Legal AI: what’s changing now?

Has Legal AI become normal everyday legal work?

Yes. Legal AI is already part of normal legal work for a large majority of lawyers, and these days the bigger question is how deeply they are willing to use it.

Clio’s latest Legal Trends Report found that 79% of legal professionals use AI in their firms. More recent LexisNexis research goes further. Its latest UK survey found that 94% of lawyers use AI for legal work, 74% use it at least weekly and 34% use it every day. Across ten Asia-Pacific markets, another LexisNexis survey of 1,715 legal professionals also found 94% AI adoption, up from 88% the previous year.

Those surveys use different samples, so 94% should not be read as a universal global adoption rate. The useful finding is the consistency of the direction. AI use is no longer confined to a small group of innovation-minded lawyers. Research, drafting, document review and summarisation have entered regular workflows across several major legal markets.

Organizations have moved more slowly than their lawyers. In the APAC survey, only 58% of respondents said their organization had an AI policy and 53% said it actually measured the value produced by AI. Thomson Reuters separately found that 34% of law-firm professionals use AI tools their firms have not approved.

Legal organizations used to worry about getting lawyers to try AI. Many are currently trying to catch up with lawyers who already use it.

Measure	Latest reported result
Clio: legal professionals using AI	79%
LexisNexis UK: lawyers using AI	94%
LexisNexis UK: weekly-or-more use	74%
LexisNexis APAC: professionals using AI	94%
APAC organizations with an AI policy	58%
Law-firm professionals using unapproved AI	34%

What can Legal AI agents actually do now?

Legal AI agents can currently take on much larger chunks of an assignment instead of helping with one prompt at a time, especially in structured research and document-heavy work.

The earlier generation was already useful for summarising documents, extracting clauses, drafting emails, producing first-pass research and answering questions about uploaded files. Those tools saved time, but lawyers still had to break a matter into steps and repeatedly tell the AI what to do next.

Thomson Reuters has built Deep Research into CoCounsel Legal so the system can plan a research process, search Westlaw, assess what it finds and assemble a supported answer. Harvey increasingly lets firms build workflows that combine documents, instructions, firm knowledge and several AI actions. A&O Shearman has worked with Harvey on agents covering areas such as antitrust filings, cybersecurity, fund formation and loan review.

A&O Shearman’s work shows what useful legal agents look like in practice. A lawyer can give the system a task, let it retrieve relevant documents or legal material, run several steps, compare findings and return a draft deliverable without prompting the AI after every intermediate action.

Harvey has also created a Legal Agent Benchmark based on complete legal matters. Instead of asking a model isolated questions, the benchmark gives it matter documents, instructions and a work product to prepare. Vendors increasingly care about whether AI can finish a meaningful piece of work rather than whether it can answer another legal quiz question.

The boundaries still matter. These agents generally operate inside defined workflows, approved data sources and specific matter contexts. A lawyer reviews the output and keeps the professional responsibility.

Why are Harvey and Legora growing so fast?

Harvey and Legora are growing so fast because major legal organizations are buying AI as a core work platform rather than as another experimental tool.

Harvey is the clearest example right now. The company has just raised $550 million at a $15.5 billion valuation, less than six months after being valued at $11 billion and roughly nine months after an $8 billion valuation. Harvey says it has now passed $400 million in annual recurring revenue, serves more than 3,000 customers and is used by 80% of the Am Law 100, 20% of the Fortune 500 and half of the Fortune 10.

The valuation almost doubled in roughly nine months, but the business expanded quickly too. Harvey was reportedly generating about $190 million in annualised revenue in January and had moved above $350 million by August before the latest disclosure put ARR above $400 million. Even with the usual caution around privately reported AI revenue metrics, that is unusually fast expansion for legal software.

Legora is following from a smaller base. It raised $550 million at a $5.55 billion valuation earlier this year, then added another $50 million extension. The company passed $100 million ARR and was reported at roughly $150 million in the second quarter, with around 1,500 customers. Major users include White & Case, Cleary Gottlieb, Goodwin, Linklaters and Deloitte.

These companies are winning much bigger budgets than a clause-extraction tool normally would because they are trying to sit across research, drafting, review and workflows. The ambition resembles an AI workspace for lawyers rather than a single legal application.

Company	Commercial scale we can verify	Recent valuation
Harvey	$400M+ ARR; 3,000+ customers; 80% of Am Law 100	$15.5B
Legora	About $150M ARR reported in Q2; around 1,500 customers	About $5.6B
Luminance	1,000+ organizations across 70 countries	Private; raised $75M Series C

Is Harvey going to dominate Legal AI?

Harvey currently has the strongest commercial lead among AI-native Legal AI platforms, but the market is still too open to call it a winner-take-all category.

Its latest numbers make the lead difficult to dismiss. More than $400 million of ARR, more than 3,000 customers and penetration across 80 of the Am Law 100 put Harvey at a scale few young enterprise-software companies reach this quickly.

Harvey is also fighting several different kinds of competitors.

Legora has crossed roughly $150 million ARR and continues to win major international firms. Thomson Reuters can attach CoCounsel directly to Westlaw and Practical Law. LexisNexis can combine AI with one of the largest proprietary legal-information businesses in the world. Google has now entered the category directly with Gemini Enterprise for Legal, covering tasks such as brief preparation, citation verification, contract work and regulatory monitoring.

The large model companies create another pressure. Legal platforms depend heavily on underlying AI capability, while companies such as Google, OpenAI and Anthropic keep making their general models better at long documents, tool use and professional workflows.

Harvey is trying to reduce that dependency. It recently released Harvey Tenet, a post-trained open-weight model built from Kimi K3, and is encouraging customers to adapt open models around their own professional knowledge. That gives Harvey another layer beyond simply wrapping somebody else’s frontier model.

Harvey is winning the Legal AI platform race today. The lead is real, but it is nowhere near permanent.

If you want more recent data on this point, please see our latest Legal Tech market report.

Are lawyers choosing ChatGPT instead of Legal AI tools?

For a lot of everyday work, yes. General-purpose AI is still more widely used than specialist Legal AI in some markets, which forces legal vendors to prove why lawyers should pay extra.

Clio found that 46% of legal professionals were using general-purpose tools such as ChatGPT, Claude, Gemini or Perplexity, compared with 40% using legal-specific AI. That specialist share had actually fallen from 58% in the previous survey while general-purpose usage increased from 36%.

Those numbers should make every Legal AI company uncomfortable. If lawyers only need help rewriting an email, summarising a generic document or brainstorming questions, ChatGPT or Claude may already be good enough.

Specialist products earn their place when the task becomes more serious. LexisNexis found that 81% of UK lawyers feel more comfortable using AI when it is grounded in legal sources, up from 70% a year earlier. Its APAC research found that only 46% use specialist Legal AI, either alone or alongside general tools, but 69% said guided AI workflows would make them more likely to use AI.

General AI handles flexible, low-friction work extremely well. Legal AI has to win on authoritative sources, matter context, permissions, auditability and workflows that a lawyer would otherwise perform manually.

Has Legal AI solved the hallucination problem?

No. Legal AI is much more usable than it was, but hallucinations remain common enough that lawyers still have to verify important output.

Independent Stanford research previously tested specialist legal-research products and found hallucination rates above 17% under its methodology. Those exact percentages should not be treated as current failure rates because the products and underlying models have improved substantially since the study.

Lawyers remain worried anyway. In LexisNexis’ latest UK research, 83% were concerned about inaccurate or fabricated information. Across APAC, 63% named accuracy and hallucination risk as their leading AI concern. Two different surveys, same operational problem.

The legal consequences make a low error rate harder to tolerate than they would be in many other professions. A fabricated authority in an internal brainstorm is annoying. A fabricated authority filed with a court can create sanctions, reputational damage and professional-conduct problems.

The American Bar Association has already made the lawyer’s responsibility clear through Formal Opinion 512. Lawyers using generative AI still have duties around competence, confidentiality and review of the work produced.

Legal AI has become good enough to use routinely before becoming good enough to trust without checking.

Is legal data becoming more valuable than the AI model?

In many Legal AI products, yes. These days the model is increasingly interchangeable, while trusted legal sources, internal knowledge and matter data are harder to reproduce.

Several strong models can already draft competent legal prose. A vendor that relies only on having access to a good model therefore has a weak moat.

The harder layer sits around it. Thomson Reuters can connect CoCounsel directly to Westlaw case law, KeyCite and Practical Law. LexisNexis can ground answers in its own legal corpus. Harvey and Legora can combine models with firm precedents, uploaded matter files, internal knowledge and custom workflows.

Lawyers themselves increasingly ask for this. LexisNexis found that comfort with legally grounded AI rose from 70% to 81% in a year.

The reason is practical. A beautifully written answer has little value if the authority does not exist, has been overturned, comes from the wrong jurisdiction or cannot be traced back to a source. The same goes for an AI that gives a reasonable answer while missing the one clause, email or precedent that changes the matter.

Harvey’s latest move toward post-trained open models makes this trend even clearer. If legal companies can swap or customise the underlying model, more of their long-term advantage has to come from the data, evaluation systems, workflows and institutional knowledge built around it.

If you want more recent data on this point, please see our latest Legal Tech market report.

Are junior lawyers actually losing work to Legal AI?

Yes. Legal AI is already taking over parts of junior lawyers’ workload, although the evidence still does not show a broad collapse in junior legal employment.

Research, first drafts, document review, chronology building, due diligence and contract comparison have traditionally consumed huge amounts of junior time. They are also exactly the jobs where AI works best today.

Recent employment data still looks strong. The National Association for Law Placement reported employment for recent US law graduates at roughly 93%, with about 83% moving into full-time, long-term lawyer jobs. Those results are far removed from a junior-lawyer employment crisis.

What worries firms is what happens next. Thomson Reuters’ latest research shows an expectation that junior positions will shrink more than mid-level and senior roles. At the same time, 78% of law-firm professionals believe early-career lawyers depend on experienced mentorship to develop skills that AI may displace. LexisNexis found an even sharper concern: 75% of UK lawyers believe juniors who rely heavily on AI may struggle to develop legal judgment.

Firms can remove several hours of tedious junior work and improve margins, but some of those hours were also how junior lawyers learned what a bad clause looks like, how cases connect and where apparently small details become dangerous.

The bigger problem for now is the training model, not mass unemployment.

Is Legal AI killing the billable hour?

Legal AI is putting real pressure on the billable hour now, especially for repeatable work where clients can see that the number of hours required has fallen.

Thomson Reuters found that 71% of in-house legal professionals expect outside firms to change their commercial model as AI usage increases. Only 28% of law firms said they had already changed pricing in response. Another measure in the same research found 62% of law-firm professionals saying their pricing structure had not changed at all.

Suppose AI cuts a ten-hour document-review task to two hours. An hourly firm has made the work much more efficient while removing eight hours it could previously bill. The client naturally expects some of that saving to flow through to the price.

Clio’s data suggests pricing had already started diversifying before the latest wave of agentic AI. It found that 59% of firms used flat fees for at least some matters while 41% billed exclusively by the hour.

Hourly billing will survive where the amount of work is genuinely difficult to predict, particularly complex litigation, investigations and bespoke advisory work. Repeatable legal production is much more exposed.

Pricing question	Reported result
In-house teams expecting firms to change how they charge	71%
Law firms already changing pricing in response to AI	28%
Law-firm professionals reporting unchanged pricing	62%
Firms using at least some flat fees in Clio data	59%

If you want more recent data on this point, please see our latest Legal Tech market report.

Are clients really choosing law firms based on AI now?

Yes. Some corporate clients are already reconsidering firms that cannot show AI-enabled value, and law firms appear to be underestimating that pressure.

Thomson Reuters found that 77% of clients consider AI-enabled quality improvements very important or essential. Only about 5% said they receive that level of improvement from most or all of their outside providers.

The commercial consequence is starting to show. Eleven percent of in-house legal professionals said they were already reconsidering relationships with firms that fail to demonstrate AI-enabled value. Another 22% expected to do so within the following twelve months.

Law firms see the situation very differently. Only 11% of law-firm professionals believed AI could begin costing their firm clients within twelve months, while half did not believe they would lose clients over it at all.

Corporate legal departments are themselves under pressure to show productivity gains: 61% told Thomson Reuters that internal stakeholders are pushing them to adopt AI faster. Once the in-house team is being asked to do more with AI, it becomes harder to accept an outside firm charging the same amount for work that should also have become faster.

Clients do not need to write “use Harvey” into an engagement letter to change the market. Asking why one firm takes three weeks when another takes four days is enough.

Is in-house Legal AI becoming a threat to outside law firms?

Yes. For routine legal work, one of the biggest threats to law firms currently comes from clients doing more of the work themselves.

AI changes the old outsourcing calculation. A corporate legal department can increasingly review contracts against a playbook, summarise large document sets, run preliminary research, monitor regulations and conduct first-pass diligence internally. Outside counsel then becomes necessary for the difficult exception rather than every step of the process.

This is particularly important because the main Legal AI platforms sell to both sides. Harvey serves law firms and corporate legal departments. Legora does the same. CoCounsel, LexisNexis and Luminance can also be deployed inside companies.

Thomson Reuters found that nearly one-third of in-house legal professionals already work in a model built around volume, consistency and competitive cost for repeatable work. As seen above, those teams are also putting pressure on outside firms to pass AI efficiencies into their pricing.

High-stakes litigation, unusual transactions, difficult negotiations and specialist advice remain much harder to bring in-house.

Routine production is more exposed because Legal AI lets a client internalise work without hiring enough lawyers to reproduce a large outside team.

If you want more recent data on this point, please see our latest Legal Tech market report.

Are contracts becoming the easiest legal work for AI to automate?

Yes. Contract work is currently one of the strongest candidates for deep Legal AI automation because the documents, rules and acceptable outcomes can often be defined in advance.

Luminance provides one example of how far the workflow has expanded. The company moved beyond contract analysis into generation, negotiation and post-signature management, and says more than 1,000 organizations across 70 countries use its platform.

Broader Legal AI platforms are heading in the same direction. A contract workflow can pull the company’s playbook, identify clauses, compare wording with approved positions, suggest changes, flag unusual terms and route only the difficult exceptions to a lawyer. Harvey and Legora increasingly support this type of multi-step document work alongside their broader research and drafting products.

Contracts fit AI particularly well because companies encounter the same legal concepts thousands of times. Governing law, indemnification, liability caps, assignment, termination and data-processing terms vary from agreement to agreement, but they are often judged against explicit internal preferences.

That gives Legal AI clearer boundaries than many open-ended legal questions. The company knows which documents matter, which rules to apply and which deviations should trigger human review.

The economics also compound quickly. Saving twenty minutes on one agreement is trivial. Saving twenty minutes across 30,000 agreements becomes 10,000 hours.

Will AI replace Westlaw and Lexis?

Probably not. AI is rapidly replacing parts of the way lawyers use Westlaw and Lexis, while the underlying legal databases remain extremely valuable.

Traditional legal research made the lawyer perform almost every intermediate step: construct searches, inspect cases, follow authorities, check treatment and assemble the conclusion.

AI can increasingly hide much of that process. A lawyer asks the legal question in ordinary language, and a research agent searches, retrieves, compares and synthesises authorities before presenting an answer.

That sounds dangerous for Thomson Reuters and LexisNexis until we look at what the agent needs underneath. Reliable research still depends on comprehensive legal material, accurate metadata, citators and systems that know whether a case remains good law.

Thomson Reuters has therefore put AI directly on top of Westlaw and Practical Law instead of defending the old search interface. LexisNexis is doing much the same with Lexis+ AI and Protégé.

The risk for both companies comes from AI-native platforms assembling strong enough alternatives around other data sources. Harvey, general AI companies and newer research providers would love to make the traditional legal database less central.

For now, though, lawyers seem to want more authoritative grounding as AI use increases, not less.

Is confidential client data becoming a bigger Legal AI problem?

Yes. Legal AI is becoming more useful by accessing more sensitive information, so confidentiality has turned into a much bigger product and governance problem.

A research chatbot can work with a relatively narrow prompt. A useful matter-level system may need contracts, emails, litigation documents, corporate records, privileged analysis, previous firm work and internal negotiating positions.

That creates much more complicated questions around permissions, retention, model training, data residency, audit trails and information barriers between clients.

The problem becomes harder when lawyers bypass approved systems. Thomson Reuters found that 34% of law-firm professionals use AI tools their organization has not authorized. Firms can spend heavily on secure enterprise systems while still having confidential information copied into products they have never reviewed.

The American Bar Association has explicitly connected generative AI use with lawyers’ existing confidentiality and competence duties. Firms therefore need to understand where client information goes and what the provider can do with it rather than relying on a generic promise that a tool is “secure.”

Legal AI vendors know this. Enterprise deployments increasingly compete on security architecture, permission controls and traceability alongside model performance.

Could Legal AI actually create more legal work?

Yes. Cheaper legal analysis can expand the amount of legal work people are willing to do, so fewer lawyer-hours per task does not automatically mean a smaller legal market.

There is plenty of unmet demand. Clio has cited research suggesting that roughly 77% of legal problems receive no help from a legal professional. Cost, time and inconvenience explain a large part of that gap.

AI lowers those barriers.

A small company that would never pay outside counsel to review every customer contract can run every agreement through an AI workflow and escalate unusual clauses. A corporate legal department can monitor regulations in twenty countries instead of five. Consumers can get an initial assessment of problems that previously felt too small to justify paying for a lawyer.

The same effect can happen inside firms. Clio found that growing firms use time-saving automation about twice as much as stable firms and almost three times as much as shrinking firms. More capacity can mean handling more matters rather than simply reducing staff.

This also changes what lawyers themselves are worth. Routine production becomes cheaper, while judgment, negotiation, strategy and responsibility make up a larger share of the human job. Google’s general counsel made that distinction explicitly when discussing Gemini Enterprise for Legal: AI can take on much more legal work, but human judgment remains essential.

What is holding Legal AI back now?

Legal AI is currently being held back more by implementation than by a lack of capable models.

Thomson Reuters found a striking gap between organizations with a clear AI strategy and those without one. Among professionals working where a clear strategy exists, 64% said AI met or exceeded expectations. Where there was no clear strategy, that figure fell to 29%.

Other data points describe the same problem from different angles. Only 18% of law firms told Thomson Reuters that they measure AI return on investment. In LexisNexis’ APAC research, just 53% said their organization assesses the value created by AI. Meanwhile, 69% said they would be more likely to use AI if it came through guided workflows.

The hard work has moved inside the organization. Firms need to decide which tasks AI should handle, which sources it can use, what quality threshold is acceptable, when a lawyer has to intervene and how the resulting efficiency should change staffing or pricing.

That explains why two firms can buy similar Legal AI tools and get completely different results. One drops the software into the existing workflow and hopes people use it. The other redesigns the workflow around what the software can reliably do.

Is Legal AI changing law firms faster than law firms are changing themselves?

Yes. Legal AI capability and adoption are moving noticeably faster than law-firm pricing, training and operating models.

The mismatch shows up in several places at once.

More than three-quarters of legal professionals use AI in Clio’s data, while recent LexisNexis surveys put adoption above 90% in the UK and APAC. Agentic products can already carry out multi-step research and document workflows. Harvey has passed $400 million in ARR and reached a $15.5 billion valuation. Corporate clients increasingly expect their advisers to show clear AI-enabled improvements.

The business model has moved far less. As pointed out above, 62% of law-firm professionals told Thomson Reuters their pricing had not changed because of AI, despite 71% of in-house lawyers expecting firms to change how they charge. Firms are also still trying to work out how juniors develop judgment when AI performs more of their traditional training work.

Even communication with clients remains surprisingly weak. Thomson Reuters’ Stand-out Lawyers Survey found that more than three-quarters of highly regarded lawyers said their firm had an AI strategy, yet fewer than half felt confident about their practice area succeeding as AI becomes more embedded. Among partners who were already heavy AI users, only about one-third had discussed AI with at least 60% of their clients.

Law firms have spent decades optimizing around hours, leverage and junior labour. AI is making all three variables less stable at the same time.

If you want more recent data on this point, please see our latest Legal Tech market report.

What’s actually changing in Legal AI now?

Legal AI is currently moving from a tool lawyers occasionally consult to a system that can perform meaningful parts of a legal matter before a lawyer steps in.

The change is already visible across adoption, product design and money. AI use has become normal in several major legal markets. Harvey has crossed $400 million in ARR and reached a $15.5 billion valuation. Legora has grown to roughly $150 million in reported ARR. Thomson Reuters and LexisNexis are rebuilding legal research around AI. Google has entered the market directly. Contract platforms are moving deeper into automated negotiation and review. Corporate legal departments can perform more work without sending every step to outside counsel.

At the same time, lawyers still worry heavily about hallucinations, confidentiality and weaker training for juniors. Those concerns are not slowing adoption enough to reverse the trend. They are shaping where humans remain necessary.

AI is increasingly strong at finding information, comparing documents, applying defined rules, producing first drafts and carrying a workflow through several intermediate steps. Lawyers still have a much stronger position when the work depends on judgment, negotiation, accountability, ambiguous facts or deciding how much risk a client should accept.

For years, the Legal AI debate centered on whether machines could write credible legal text. Then the question became whether lawyers would actually use them. Those arguments are largely settled.

What matters now is how much paid human work remains between the beginning of a legal problem and the moment when professional judgment becomes indispensable.

That distance is getting shorter.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.