What problem does it solve?
A large part of a lawyer's time goes into work that is necessary but repetitive: finding the relevant authorities, reading them, extracting what matters from long documents and turning it into a first draft of a memo, a briefing note or a filing. Much of it falls to junior lawyers, clients question paying for it by the hour, and in public bodies the same work competes with heavy caseloads and vacancies.
General purpose chatbots look like an answer but are dangerous here. They produce fluent text with invented case citations, and courts have started to take action against lawyers who filed them. The High Court of England and Wales stated in 2025 that freely available generative AI tools trained on a large language model are not capable of conducting reliable legal research. The useful version is an assistant grounded in authoritative legal sources and the organization's own precedents, that shows where every statement comes from and leaves the judgment to a lawyer.
How does it work?
- Ask in plain language. The lawyer asks a question or describes the task ("summarise the limitation rules for this claim in these three jurisdictions", "draft a first briefing on this article of association").
- Retrieve from trusted sources. The assistant searches licensed legal databases, the organization's precedent bank and the documents of the matter, not the open web by default.
- Answer with citations. It answers or drafts with a citation for each proposition, linked to the passage it relied on, and says when it found nothing.
- Extract and compare. For document heavy tasks it pulls defined points from many documents into a table, for example clauses, dates or obligations, with references back to the source.
- Verify and finish. The lawyer checks every citation and conclusion, edits the draft and records the result in the matter file.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools, Microsoft Teams
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | about 40,000 | 1 | 1 organization |
| Productivity gain | Too few to pool | about 45% | 1 | 1 organization |
| Users served | Not pooled | about 3500 | 1 | 1 organization |
Value drivers: Employee productivity, Speed and cycle time, Lower cost to serve, Risk and loss reduction.
Indicative value
A law firm or legal department with 100 lawyers
USD 450,000 to USD 3 million
Lawyer time released per year
How this is calculated
Formula: lawyers * hoursPerLawyer * timeSaved * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Lawyers using the assistant lawyers, lawyers | 100 | 100 | The reference organization. |
| Hours per lawyer per year on research, extraction and first drafts hoursPerLawyer, hours per lawyer per year | 300 | 500 | Editorial assumption, replace with your own time recording data. |
| Share of those hours saved timeSaved, fraction of hours | 0.15 | 0.3 | Conservative against Ashurst's controlled experiments on this page (about 45% time saved on first draft briefings), because verification of citations takes time and not every task suits the tool. |
| Internal cost of a lawyer hour hourlyCost, USD per hour | 100 | 200 | Editorial assumption for a blended internal cost, not the billing rate. |
What it leaves out: Values released lawyer time at internal cost. It leaves out licence and legal database costs, the time spent verifying outputs, any change in billable revenue under hourly billing, and the cost of an error that verification misses.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
U.S. Securities and Exchange Commission
United States · Government and public sector · 2025
The SEC's Division of Enforcement, Division of Examinations, Office of the General Counsel and Office of the Chief Data Officer are piloting Westlaw CoCounsel, a generative AI assistant, since October 2025 for legal research, document analysis and document production; its outputs are document summaries, draft documents and answers to user selected questions. Separately, the agency has used AI features in Lexis+ and Westlaw Precision for legal research since January 2020, and Enforcement uses Westlaw Quick Check to cross check SEC documents against case law. No outcome figures are published.
No outcome disclosed.
Ashurst Perkins Coie
United Kingdom · Professional services · 2024
Ashurst, which has since combined with Perkins Coie and now operates as Ashurst Perkins Coie, ran three global generative AI trials from November 2023 to March 2024 with 411 partners, lawyers and staff in 23 offices and 14 countries, and published the results in its report Vox PopulAI. In controlled experiments with small groups within those trials, it measured approximate time savings of 45% on first draft legal briefings, 59% on sector research reports and 80% on UK corporate filings that required extracting information from articles of association. In a blind study, an expert panel correctly identified all lawyer written outputs, but half of the AI generated outputs were either mistaken for human work or could not be classified. The trials used only publicly available data, not client data.
- Productivity gain: about 45%, controlled experiments, November 2023 to March 2024
"In our controlled experiments, we measured approximate time savings of 80% to draft UK corporate filings requiring review and extraction of information from company articles of association, 59% to draft industry/sector-specific research reports that require reviewing and extracting key information from public company filings (Form 10-Ks), and 45% on creating first draft legal briefings."
Claimed by: organization
U.S. Department of Justice
United States · Government and public sector · 2024
The Department of Justice lists AI assisted legal research in Westlaw and LexisNexis as deployed department wide since 2019 and 2020. These features recommend case law, statutes, regulations and scholarly articles and in some circumstances give an overview of the law in response to a query; the inventory classifies them as NLP and classical machine learning, so they are AI assisted search and recommendation rather than a generative drafting assistant. The generative tool on record is a CoCounsel pilot at the Executive Office for United States Attorneys since March 2024, meant for document review and other tasks as a comparison point to other commercial alternatives. Its listed outputs are automated transcriptions, metadata tagging suggestions and object and facial recognition in media files, not legal research or drafting. No outcome figures are published.
No outcome disclosed.
A&O Shearman
United Kingdom · Professional services · 2023
Allen & Overy, which merged with Shearman & Sterling in 2024 to form A&O Shearman, trialled Harvey, a generative AI platform built on OpenAI models and adapted for legal work, from November 2022 and in February 2023 integrated it into its global practice for more than 3,500 lawyers in 43 offices. The release describes Harvey as supporting work such as contract analysis, due diligence, litigation and regulatory compliance, and a quoted A&O executive says it can work in multiple languages; it does not say which of these tasks A&O lawyers use it for. The firm states that the output needs careful review by an A&O lawyer. The published figures describe usage in the trial, not time saved.
- Users served: about 3500, beta trial, November 2022 to February 2023
"At the end of the trial, around 3500 of A&O’s lawyers had asked Harvey around 40,000 queries for their day-to-day client work."
Claimed by: organization - Interactions handled: about 40,000, beta trial, November 2022 to February 2023
"At the end of the trial, around 3500 of A&O’s lawyers had asked Harvey around 40,000 queries for their day-to-day client work."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Licensed access to the legal databases lawyers already rely on
- A curated precedent and know how bank with owners and review dates
- Matter documents with access rights that the assistant respects
- A written policy on permitted uses, client consent and verification
Systems to integrate
- Legal research databases
- Document management system and precedent bank
- Matter management and time recording
- Identity and access management for matter level permissions
Complexity: Medium
Buying a legal AI product is easy; making it trustworthy is not. The work is in connecting authoritative sources and the organization's own know how, keeping client data confidential, training lawyers to verify, and defining which tasks the tool may be used for.
- 1
Start with a structured trial
Choose a few task types, such as first draft briefings, research summaries and document extraction, and measure time and quality against a control group. Ashurst ran three trials with 411 people and measured time savings in controlled experiments, and A&O ran a beta from November 2022 before its firm wide rollout in February 2023.
- 2
Ground it in trusted sources
Connect licensed legal content and the organization's precedents, and turn off open web answers for legal questions. Require a citation for every proposition.
- 3
Make verification part of the workflow
Treat every output as a trainee's draft. Build citation checking into the process and make the reviewing lawyer's sign off visible in the matter file.
- 4
Protect confidentiality
Use a deployment where client data is not used to train models, respects matter access rights and stays in the required region. Record client consent where engagement terms require it.
- 5
Train, then widen
Teach lawyers what the tool is good and bad at, share good prompts per practice area and track adoption and quality before opening it to more teams.
Guardrails
- Citation for every legal proposition, linked to the source passage
- Refusal when no authority is found, instead of a plausible guess
- Matter level access control so the assistant never mixes clients
- No client data used for model training, with data kept in the required region
- Lawyer verification and sign off before anything leaves the organization or reaches a court
KPIs to instrument
- Time to first draft for defined task types, against a control group
- Share of citations that fail verification on a sample
- Weekly active lawyers as a share of licensed lawyers
- Reviewer edits per draft
- Client or court complaints linked to AI assisted work
Human in the loop
A lawyer owns every output: they check each citation against the source, decide the legal position and sign the document. The assistant never files, sends or advises on its own.
Common failure modes
- Invented or misquoted authorities
- The model cites cases that do not exist or do not say what it claims. Courts in the US and the UK have taken action against lawyers for this. Ground answers in authoritative sources and verify every citation.
- Automation bias
- Fluent drafts are accepted without enough scrutiny, especially under time pressure. Set a fixed verification checklist per task type and sample reviewed work, instead of relying on reviewers to notice when something reads wrong.
- Confidentiality breach
- Client documents are pasted into a consumer tool. Provide an approved tool and block the rest.
- Adoption without measurement
- Licences are bought but usage and quality are not tracked. Measure time saved and error rates per task type and cut the tasks where the tool does not help.
What are the risks and rules?
EU AI Act
Depends on design
Research and drafting support for lawyers in firms and companies is not listed in Annex III, so it is normally minimal risk with AI literacy duties. Annex III point 8(a) makes it high risk when a judicial authority, or someone on its behalf, uses AI to research and interpret facts and the law and to apply the law to a concrete set of facts, or when it is used in a similar way in alternative dispute resolution, so a deployment for courts, tribunals or arbitration needs its own classification.
Guidance
- Risk Outlook report: The use of artificial intelligence in the legal market (Solicitors Regulation Authority, Europe). Explains hallucination and says that however well controlled a system is, firms still need to check its outputs for accuracy.
- Generative AI, the essentials (The Law Society of England and Wales, Europe). Practical guidance for solicitors on risks such as hallucination, confidentiality and professional duties when using generative AI.
- Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin) (High Court of England and Wales (Divisional Court), Europe). Sets out lawyers' duties when using AI for research and warns of sanctions, including referral to regulators, for citing fictitious authorities.
Controls to put in place
- Approved tool list and a written acceptable use policy for generative AI
- Citation verification step recorded in the matter file
- Data protection impact assessment and supplier due diligence on data use and location
- Training for lawyers before access, with refreshers when tools change
- Sampling of AI assisted work by a senior lawyer
When it went wrong elsewhere
- ChatGPT Reportedly Produced False Court Case Law Presented by Legal Counsel in Court. AI Incident Database entry 541. In Mata v. Avianca, lawyers filed a brief citing cases invented by ChatGPT, and a US federal judge ordered them to show cause why they should not be sanctioned.
- Ayinde and Al-Haroun, fictitious citations in court documents. The Divisional Court dealt with two cases in which documents placed before the court cited authorities that did not exist, and referred lawyers to their regulators.
Frequently asked questions
- How much time does a legal AI assistant save?
- In Ashurst's controlled experiments, the firm measured approximate time savings of 45% on first draft legal briefings, 59% on sector research reports and 80% on UK corporate filings that required extracting information from articles of association. Those were trial conditions with small groups; real savings depend on how much verification the task needs.
- Can lawyers rely on ChatGPT for legal research?
- Not on its own. The High Court of England and Wales said in 2025 that freely available generative AI tools are not capable of conducting reliable legal research, and courts in the UK and the US have taken action against lawyers over fictitious citations. Use a tool grounded in authoritative sources and verify every citation.
- Who uses it at scale?
- Allen & Overy rolled out Harvey to more than 3,500 lawyers in 43 offices in 2023 after a trial in which lawyers asked around 40,000 queries. In the US federal government, the Department of Justice and the SEC use AI features in legal research services, and the SEC is piloting a generative assistant.
- Is a legal research assistant high risk under the EU AI Act?
- For law firms and legal departments, normally not. It becomes high risk when a judicial authority uses it to research and interpret facts and law and apply the law to a case, or when it is used in a similar way in alternative dispute resolution (Annex III point 8(a)).
How to cite this page
Blits.ai AI Use Case Library, "AI legal research and drafting assistant for lawyers", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/legal-research-and-drafting-assistant. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published