Legal AI stopped being experimental and became, in the words of one firm, "table stakes." But the honest picture is narrower and more useful than the marketing: the value shows up in specific tasks — drafting, summarisation, due diligence, data extraction, translation — the moat is fluency, not the tool, and the real disruption is to the billable-hour business model. This is a category map, not a product pitch: what the tools do, how to read the evidence, and the POPIA-and-privilege reality of running any of it in a South African firm — a domain where getting it wrong isn't an option.
The single most useful thing to understand about legal AI is that it is not, yet, good at "being a lawyer." It is good at a specific, recurring set of tasks inside legal work — and knowing which ones is the whole buying decision. Across the largest independent usage study to date (RSGI, commissioned by Harvey, November 2025), the tasks that consistently produced value were the same five, in roughly this order.
| Task | What the tool does | Where it lands hardest |
|---|---|---|
| Drafting | First-pass clauses, correspondence, memos from a brief | Transactional practice; high-volume, template-adjacent work |
| Summarisation | Condensing long documents and matters to their operative points | Litigation, board packs, incoming instructions |
| Document review / due diligence | Reading a data room, flagging issues, building the issues matrix | M&A, financing, the top in-house use case |
| Data extraction | Pulling defined fields from many contracts at once | Data-heavy matters; where the biggest time savings are reported |
| Translation | Working across languages in client correspondence and documents | Cross-border firms; non-English-first jurisdictions |
Notice what is not on that list: judgement, strategy, advocacy, and the client relationship. The study's participants were emphatic that the top benefits were about work fulfilment, time, and client time — not the work product replacing the lawyer. The value is in clearing the drudgery so the judgement has room to happen.
The most cited data point in the market right now is the RSGI study — 40 organisations (29 law firms, 11 in-house), interviewed September–October 2025. Its headline findings are striking: 68% of organisations saw measurable benefit within three months, "power users" (typically 20–30% of lawyers) reported roughly double the time savings of standard users, average monthly usage of purchased licences ran around 92%, and 100% of law-firm participants agreed their lawyers would be upset if the tool were taken away.
Those numbers are useful, and they are not neutral. The study was commissioned by Harvey and scoped to Harvey's own customers — so there is no control group, no comparison against firms that didn't buy, and built-in survivorship bias (only customers who bought and kept the tool were asked). Time savings are self-estimated by participants, usage data was supplied by the vendor, and the in-house sample (n=11) is small enough that the report itself calls it "less robust." Read it as a well-run study of satisfied buyers, not as proof of category-wide ROI. The structural findings — task fit, the power-user curve, the business-model shift — corroborate what practitioners report elsewhere; the specific percentages should carry the asterisk, not the argument.
"Legal AI" is several different products wearing one label. The buying decision turns on whether your work ends in a draft, a data-room review, or a filing that needs a defensible source trail. The strengths below are drawn from independent market coverage (2026) — comparison analysis, not primary benchmarks — and the market moves quarterly.
| Tool | Shape | Strongest at | Grounding |
|---|---|---|---|
| Harvey | Enterprise firm + in-house platform | High-volume review, issue tracking, breadth; the "Swiss army knife" | Model-based; Vault for documents |
| Legora | Enterprise platform | Tabular / matrix review, cross-border and multilingual matters | Model-based |
| CoCounsel (Thomson Reuters) | Research engine | Answers grounded in trusted legal content; the defensible source trail | Database-grounded |
| Lexis+ AI / Protégé (LexisNexis) | Research engine | Citation-backed research where the answer ends in a filing | Database-grounded |
| Spellbook | Word-native drafting/review | Transactional drafting and markup inside the document | Model-based, in-Word |
| Luminance · Robin AI | Contract specialists | Contract review, negotiation, and lifecycle management | Model-based, contract-tuned |
The sharpest line in the field is model-based vs database-grounded. Harvey, Legora and Spellbook reason with a model over your documents — excellent for drafting and review, but the authority is only as good as what you fed it. CoCounsel and Lexis+ AI ground answers in a licensed corpus of case law and statute, so the output carries a citable source trail. If the work ends in advice or a filing that has to survive scrutiny, database grounding changes the decision. Many firms run both, for different jobs.
The finding that matters most is not a productivity number — it's a market one. A US firm in the study described the tool as "table stakes, not a differentiator." When the capability is table stakes, the competition moves to how you price and package it, and legal AI is quietly prising apart the billable hour. The study documented firms moving to fixed-fee and managed-service models (A&O Shearman's ContractMatrix and Vantage), subscription pricing, first-pass due diligence as a standalone service, and workflow-based pricing for repetitive work — all priced on outcome, not hours.
The mechanism is simple and uncomfortable: AI is absorbing the commodified mid-tier work that used to fill associate hours, which lets firms price that work competitively and reserve the hourly premium for genuine judgement. The firms leaning in are taking on previously unprofitable mid-tier mandates to win larger ones later. This is the same move the strangle-the-licence pattern describes, arriving in the one profession whose entire economics were built on billing time.
Across the study, the same pattern appeared everywhere: value compounds with fluency. "Power users" — roughly 20–30% of lawyers at a firm — delivered about double the time savings of standard users, and firms with the longest tenure showed the highest impact. This is the "Harvey fluency curve," and it is the same shape the software branch describes for coding agents: the tool is undifferentiated; the skill of the operator is where the returns live.
If power users deliver twice the return and are a minority of the firm, then the highest-leverage investment is not more licences — it's deliberately growing the power-user cohort: internal case studies, named champions, and the "use it or lose it" licence reallocation the best-run firms already practise. The tool is bought once; the fluency compounds. This is the legal-sector expression of the tree's recurring thesis: own the harness and the talent, rent the model.
The uncomfortable subtext of the whole study is that the sector cannot yet measure ROI cleanly — only a fifth of participants had a formal framework. Firms fall back on adoption rate, intensity of usage, and power-user performance as proxies, plus self-estimated time savings. One participant put the problem exactly: "how would you measure the value of Microsoft Word?" Once a tool is embedded in the workflow, its value becomes real and unmeasurable at the same time.
Treat self-reported hours saved as directional, never as banked savings — time saved on a first draft is partly repaid in verification, and the net is what matters. Watch adoption and intensity as leading indicators, but hold that usage is not value: a licence used daily on low-stakes work is not the same return as one used weekly on a bet-the-company matter. The defensible metrics are the boring ones — matters delivered, cycle time on merged work product, realisation on fixed-fee mandates — not a headline hours-saved figure that no one can reconcile.
This is a domain where getting it wrong isn't an option, so the "skip" cases carry more weight than the "use" cases.
Model-based tools can invent a plausible clause, citation, or fact. In most fields that's an annoyance; in law it's a filed error with a client's name on it. Grounding and verification aren't optional extras — they're the product.
Every draft the model produces still has to be read by someone accountable. The honest time saving is drafting time minus checking time — real, but smaller than the headline, and the study's participants said aggregating it across complex matters is genuinely hard.
The enterprise tools run four-to-five figures per seat on annual contracts, and the market leader doesn't publish pricing at all — you negotiate. Workflows, prompt libraries, and integrations build switching cost fast. Own your document layer and evals so the model stays swappable.
Legal professional privilege can be jeopardised if privileged material is processed somewhere it shouldn't be, or in a way that arguably shares it with a third party. The data path is a legal question before it's a technical one — design it before week one.
High usage numbers are easy to celebrate and easy to fake as value. A firm can hit 92% monthly usage on low-stakes drafting and see no change in realisation or client outcomes. Measure what shipped, not how many prompts ran.
The evidence base for legal AI is almost entirely US, UK, and European — the RSGI study included no South African firms — so the SA reality has to be reasoned from first principles, and it changes the buying order. For an SA firm, the first question is not "which tool is best" but "where do the privileged documents sit, and who can reach them."
Client matter data is special personal information under POPIA and is often privileged. Sending it to a US-hosted model is both a cross-border transfer question (POPIA s.72) and a privilege-preservation question. The practical consequence is that the SA-viable options narrow to tools that can keep inference and storage in country — which is the same in-country footing the Copilot-in-SA and Dataverse leaves describe, and the reason the Keep the Data Home briefing exists. A brilliant tool you can't defend under POPIA is not an option for a regulated firm.
The business-model shift the study documents — fixed-fee, managed services, subscription — assumes a market that already prices sophisticated legal work at a level where absorbing commodity work pays off. Much of the SA market still runs on time-and-materials and fixed-scope mandates that incentivise billing hours, not compressing them. An SA firm adopting legal AI to cut drafting time, without changing how it prices, hands the saving to the client and keeps none of it. The tooling is the easy part; the commercial model is the decision.
Legal AI is the legal-sector expression of theses the tree already holds: ground the model in owned data, own the harness not the weights, and keep the data in the country.
The RSGI report is the load-bearing data source and is labelled for what it is — vendor-commissioned. Vendor sites are primary for what each tool claims to do; the market coverage is secondary and moves fast.