Legal AI: what is actually working now?
In our Legal Tech market deck, you will find everything you need to understand the market
SUMMARY
Legal AI: what is actually working now? Research, document analysis, contract review, first drafts, due diligence and litigation preparation are already working in serious legal workflows; autonomous legal judgment still is not.
The clearest sign that legal AI has moved beyond novelty is repeat usage. Major deployments now show lawyers coming back week after week, sometimes at usage rates above 90%, long after the pilot phase.
The strongest legal AI use cases share one practical feature: the task is bounded and the output is easy to verify. Summaries, clause comparisons, document extraction and first-pass review all give lawyers something concrete to check against the source material.
Legal research has improved enough to become a serious first pass, but the reliability ceiling still matters. Current systems can find and synthesise legal material quickly while still making enough mistakes that the underlying authority has to be opened and checked.
Contract work may be the best fit for legal AI today because speed and consistency compound across large volumes. On one agreement the gain may be modest; across hundreds of agreements, the same review logic can suddenly cover the full population instead of a sample.
Specialist legal AI is not winning simply because it always has a better model. Its real advantage increasingly comes from workflow: permissions, firm knowledge, playbooks, legal sources, Word integration, document context and traceable outputs.
The productivity gains are now large enough to affect legal economics. Several real deployments report four, six, eight or more hours returned to lawyers each week, and some in-house teams are already using that capacity to reduce outside-counsel work.
That creates a tension for law firms. In-house teams can capture AI savings directly, while hourly-billing firms may destroy some of their own billable time unless pricing changes toward fixed, capped or outcome-linked structures.
Agents are useful when they chain together defined steps such as intake, review, extraction, redlining and reporting. The evidence becomes much thinner once the workflow reaches a decision shaped by commercial priorities, litigation strategy or professional responsibility.
The biggest unresolved issue is training. Legal AI is taking over many tasks junior lawyers historically learned from, so firms may need to replace passive apprenticeship through repetitive work with much more explicit coaching around judgment, context and why an AI answer is good or bad.
Why does legal AI suddenly feel real now?
Legal AI feels real today because lawyers are using it repeatedly for actual work, and some of the biggest deployments are now old enough to show whether people kept using the tools after the novelty wore off.
The latest LexisNexis survey of 543 legal professionals is unusually clear on this. Ninety-four percent said they now use AI for legal work, 74% use it at least weekly and 34% use it every day. Legal research leads at 69%, followed by document summarisation at 62%, while drafting and document review are both at 53%.
The numbers are high enough that we should be careful about comparing them directly with every other survey because samples and definitions differ. The direction, though, keeps repeating. Thomson Reuters' latest Future of Professionals research found 74% of professionals across legal and other professional services using AI several times a week. In Asia-Pacific, a separate LexisNexis survey of 1,715 legal professionals across ten markets found 94% using either general-purpose AI, specialist legal AI, or both.
Actual deployments give us a better test than surveys. Repsol started with 50 lawyers across 12 jurisdictions in early 2024. Two years later, the company says 96% of its roughly 200-person legal department uses Harvey, with each lawyer saving four to six hours a week. GÖRG, a German commercial law firm with more than 380 lawyers, reports 96.81% weekly usage after rolling out Legora across five offices.
The vendors have also become substantial businesses. Harvey's CEO recently said the platform is used by roughly 200,000 lawyers and generates around $350 million in annual recurring revenue. Legora says it has passed $100 million in ARR and serves more than 1,000 organisations.
Those figures do not tell us whether AI can handle a lawsuit or negotiate an acquisition by itself. They tell us something simpler and more useful: lawyers have found enough recurring value to keep opening these products after the pilots ended.
And that is where the interesting part starts.
What should count as legal AI actually working?
Legal AI is working when it reliably removes a meaningful chunk of lawyer time from a real task and the remaining review takes less effort than doing the whole job manually.
That sounds obvious, but it rules out a lot of impressive demos.
A system that produces a beautiful legal memo in 30 seconds has little value if a lawyer then spends two hours checking every proposition. A contract tool can be extremely valuable even if the lawyer rewrites part of its first draft, provided it has already found the right clauses, spotted deviations and built a usable starting point.
Verification changes the answer dramatically.
Summarising a 150-page brief is attractive because the lawyer still has the document. If the AI says something important happened on page 73, the lawyer can inspect page 73. Comparing 300 contracts against the same change-of-control rule works for the same reason. The machine does the repetitive search and the lawyer checks the exceptions that matter.
An open-ended question such as whether a company should sue a former distributor is much harder. Facts may be incomplete. Several legal theories may apply. Commercial relationships matter. Strategy matters. A beautifully written answer can still point the client in the wrong direction.
Across current deployments, the strongest dividing line is surprisingly practical: legal AI works best when the job is bounded and the output is easy to verify.
| Legal task | How well legal AI works today | What the lawyer still does |
|---|---|---|
| Document summaries | Very well | Checks important passages and omissions |
| Contract extraction and comparison | Very well | Reviews exceptions and commercial significance |
| First-pass contract review | Very well | Decides what risk is acceptable |
| First drafts and redlines | Well | Adjusts judgment, leverage and deal context |
| Legal research | Well, with caveats | Verifies authorities and propositions |
| Litigation-file analysis | Well | Decides case strategy |
| Multi-step legal workflows | Increasingly useful | Approves important decisions |
| Autonomous legal advice | Still weak | Lawyer effectively remains in charge |
If you want more recent data on this point, please see our latest Legal Tech market report.
This market map, featured in our Legal Tech market deck, highlights top companies and startups in the Legal Tech market
Can lawyers trust AI for legal research now?
Legal AI research is good enough to use as a serious first researcher today, but treating it as the final check is still reckless.
The improvement from the first generation of products has been large.
Stanford researchers previously tested Lexis+ AI, Ask Practical Law AI and Westlaw's AI-assisted research system and found incorrect information in more than 17% of responses from the first two products and more than 34% from Westlaw's tested system. Those results helped puncture the idea that adding a legal database automatically solved hallucinations.
Newer testing looks much better.
Vals AI built a private benchmark of legal-research questions and scored answers on substantive accuracy, authoritative sourcing and usability. Specialist legal products landed around the mid-to-high 70% range, while ChatGPT reached a similar range and the tested human-lawyer baseline came in lower.
Vals' separate CaseLaw benchmark gives us another useful reality check. On difficult questions drawn from court cases, even leading frontier models currently sit around the low-70% range rather than approaching flawless performance.
That combination is more informative than a claim that AI has “beaten lawyers.” A model can now be extremely competitive at finding and synthesising legal material while still getting enough questions wrong that blind reliance would be absurd.
Lawyers seem to have reached much the same conclusion. In LexisNexis' latest survey, 69% said they use AI for legal research. At the same time, 81% said they are more comfortable when AI answers are grounded in legal sources, up sharply from the previous year.
This is why useful legal-research products increasingly look less like an empty chat box. Westlaw, LexisNexis and other specialist tools surface authorities alongside the answer. Harvey connects research with firm knowledge and source material. Lawyers can move quickly from the generated conclusion to the underlying case, statute or document.
For now, that is the sweet spot. Let AI find the path through the library. Before anything consequential leaves the firm, a lawyer still needs to open the books.
Can legal AI draft contracts lawyers would actually use?
Legal AI can already create useful first drafts and redlines for routine commercial contracts, especially when it has access to the organisation's own templates and negotiation rules.
Independent testing gives us a clearer picture than vendor demos.
LegalBenchmarks.ai compared 13 AI systems with human lawyers across 30 contract tasks, producing 450 evaluated outputs. Human lawyers met all of the benchmark's reliability criteria on 56.7% of tasks. AI averaged 57%, while the best individual systems reached roughly 73%. The best tested human reached 70%.
Speed was the huge difference. Lawyers averaged 12 minutes and 43 seconds per task. AI produced drafts in under a minute.
The result should not be read as “AI is a better contract lawyer.” Human lawyers still scored slightly higher on overall usefulness and were stronger when drafting required commercial context, nuanced instructions or deciding when conceding a point would be a bad idea.
What the test does show is that first-draft generation has become real work rather than a toy use case.
The products are also moving toward the way contracts are actually negotiated. Spellbook runs inside Microsoft Word. Harvey and other platforms can use internal precedents and playbooks. A legal team can ask the system to compare a counterparty's clause against its preferred language, explain the difference and draft a replacement.
This is a much stronger workflow than asking a generic chatbot to “write an NDA.”
The benchmark found something else that helps explain why specialist products survive even when general models draft well: two-thirds of the legal AI tools tested integrated with Word, while specialist products were much better at carrying templates, context and internal rules through the workflow.
Contract drafting today is therefore less about generating legal prose. The useful part is generating the right prose inside an existing negotiation process.
As this chart shows, and as featured in our Legal Tech market deck, search interest in Legal Tech has been growing steadily
Is contract review the clearest legal AI success today?
Contract review is currently one of the clearest legal AI successes because machines are very good at doing the same check hundreds of times without getting tired.
The basic workflow fits current AI unusually well.
A lawyer reviewing a supplier agreement may want to know whether the limitation of liability matches policy, whether indemnities are reciprocal, whether assignment requires consent, whether auto-renewal is allowed and whether data-security terms deviate from the company's standard.
Those questions can be written down. The relevant text already exists. Every AI finding can be linked back to a clause.
Once we move from one agreement to 500, the advantage becomes much larger.
Danway says CoCounsel cut its contract-review time by about 30% to 40%, saving roughly 11 hours each week while allowing the team to handle more tenders simultaneously. A financial-services legal department interviewed by Thomson Reuters reported cutting document-review time by 50% to 75%. A 100-page brief that previously took two to three hours to assess could be triaged from an AI summary in about 15 minutes.
A Forrester study commissioned by Thomson Reuters provides a broader economic example. Across 26 active users at the organisation it modelled, lawyers reportedly saved roughly half an hour to one and a half hours per day on legal research, while contract-review time fell by as much as 75%. The study calculated a 222% three-year return on the CoCounsel investment after adjusting its assumptions for risk.
Commissioned studies and vendor customer stories show what is possible; they should not be treated as an industry average. Still, the same workflow keeps producing large gains across unrelated organisations.
There is also a quality angle that gets less attention. Human teams facing 1,000 agreements often sample because checking everything costs too much. AI makes it realistic to run the same question across the full population.
A system that helps a lawyer examine all 1,000 contracts can sometimes improve the review even if the machine itself is less perceptive than the best lawyer in the room.
If you want more recent data on this point, please see our latest Legal Tech market report.
Is legal AI actually useful in litigation?
Legal AI is already useful in litigation for finding facts, rebuilding timelines, analysing large case files and preparing first drafts, while the real case strategy still sits with the lawyers.
Litigation contains a huge amount of work that happens before anyone makes the clever argument.
Teams read emails, witness statements, contracts, expert reports and deposition transcripts. They reconstruct who knew what and when. They compare testimony. They search thousands of pages for one conversation. They prepare chronologies and lists of contradictions.
Current AI can compress that work aggressively.
LPHS, a Dallas litigation boutique, says lawyers use Harvey for factual extraction, research, drafting, deposition preparation and case analysis, saving more than eight hours a week in some cases. In one competitive pitch involving hundreds of documents, the firm says it responded within 48 hours while rival firms expected roughly a week.
The same firm described receiving hundreds of documents on the morning of a mediation and using AI to analyse the batch immediately. Without that capacity, lawyers would have had to decide which fraction of the material they had time to read.
Thomson Reuters reports a similar pattern among CoCounsel users. Its current litigation material estimates savings of roughly 240 hours a year for some litigators, or around six working weeks, with certain large document reviews falling from days to less than an hour.
The headline number will vary wildly by matter, and vendor estimates should be treated carefully. The underlying advantage is easier to believe: AI lets a litigation team search and synthesise far more material before the deadline arrives.
What AI still struggles to own are the questions that make litigation litigation. Which witness will survive cross-examination? Which fact will bother the judge? Should we settle at $4 million? Is the technically stronger argument worth making if it distracts from the cleaner story?
A model can now put much more of the case on the table. Experienced litigators still decide what to do with it.
This chart, featured in our Legal Tech market deck, illustrates yearly venture capital funding for Legal Tech startups
Are Harvey and other legal AI tools really better than ChatGPT or Claude?
Specialist legal AI earns its place today mainly by making general AI usable inside serious legal work, rather than through some permanently superior legal brain.
The independent comparisons have made that increasingly difficult to ignore.
In LegalBenchmarks.ai's contract study, general-purpose tools averaged 58.3% reliability and specialist legal AI averaged 57.6%. The usefulness scores were almost identical. Some of the strongest individual drafting results came from general frontier models.
Vals found a similar pattern in legal research, where a general AI system came surprisingly close to the specialist platforms.
So why are organisations paying for Harvey, CoCounsel, Legora, LexisNexis, Spellbook and similar products?
Because the model is only one layer of a legal workflow.
The lawyer needs access to the correct matter. The system has to respect permissions and confidentiality. It may need the firm's precedents, a customer's negotiating playbook, reliable case law, document versioning and Microsoft Word. Ideally, every important factual or legal statement should point somewhere the lawyer can inspect.
LegalBenchmarks.ai's contract research makes this concrete. Specialist tools were much stronger on workflow support, and 66.7% integrated directly with Word. In high-risk drafting scenarios, specialist systems also raised explicit legal-risk warnings in 83% of outputs, versus 55% for general AI.
The model underneath these products can change. The firm's precedents, permissions, integrations and workflow history are much stickier.
That is where a lot of the durable value in specialist legal AI is moving.
If you want more recent data on this point, please see our latest Legal Tech market report.
Are lawyers really saving enough time for legal AI to matter?
Yes. Legal AI is now saving several hours a week in enough real deployments that the productivity gain is too large to dismiss as convenience.
There is no universal “hours saved” figure. A litigator working through discovery and a partner negotiating bespoke M&A agreements will get very different results.
Still, the current range is informative.
Repsol reports four to six hours saved per lawyer each week. LPHS says some litigators save more than eight. Danway reports about 11 hours across its contract workflow. The financial-services department using CoCounsel says the tool saves several hours weekly. Forrester's recent commissioned study found between half an hour and one and a half hours saved per day on research among 26 active users.
Even three hours a week becomes large over a year.
Assume 46 active working weeks. Three hours a week gives one lawyer 138 hours back. Across 100 lawyers, that is 13,800 hours. At five hours weekly, the same team recovers 23,000 hours.
We should then ask where those hours go, because “time saved” can be misleading. Lawyers may use the capacity to handle more matters, reduce outside-counsel work, answer the business faster or simply work fewer late evenings. Law firms may convert some of it into higher caseloads.
The order of magnitude is what matters. We have moved well beyond saving five minutes on an email.
| Reported use of legal AI | Reported time gain |
|---|---|
| Repsol with Harvey | 4–6 hours per lawyer per week |
| LPHS with Harvey | More than 8 hours per lawyer per week in some cases |
| Danway with CoCounsel | About 11 hours per week |
| Forrester CoCounsel study | 0.5–1.5 hours per day on research among 26 active users |
| Financial-services CoCounsel deployment | 50–75% less document-review time |
This chart, featured in our Legal Tech market deck, looks at Clio’s strategy in Legal Tech
Will in-house legal teams get more from AI than law firms?
In-house legal teams currently have the easiest way to turn legal AI productivity into money because every task kept inside the department can reduce backlog, outside-counsel bills or pressure to hire.
The economics for a law firm are more awkward.
Imagine AI reduces a six-hour associate research job to two hours. The corporate legal department has recovered four hours immediately. Under traditional hourly billing, the law firm may have just removed four billable hours from its own invoice.
That difference is already showing up in client behaviour.
A financial-services legal department using CoCounsel estimates that it has cut outside-counsel spending by $100,000 to $200,000 a year. Other corporate legal teams describe keeping research and contract review in-house because AI makes work manageable with existing headcount.
The latest Thomson Reuters legal report shows the pressure moving directly toward law firms. Seventy-one percent of in-house legal professionals expect outside firms to change how they charge as AI use grows. Only 28% of law firms say they have already changed their pricing in response.
Clients also seem unimpressed with the benefits they currently receive. Seventy-seven percent say AI-enabled quality improvement from outside firms is very important or essential, yet only 5% say they are getting that from most or all providers. Thirty-three percent have already reconsidered, or expect within 12 months to reconsider, relationships where firms fail to show AI-enabled value.
Those numbers make the billing debate much less theoretical.
Hourly billing will survive where scope and outcomes remain genuinely unpredictable. Complex litigation can expand overnight. An acquisition can collapse and restart. Regulatory work can take unexpected turns.
Routine document-heavy work is harder to defend the same way. Once a review that used to consume five hours reliably takes one, clients will eventually want to know who keeps the other four.
That pressure favours fixed fees, capped fees and other arrangements where a firm can keep some of the productivity gain while the client pays less for the outcome.
Legal AI may therefore hit the economics of legal services before it hits lawyer headcount.
If you want more recent data on this point, please see our latest Legal Tech market report.
Can legal AI agents run a whole legal matter now?
Legal AI agents can now handle useful chains of work on their own, but we still have very little evidence that they can safely own an entire legal matter from beginning to end.
The word “agent” makes the current products sound more autonomous than many of them really are.
A useful contract agent can receive an agreement, compare it with a playbook, flag unacceptable clauses, produce redlines and prepare a summary. A due-diligence workflow can inspect hundreds of documents, extract requested information and assemble the findings into a table or report.
That is already more valuable than prompting a chatbot five times manually.
The latest products are moving aggressively in this direction. CoCounsel now markets multi-step workflows across research, document analysis and drafting. Legora is building agents that can work across large document sets. Spellbook has pushed contract workflows toward systems that move from intake through review and suggested revisions.
There is strong demand for that structure. LexisNexis' latest Asia-Pacific survey found that 69% of legal professionals would use AI more if it came through guided workflows.
The harder step appears when the workflow reaches a decision that depends on goals nobody fully wrote down.
Should we concede this warranty because the customer is strategically important? Is this litigation worth continuing after the new evidence? Does a technically available argument create a reputational problem? Should the company accept a higher liability cap to close before quarter-end?
Today's systems can feed those decisions with far better preparation.
Giving the decision itself to the agent is where the evidence becomes thin.
This chart, featured in our Legal Tech market deck, illustrates yearly funding for Legal Tech startups
Are hallucinations still a serious problem in legal AI?
Yes. Hallucinations remain a serious legal AI problem today, and their importance actually grows as lawyers use AI more often.
The newest surveys show how firmly accuracy remains at the top of lawyers' concerns.
LexisNexis found 83% of legal professionals worried about lawyers relying on inaccurate or fabricated AI information. Confidentiality came well behind at 53%. Its Asia-Pacific study found accuracy and hallucination risk leading concerns there too, cited by 63% of respondents.
Those fears are easy to understand because legal errors have a nasty asymmetry.
If AI summarises 99 contracts correctly and misses one unusual clause, that one clause can still contain the provision that blows up the deal. If a research assistant finds nine real cases and invents the tenth, the fabricated case can still end up in a court filing.
Courts have now seen the second problem repeatedly. Judges in several jurisdictions have sanctioned lawyers for submitting invented cases, fake quotations and misrepresented authorities produced with generative AI. The professional line emerging from those decisions is straightforward: using AI does not excuse the lawyer from checking what gets filed.
The recent benchmark evidence leads us to the same conclusion from another direction.
LegalBenchmarks.ai's newly updated contract-workflow leaderboard uses a tougher pass criterion than its earlier study: the output passes only when every applicable substantive requirement is satisfied. On those harder workflows, even leading applications currently pass only around a third of tasks.
That number should not be compared directly with the earlier 57% reliability scores because the methodology changed. It is still a useful antidote to the idea that modern models have quietly solved legal reliability.
Lawyers have discovered a tool worth using constantly before the tool has become safe to trust constantly.
That tension defines legal AI today.
Can legal AI actually make the work better, rather than just faster?
Legal AI can improve the completeness and consistency of some legal work, especially when the alternative is rushed human review or sampling.
This quality argument is strongest on repetitive tasks.
Suppose a team has 800 agreements and three days to identify every contract with an unusual termination provision. A senior lawyer may be better than AI at interpreting a difficult clause. The senior lawyer still cannot realistically inspect all 800 agreements alone.
AI can.
That changes what “better” means. Instead of sampling 80 agreements carefully, the team can run the first review across all 800 and send questionable cases to lawyers.
Independent drafting research found another interesting pattern. In scenarios containing significant legal risks, specialist legal AI explicitly warned about those risks in 83% of outputs, compared with 55% for general AI. The human participants in that particular test did not explicitly flag the risks.
The sample is too small to prove that machines have better judgment. It does illustrate one advantage that scales well: software can apply the same checklist every single time.
There is a downside. AI can also make mediocre thinking look polished. A convincing summary may discourage someone from reading the underlying document. A clean first draft can anchor the lawyer around the machine's framing. Research presented with impressive confidence can make the missing case harder to notice.
The best workflow we see today uses AI for breadth and humans for importance.
Have the machine read everything. Have the lawyer decide what deserves attention.
This chart, featured in our Legal Tech market deck, compares the main business model options for Legal Tech SaaS platforms
Will legal AI replace junior lawyers first?
Legal AI is already eating into junior-lawyer tasks, and the bigger unresolved problem is how firms train future senior lawyers when AI does much of the work people used to learn from.
Document review, narrow research, chronologies, citation checks and first drafts have historically filled a large part of junior practice.
They are also among today's strongest AI use cases.
That does not automatically mean firms can delete the junior layer. Someone still has to become the fifth-year associate who knows when an argument is weak, the partner who spots the strange indemnity or the litigator who notices that a witness's story feels wrong.
Traditionally, much of that judgment was built while doing lower-level work.
Thomson Reuters' recent research suggests lawyers themselves see the problem. Forty-eight percent of legal professionals said they are concerned about AI's effect on developing independent judgment. The report estimates that professionals expect AI to lengthen the path toward trusted professional judgment by almost two years.
At the same time, preserving inefficient work purely as a training exercise will become increasingly difficult. Clients are unlikely to pay ten associate hours for something the firm can complete in two with AI.
Firms therefore need a different apprenticeship model. Junior lawyers can move into higher-level analysis earlier, but senior lawyers may have to explain more explicitly why an AI answer is good, incomplete or commercially stupid.
The staffing effect is still unclear. Firms could hire fewer juniors. They could keep similar numbers and handle much more work. Smaller teams could compete for matters that once required large associate pyramids.
What already looks likely is a change in what “junior work” means.
If you want more recent data on this point, please see our latest Legal Tech market report.
So, what legal AI is actually working now?
Legal AI is clearly working today for research, document analysis, contract review, first drafts, due diligence and litigation preparation; autonomous legal judgment remains far behind those use cases.
We have enough evidence now to say more than “AI looks promising.”
Lawyers are using it every week and increasingly every day. Large deployments still show usage above 90% long after the initial pilot. Independent benchmarks find frontier AI competitive with human lawyers on bounded drafting and research tasks. Corporate teams report contract-review reductions of 50% to 75% in some workflows. Several deployments report four, six, eight or more hours returned to lawyers each week. Clients are already asking law firms to pass some of those productivity gains back to them.
At the same time, recent evidence gives us a clear ceiling. The toughest independent contract benchmarks still expose large failure rates. Lawyers remain deeply worried about fabricated information. Courts continue to see AI-generated fake authorities. Multi-step agents are getting much better at completing a process, while final legal and commercial decisions still depend heavily on context that no prompt captures neatly.
The pattern across all of this is unusually consistent.
Legal AI works best when there is something concrete to read, a specific job to perform and a practical way for a lawyer to check the answer. The further we move toward ambiguous facts, strategy, negotiation and professional responsibility, the more important the human lawyer becomes.
So the current legal AI revolution is happening one task at a time.
That may sound less dramatic than replacing lawyers. In practice, it is already enough to change how thousands of lawyers spend several hours of every working week, which work legal departments keep in-house, how firms staff repetitive matters and eventually how clients expect to be billed.
| Legal AI use case | Our judgment today |
|---|---|
| Summarisation and extraction | Clearly working |
| Contract review and comparison | Clearly working |
| Large-scale due diligence | Clearly working |
| First drafts and redlining | Working well with lawyer review |
| Legal research | Strong, but verification remains essential |
| Litigation preparation | Working well on documents, facts and drafts |
| Guided multi-step agents | Useful and improving quickly |
| Independent legal strategy | Still weak |
| Autonomous legal advice | Premature |
| Replacing lawyers end to end | Unsupported by current evidence |
This chart, featured in our Legal Tech market deck, breaks down revenue across customer segments in the Legal Tech market
OUR METHODOLOGY
This analysis tests what legal AI is actually working now by separating sustained adoption, task performance, ease of verification, time saved, workflow integration, economic impact and the point at which human judgment still becomes decisive.
We cross-checked several forms of evidence instead of letting one type dominate. Broad surveys were used to establish how widely and frequently legal AI is being used. Mature deployments showed whether lawyers kept using the tools after pilots. Independent benchmarks tested research, drafting and contract workflows under controlled conditions. Product material showed what current systems can actually do inside legal workflows, while court decisions showed what happens when reliability failures reach real proceedings.
Evidence was weighted according to what it could genuinely establish. Independent research and broad surveys carried more weight for general adoption or performance. Customer deployments were used to show what is already achievable in real legal work and the scale of gains being reported, without treating one customer story as an industry average.
Benchmarks with different tasks, scoring systems or pass criteria were kept separate. We did not treat a legal-research score, a contract-drafting reliability score and a stricter end-to-end workflow pass rate as directly comparable measures of the same thing.
We gave particular weight to whether the legal task could be clearly defined, whether the AI output could be traced back to a case, clause, document or other source a lawyer could inspect, and whether checking the result still saved meaningful time compared with doing the whole task manually.
The final judgments come from convergence. A use case ranked more strongly when independent testing, sustained real-world deployments and current product capability all pointed in the same direction. No single survey percentage, benchmark or customer story determined where a task landed.
Key sources include LexisNexis' legal AI survey, LexisNexis' APAC AI Sentiment Survey, Thomson Reuters' Future of Professionals research, Stanford Law School's testing of leading AI legal-research tools, Vals AI's legal benchmarks, LegalBenchmarks.ai's human-versus-AI contract research, Harvey's Repsol deployment, Legora's GÖRG deployment, Forrester Consulting's commissioned CoCounsel study, and the U.S. District Court material on AI-generated authorities and Rule 11 sanctions.
This chart, featured in our Legal Tech market deck, shows how AI contract review platform technology has evolved over time
Related blog posts
- LegalTech: what are startups building now?
- Legal AI: what’s changing now?
- Legal AI: what is getting real adoption now?
- The latest funding news in the Legal Tech market
