• Welcome to Digital Expressway Consulting
  • +91 98180 30135
  • info@digitalExpressway.net

Legal AI Cut First-Draft Time by 45% in One Firm’s Trials: How Law Firms and Legal Teams Get the Gains Without the Fake Cases

← Industry Foresight & Innovation Hotbeds · Professional services

Legal AI Cut First-Draft Time by 45% in One Firm’s Trials: How Law Firms and Legal Teams Get the Gains Without the Fake Cases

In trials run by the global law firm Ashurst with more than 400 people, generative AI cut the time to produce a first-draft legal briefing by 45% on average, and cut time on certain corporate filings by 80%, according to Ashurst. Allen & Overy’s lawyers put about 40,000 queries to the legal AI platform Harvey during a beta trial before the firm rolled it out to more than 3,500 lawyers in 43 offices, the Law Gazette reported.

Now the other side. In June 2023, a New York federal judge fined two lawyers and their firm $5,000 for filing a brief built on court decisions that ChatGPT had invented, according to the court’s sanctions order in Mata v. Avianca. Two years later, the High Court in London dealt with a case in which a solicitor relied on research from his client that apparently included 18 fictitious cases, Legal Cheek reported.

The technology is broadly the same in both stories. What differs is how it was used, checked and governed. If you run a legal team or buy legal services, that difference is where both the value and the risk sit.

What you will take away from this article

  • What legal AI does well today: contract review, due diligence, research summaries and first drafts, with sourced numbers.
  • Why legal AI goes wrong: invented citations, overtrust in specialist tools, confidentiality slips and missing firm policies.
  • What the professional rules now say: the American Bar Association, English courts and regulators, and Indian courts.
  • How AI in audit compares: what the UK and US audit regulators have found.
  • A 30-day plan: a practical checklist and the habits a trained team uses instead of shortcuts.

About a 12-minute read.

First, a quick definition. “Legal AI” here means software built on large language models (systems trained on huge amounts of text to predict and generate language) that is used for legal work. Some tools are general chat assistants. Others are specialist products that search a legal database first and then write an answer from what they find, a design called retrieval-augmented generation. Most AI for lawyers falls into one of these two groups. Either way, the tool produces fluent text, not verified legal advice. A “hallucination” is when the tool states something false with confidence, such as a case that does not exist.

Four Moves the Leaders Make

  • Run a measured trial before a firm-wide rollout. Ashurst ran trials of three generative AI platforms with more than 400 people across 23 offices in 14 countries, Ashurst says. It measured time saved on specific tasks, such as about 2.5 hours saved per first-draft briefing. It also had an expert panel judge outputs blind, the firm reports. What to borrow: pick named tasks, time them before and after, and have experts judge quality without knowing which drafts came from AI.

  • Compare platforms on your own work before choosing one. Cyril Amarchand Mangaldas, an Indian law firm, tested three generative AI platforms with 380 lawyers before selecting Legora as its firm-wide platform, Bar & Bench reported in July 2025. The firm judged them on usability, accuracy, efficiency, training and how well they fitted existing systems, according to the same report. What to borrow: write your evaluation criteria down before the pilot starts, and include accuracy and training, not just speed.

  • Make human review part of the rollout message. When Allen & Overy launched Harvey for work including contract analysis, due diligence, litigation and regulatory compliance, the firm said the output needs “careful review” by a lawyer, the Law Gazette reported. What to borrow: tell every user, from day one, that the tool drafts and the lawyer decides.

  • Fence off the riskiest tasks. When England’s Solicitors Regulation Authority approved Garfield.Law, an AI-driven firm that helps businesses recover unpaid debts through the small claims process, it set conditions, the regulator says. The system may not propose case law, because that is a high-risk area for language models. It takes a step only after the client approves it, and named solicitors stay responsible for every output. What to borrow: list the tasks your tool must never do, and build approval steps into the workflow.

Why Most Legal AI Efforts Stall

1. Invented citations reach the court. The most visible failure is a filing that cites cases that do not exist. In Mata v. Avianca, the judge wrote that “there is nothing inherently improper about using a reliable artificial intelligence tool for assistance”, but that existing rules give lawyers “a gatekeeping role” over the accuracy of their filings, according to the sanctions order. The problem has grown since. A Stanford Center for Internet and Society analysis of a public database of AI hallucination cases, run by the researcher Damien Charlotin, counted 114 US cases involving lawyers between June 2023 and October 2025, plus 160 involving people representing themselves, published in October 2025. About 90% of the firms involved were solo practices or firms of up to 25 lawyers, the analysis found.

2. Specialist tools are trusted more than they deserve. It is tempting to assume a legal research product cannot make things up. Stanford researchers tested leading tools on 202 legal questions and published the results in the Journal of Empirical Legal Studies in 2025, according to the paper. Lexis+ AI and Ask Practical Law AI gave misleading or false answers to more than 1 in 6 queries (about 17%), and about one-third of Westlaw AI-Assisted Research answers contained a hallucination. These tools made fewer errors than a general model, GPT-4, Stanford HAI reported, but the authors concluded that claims of “hallucination-free” legal AI remain “at best, ungrounded”. Notably, the Stanford analysis above found Westlaw and Lexis named among the tools behind some faulty filings.

3. No policy, so everyone improvises. In the 2025 Clio Legal Trends Report, 79% of legal professionals said they use AI, yet 53% said their firm had no AI policy or that they did not know of one, according to a summary of the report. The surveys covered more than 1,700 US legal professionals, and Clio, a software supplier, is not a neutral party. Without a policy, nobody knows which tools or data are allowed.

4. Confidentiality and privilege are treated as afterthoughts. Client information typed into a public chatbot may be stored or used by the provider. The Kerala High Court’s policy for district courts, reported in July 2025, warns judges and staff against putting case facts, personal identifiers or privileged communications into cloud-based tools such as ChatGPT and DeepSeek, MediaNama reported. The same logic applies to any law firm or in-house team.

5. Culture lags behind use. In a survey of UK lawyers published by LexisNexis in September 2025, 61% said they use AI in their daily work, but only 17% said it was fully embedded in their organisation’s strategy and operations, according to LexisNexis. LexisNexis sells legal AI tools and did not state the sample size on that page.

Case in point: two English cases, one warning to the whole profession

What happened: In June 2025, the High Court in London ruled on two cases heard together, Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank. In one, a pupil barrister (a trainee) cited five authorities that did not appear to exist. In the other, a solicitor relied on research supplied by his client that apparently included 18 fictitious cases, Legal Cheek reported.

What went wrong: The citations reached the court without being verified. The court did not find the solicitor knew the cases were false, and neither lawyer faced contempt proceedings. But it said relying on his client was no excuse, because “it is the lawyer’s professional responsibility to ensure the accuracy”, according to the same report.

What changed: The court found existing guidance for lawyers “insufficient to address the misuse” of AI. It said that in future hearings it may ask whether managing partners and heads of chambers have met their leadership responsibilities, according to the same report.

The lesson: Checking AI output is now a leadership duty, not just an individual one.

What the Successful Few Do Differently

What they do What it means in practice The evidence
Set a visible AI strategy Leaders name the goals, approved tools and rules, and tell staff In Thomson Reuters’ 2025 survey of 2,275 professionals, organisations with a visible AI strategy were twice as likely to report AI-linked revenue growth; only 22% had one (Thomson Reuters)
Adopt AI widely, not in pockets Tools are part of normal matters, not a side experiment Among professionals who had widely adopted AI, 69% said it helped revenue, against 36% overall (Clio, via 2Civility)
Verify every citation A person opens each cited case in a trusted database before filing Leading legal research tools still hallucinated on 17% to 33% of test questions (Stanford, Journal of Empirical Legal Studies)
Keep named humans accountable A specific lawyer signs off every AI-assisted output A condition of the regulator’s approval of Garfield.Law (SRA)

These surveys show correlation, not causation. Firms that grow may simply be better run in many ways, including how they adopt AI.

Where Legal AI Already Works

Extracting information from long documents (company-reported trial). Reviewing company articles of association to pull out information for corporate filings took 80% less time with generative AI, Ashurst reports. This is classic due diligence work: repetitive and easy to check.

First drafts (company-reported trial). First-draft legal briefings took 45% less time on average, Ashurst says. In a blind review, experts correctly identified only half of the AI drafts as AI-made, the firm reports. That says something about style, not about legal accuracy, which still needs a lawyer.

India view. Indian firms are moving from pilots to firm-wide platforms. Bar & Bench reported in July 2025 that Cyril Amarchand Mangaldas chose Legora after a 380-lawyer pilot, that Shardul Amarchand Mangaldas had announced a partnership with Harvey the month before, and that firms including Trilegal, IndusLaw and S&R Associates had adopted other legal AI tools (Bar & Bench). In the courts, Kerala’s July 2025 policy bars district judges from using AI to decide any order or judgment, and requires AI-generated citations to be verified by judicial officers (MediaNama).

How It Works

A legal AI workflow has five parts.

  1. Input. A lawyer supplies documents, such as a contract or a data room (the set of files shared in a deal), plus a clear instruction.
  2. Retrieval. Specialist tools search a trusted source, such as a case-law database or the firm’s own precedents, for relevant material.
  3. Generation. The language model writes a summary, a clause comparison or a first draft based on what it found.
  4. Verification. A lawyer checks the output against the source documents and opens every cited authority.
  5. Approval and record. A named lawyer signs off, and the firm logs which tool was used and how it was checked.

Illustrative workflow (a composite, not a real organisation). An in-house legal team at a mid-sized manufacturer receives 60 supplier contracts in an acquisition. A paralegal uploads them to an approved tool inside the company’s secure environment. The tool flags change-of-control clauses, which let a supplier end the contract if the company changes owner, and lists the clause and page for each. A senior lawyer checks every flagged clause against the original and samples ten unflagged contracts for misses. Only then does a risk summary go to the deal lead.

[DIAGRAM: Contracts uploaded to approved tool → Tool flags key clauses with page references → Lawyer checks flags against originals → Lawyer samples unflagged contracts → HUMAN APPROVAL: senior lawyer signs off → Risk summary sent to deal lead → Tool use logged]

Value It Unlocks

The measured gains cluster in reading, extracting and first drafting, as Ashurst’s company-reported figures show. Professionals in Thomson Reuters’ 2025 survey expected AI to free up about five hours a week within a year, worth about $19,000 per professional each year, Thomson Reuters says. Thomson Reuters sells legal AI tools, and these are expectations, not measurements.

The gains are uneven. In the Clio report, 87% of legal professionals in large firms used AI against 71% at solo firms, and a third of respondents said it was too early to judge the effect on revenue, according to 2Civility. Time saved also creates a pricing question. The American Bar Association’s Formal Opinion 512 says lawyers may bill for time spent entering information and reviewing AI output, but generally not for learning the tool, according to the ABA. Opinion 512, issued in July 2024, also ties AI use to existing duties of competence, confidentiality and client communication.

AI in audit. Audit shows a similar pattern of fast adoption and thin measurement. In June 2025 the UK’s Financial Reporting Council (FRC), the audit regulator, reviewed the six largest audit firms and found that, apart from one, none had set measures of how their automated tools affect audit quality, Accountancy Age reported. The tools included KPMG’s transaction scoring, which flags unusual entries, and Deloitte’s use of AI to review board minutes and contracts, according to the same report. The FRC published its first guidance on AI in audit at the same time, the regulator says. In the US, staff at the Public Company Accounting Oversight Board (PCAOB), the audit regulator, found in July 2024 that firms’ generative AI use was mostly administrative and research work, with a stated need for strong supervision, according to the PCAOB.

Where Claude Fits, and Where It Doesn’t

Claude, the AI assistant made by Anthropic, is a general-purpose tool rather than a legal database. It fits work where the source material is in front of it: summarising a long contract you provide, comparing two versions of a clause, turning meeting notes into a first-draft memo, or producing a checklist from a policy document.

A person must stay in charge in four places. First, any legal authority: a general assistant is not a case-law database, so every case, statute and quotation must be checked in a trusted source. Second, client confidentiality: only approved, contracted environments should receive client data. Third, legal judgement and advice: deciding what a clause means for this client is the lawyer’s job. Fourth, anything filed with a court or regulator, which needs a named person’s sign-off.

The principle: let AI do the reading and first drafting; keep the judgement, the citations and the signature human.

How to Start: A 30-Day Checklist

  • [ ] Week 1: Write a one-page AI policy. Name approved tools, banned uses and what client data may be used.
  • [ ] Week 1: Pick two low-risk, high-volume tasks. Good candidates are clause extraction and first-draft summaries.
  • [ ] Week 2: Time the tasks as they are done today. Record hours and error rates as a baseline.
  • [ ] Week 2: Set a citation rule. Every authority is opened in a trusted database before anything leaves the team.
  • [ ] Week 2: Check confidentiality terms. Confirm with the supplier where data is stored and whether it is used for training.
  • [ ] Week 3: Run the pilot with a small group. Time the same tasks with AI and log every error found in review.
  • [ ] Week 3: Run a blind quality check. Have a senior lawyer rate AI-assisted and manual drafts without knowing which is which.
  • [ ] Week 4: Review results and decide. Expand, adjust or stop, based on time saved and errors caught.
  • [ ] Week 4: Brief leadership and clients. Explain how AI is used, how it is checked, and how billing reflects it.

Why Trained Teams Succeed

An “anti-pattern” is a habit that looks efficient in the moment but causes problems later. In legal AI, most anti-patterns come from treating fluent text as checked text.

Short-sighted habit What a trained team does instead
Pasting a citation from the tool straight into a brief Opens every case in a trusted database and reads the relevant passage
Assuming a specialist legal tool cannot hallucinate Checks specialist-tool output with the same care as a general chatbot
Pasting client documents into a free public chatbot Uses only approved tools with agreed confidentiality terms
Letting a junior lawyer use AI unsupervised on court work Has a named senior lawyer review and sign off
Billing as if AI had not been used Agrees transparent pricing that keeps fees reasonable, in line with ABA Opinion 512

Training makes a measurable difference to adoption. In BCG’s AI at Work 2025 survey of 10,635 workers, 79% of people with more than five hours of training used AI regularly, against 67% of those with less, according to a summary by UNLEASH. Yet in BCG’s 2026 edition, only 36% of workers felt properly trained, according to BCG. In legal work, where one invented citation can lead to a court sanction, that training gap matters more than in most fields.

Build the skills to use legal AI safely and productively. The CCAO-F prep course from Digital Expressway Academy prepares professionals who run AI work without writing code (consultants, managers, analysts, operations and delivery leads) for Anthropic’s Claude Certified Associate exam, including how to choose appropriate uses, keep a human in charge and check outputs before they go out. See the course

Bringing AI to your whole team? Talk to us about our Industry Advisory, Innovation and Pilot Build, and Mentoring and Skill Development services.

Sources

All sources were opened and checked in October 2026. Figures described as company-reported come from the organisation itself.

Digital Expressway Insights. Published November 2026; last reviewed October 2026. Not affiliated with or endorsed by Anthropic. Claude is a product of Anthropic. Examples marked “illustrative” are composites created to explain an idea and do not describe a real organisation.