown in one direction: broader, faster, and more automated at the general layer. The dominant players — Thomson Reuters, LexisNexis, Harvey, Luminance, Relativity — have invested overwhelmingly in research automation, document review, drafting assistance, and e-discovery. What they have not built is a layer that takes the output of all of that work and converts it into a structured, calibrated, stress-testable prediction for a single specific case.
That is the gap. And it is structural — not a product gap that the incumbents will close next quarter. It is an architectural gap, because building case-specific outcome intelligence requires something the general platforms are not designed to do: anchoring to the specific facts, legal constructions, conduct patterns, and empirical comparators of one matter.
What existing services cannot do
The existing legal AI market has a clear and acknowledged limitation. Stanford HAI research found error rates of 17% for Lexis+ AI and 34% for Westlaw AI-Assisted Research on legal queries — these are tools designed specifically for legal work. General-purpose models perform far worse on the same queries (58–82% error rate). Over 700 court cases worldwide have now involved AI hallucinations, with sanctions ranging to $31,100 per incident.
The error rates are not a product defect. They are a structural consequence of asking general platforms to do something they are not designed to do: translate legal analysis into a specific, calibrated, defensible probability estimate for a single case. General platforms generate plausible-sounding analysis. AIOOJ generates a probability number grounded in named assumptions, testable against real anchor cases, and stress-tested across a defined sensitivity range.
| Capability | General legal AI (Lexis, CoCounsel, Harvey) | Document platforms (Relativity, Luminance) | AIOOJ |
|---|---|---|---|
| Legal research & case law analysis | ✓ Strong | × Not designed for this | → Consumed as input |
| Document review & e-discovery | ∼ Partial | ✓ Strong | → Consumed as input |
| Drafting & document generation | ✓ Strong | ∼ Partial | × Not in scope |
| Named probability estimate for a specific outcome | × No — generates analysis, not probability | × No | ✓ Core function |
| Sensitivity analysis — which inputs matter most | × No | × No | ✓ Shapley + P2×P4 heat map (3 tabs incl. P10 stress) |
| Stress-tested downside scenarios | × No | × No | ✓ Scenario A, B + P2×P4 heat map |
| Settlement band distribution | × No | × No | ✓ Five bands, live probability |
| Calibration to empirical case law anchors | ∼ General benchmark only | × No | ✓ 14 real anchor cases, 60/40 blend |
| Rational settlement corridor with PV | × No | × No | ✓ PV floor to costs-adj ceiling |
| Real-time recalculation on input change | ∼ Query-response only | × No | ✓ Instant on every slider move |
| Defensible audit trail for counsel & funders | ∼ Depends on prompt quality | × No | ✓ Named assumptions, documented methodology |
What makes AIOOJ different
The distinction is not that AIOOJ uses better AI. It is that AIOOJ uses a fundamentally different analytical approach. General legal AI platforms are language models applied to legal content — powerful at retrieving and synthesising what has been said in prior cases and statutes. AIOOJ is a structured probability engine built from first principles for a specific matter — it models what is likely to happen, not what has been written about what could happen.
Who should use AIOOJ
AIOOJ is not a replacement for Lexis+ AI or CoCounsel. It is the layer above them — the tool you use after you have done the research and analysis, to convert that work into a defensible probability estimate that can be used in settlement negotiations, counsel briefings, and litigation funding discussions.
-
Litigants in person and self-represented parties facing complex commercial disputes where legal costs make comprehensive external advice prohibitive. AIOOJ provides the structured probability framework that solicitors and barristers would otherwise construct implicitly — making it explicit, testable, and documented. This is Harrison v Aegon: the model was built precisely for this use case.
-
Instructed solicitors (Corrs Chambers Westgarth and equivalent firms) who need to provide clients with a defensible, documented basis for settlement positioning and litigation sequencing advice. The model gives Corrs a structured instrument to present to clients — not a recommendation, but a probability framework that clients can interrogate and understand.
-
Briefed counsel preparing for mediation or settlement conferences. The P2×P4 stress table, the Shapley driver ranking, and the stress floor scenarios are specifically designed for counsel briefing — they show which inputs counsel needs to assess, where the model is most sensitive, and what the defendant's rational negotiating range is.
-
Litigation funders assessing whether a case merits third-party funding. The model provides the probability distribution, the expected value, the sensitivity analysis, and the downside floor in a format that funding analysts can interrogate directly — replacing the qualitative "strong case / weak case" assessment with a structured, stress-tested probability engine.
-
In-house counsel and corporate legal departments managing complex commercial litigation on significant claims where settlement decisions need to be documented and defensible at board level. The export function produces a board-ready summary with all key metrics and current assumptions in a single print-ready page.
The market signal AIOOJ is responding to
The legal AI market's own data reveals the gap AIOOJ fills.
The emerging competitive structure of legal AI — as described by market analysts in early 2026 — is characterised by integrated research platforms (Thomson Reuters, LexisNexis), specialist AI providers (Harvey, Legora), and document intelligence platforms (Relativity, Luminance). No major player currently occupies the case-specific outcome intelligence layer. Litigation prediction is listed as a market segment in analyst reports, but the dominant players remain focused on the research and document layers where the volume is higher and the technical requirements are lower.
This is AIOOJ's structural opportunity. The case-specific outcome intelligence layer is not a niche — it is the layer where the highest-value decisions in litigation are made: whether to file, whether to settle, what to accept, when to escalate. Those decisions deserve better than a qualitative assessment. They deserve a calibrated, stress-tested, empirically anchored probability engine.
What AIOOJ is not
Precision about what AIOOJ does not do is as important as what it does. Being clear about this is what makes the outputs trustworthy.
- AIOOJ is not a legal research tool. It does not retrieve case law, search databases, or generate legal analysis. It consumes the output of those tools as inputs and converts that analysis into probability estimates.
- AIOOJ is not a prediction. The probability outputs are modelling estimates — calibrated expert judgements expressed quantitatively. They are not statistical frequencies from a large sample of identical cases. There is only one trial.
- AIOOJ is not a substitute for legal advice. The model structures and quantifies legal analysis. It does not replace the analysis itself. The quality of the model's outputs depends entirely on the quality of the legal judgements entered as inputs.
- AIOOJ does not hallucinate. Every number in the model is derived from a named formula, a documented assumption, or an empirical anchor. There is no generative AI component producing plausible-sounding outputs that may or may not be accurate. The model does arithmetic, logic, and statistics — not language generation.
- AIOOJ is not general-purpose. The current model is built specifically for Harrison v Aegon in the NSW Supreme Court Commercial List. The architecture is transferable — the specific inputs, calibration, and legal constructions are not. A new matter requires a new model built on the same architecture with new case-specific inputs.