AI in Legal Practice: What It Can and Cannot Do
Deliberate Academy Editorial Team
Reviewed for accuracy and professional relevance
- Map the categories of legal work where AI is demonstrably useful against those where it is unreliable or inappropriate
- Explain why professional duties including competence and confidentiality are non-delegable to AI tools regardless of output quality
- Apply a two-axis risk matrix to categorize legal tasks by consequence of AI error and degree of judgment required
- Evaluate a legal AI tool by probing its error behavior in your jurisdiction and practice area rather than relying on vendor demonstrations
Legal professionals need a more precise and more critical understanding of AI than most professions. In marketing, an AI error costs you a poor campaign. In accounting, it costs you a restatement. In law, it can cost you a professional conduct finding, a client's case, and your practicing certificate. The starting point for responsible AI use in legal practice is a clear-eyed assessment of what these tools actually do, where they are genuinely useful, and where they are reliably unreliable.
The Current State of AI in Law Firms and In-House Teams
AI adoption in legal practice is accelerating across firm sizes and practice types. Large commercial firms are deploying AI-native platforms — Harvey, Luminance, Kira Systems, CoCounsel — for document review, due diligence, and first-draft contract generation. In-house legal teams are using general-purpose AI tools for research support, policy drafting, and contract summarization. Legal operations functions are using AI to manage high-volume contract workflows at a scale that was previously impossible with legal headcount alone.
The adoption curve has been steeper than in many professional services sectors, partly because law has always been document-intensive, and AI is particularly good at processing documents at speed. The risk curve has also been steeper — because the professional standards that govern legal practice do not bend to accommodate tool failure.
AI-Assisted Due Diligence — Mid-Size Commercial Law Firm
Context
A mid-size commercial law firm was routinely involved in M&A transactions requiring review of 1,500 to 2,500 contracts per deal. Junior associates were spending six to eight weeks on initial document review for a typical transaction, at significant cost to the client and under constant deadline pressure from the deal timeline.
Action
The firm deployed Luminance for AI-assisted contract review on three consecutive transactions over a six-month pilot. Junior associates used the AI to flag high-risk clauses, non-standard terms, and missing provisions across the full document set. Each flagged item was then reviewed by a qualified lawyer before being included in the due diligence report. A structured sign-off protocol required a supervising solicitor to countersign any section of the report that relied on AI-flagged analysis.
Outcome
Initial document review time fell from six weeks to nine working days across all three transactions. Client-facing fee estimates for document review dropped by 35%. The structured review protocol identified that the AI missed approximately 4% of high-risk clauses — primarily those expressed in non-standard drafting that the model had limited exposure to in its training corpus. The legal operations manager noted that without the review protocol, those misses would have gone undetected. The protocol was retained as a non-negotiable element of the firm's AI use policy.
The categories of legal work where AI is demonstrably useful include: volume document review for patterns, defined clause types, or red flag identification; legal research as a first-pass orientation tool for unfamiliar areas; contract and document drafting as a starting framework; due diligence checklist compilation and initial execution; summarisation of long documents; and translation of complex legal text into plain English for client communication.
The categories where AI is unreliable or inappropriate include: bespoke legal advice that turns on the specific facts of a client's situation and the professional judgment of the advising lawyer; court advocacy, where strategic and interpersonal judgment are central; complex multi-jurisdictional analysis where the interaction of different legal frameworks requires qualified understanding of each; and any advice where the legal or factual position may have changed since the AI model's training cutoff.
Why Legal AI Literacy Is Categorically Different
Most AI literacy frameworks focus on efficiency and risk management. For legal professionals, AI literacy also intersects with professional duties that are non-delegable: the duty of competence, the duty of candour to the court, the duty of confidentiality to clients, and the fiduciary obligations owed to clients in many practice contexts.
A solicitor who submits an AI-generated document to a court without review, and that document contains a fabricated case citation, has not delegated the error to the AI tool. They have committed a professional conduct breach — and potentially a contempt of court — personally. The AI tool has no professional standing. The lawyer does.
This is the framing that must underpin every decision about AI use in legal practice. AI outputs are drafts, research orientations, and processing aids. They are under your professional supervision. You are responsible for what leaves your office bearing your name.
A practical risk matrix for categorizing legal tasks by AI suitability has two axes: the consequence of an AI error (low to high) and the degree of judgment required (routine to complex). High-volume routine tasks with low consequence for individual errors — scanning a hundred NDAs for non-standard terms — sit in the green zone. Bespoke advice to a client on a fact-specific situation with significant consequence — tax structuring, litigation strategy, regulatory exposure assessment — sits in the red zone. Most AI use in legal practice should start in the green zone and extend carefully into amber only with robust review protocols.
A law firm partner argues that if an AI tool produces output that turns out to be accurate, the fact that it was not reviewed before being sent to a client is a process problem, not a professional responsibility problem. How does the lesson's framework respond to this argument?
Select one answer.
Evaluating Legal AI Tools
The legal AI market has matured significantly but remains uneven. Not all tools marketed to law firms have been built with the rigor that legal work demands. When evaluating a legal AI tool, the questions that matter are:
What was the model trained on, and does that training corpus include jurisdiction-specific legal materials relevant to your practice area? A tool trained primarily on US case law will be less reliable for English and Welsh common law research. A tool trained on general commercial contracts may handle NDA review well but struggle with construction contracts.
How does the tool handle uncertainty? A well-designed legal AI tool should indicate when it is uncertain, flag when a question falls outside its reliable scope, and invite verification rather than presenting outputs as definitive. A tool that always produces confident answers is more dangerous in a legal context than one that flags its own limitations.
What is the vendor's data handling commitment? For any tool that processes client information — even for review and summarisation — the vendor's data handling terms determine whether use is compatible with your confidentiality obligations. This is addressed in detail in Lesson 5.
Do not adopt a legal AI tool based solely on a vendor demonstration using prepared examples. Demonstrations are optimized to show the tool at its best. Before deployment, test the tool against your actual work — your jurisdiction, your practice area, your document types — and specifically probe its error behavior. Ask it questions where you know the answer is wrong or where you know the correct answer requires nuanced jurisdictional knowledge. How it fails tells you more than how it succeeds.
A junior solicitor uses an AI legal research tool to identify relevant case law for a client matter and sends the results to the supervising partner without review, noting only that 'the AI found these cases.' Why is this workflow professionally problematic, even if the cases turn out to be accurate?
Select one answer.
Exercise
Your Task
Select a legal AI tool you currently use or plan to evaluate. Design three test queries: (1) one that falls clearly within the tool's stated competency and training domain; (2) one at the boundary of its scope, such as a jurisdiction-specific procedural question; (3) one where you already know the correct answer requires nuanced jurisdictional knowledge. Run all three queries and document how the tool handles uncertainty — does it flag its own limitations or present all outputs with equal confidence?
Success looks like
- You can articulate which query type the tool handled most reliably and why
- You have identified at least one instance where the tool presented uncertain or jurisdiction-specific content with unjustified confidence
- You have a written note comparing error behavior across all three query types — this is your starting calibration for where to trust this tool in practice
Watch out for
- Selecting only queries where you don't know the correct answer — you need at least one where you can verify accuracy independently
- Accepting confident AI output without checking against a primary source such as the relevant legislation, case report, or regulatory guidance
- Choosing a tool domain where the AI performs well to avoid negative results — the boundary and adversarial queries are the ones that teach you the most
Hint
For the boundary query, try a question that requires jurisdiction-specific procedure knowledge (e.g., an English Civil Procedure Rules point if you practice in England and Wales). Tools trained primarily on US case law will often answer such questions with false confidence.
- AI is demonstrably useful in legal practice for volume document review, research orientation, drafting scaffolding, due diligence execution, and plain-English summarisation — but unreliable for bespoke advice, complex multi-jurisdictional analysis, and any task requiring qualified professional judgment on specific facts.
- Professional duties in legal practice — competence, candour, confidentiality, fiduciary obligations — are non-delegable to AI tools, which means the lawyer remains professionally responsible for every output that leaves their office regardless of how it was produced.
- Use a risk matrix with two axes — consequence of AI error and degree of judgment required — to categorize legal tasks: high-volume routine tasks are appropriate AI candidates; bespoke advice on fact-specific situations with significant consequences is not.
- Evaluate legal AI tools by probing their error behavior on tasks within your jurisdiction and practice area — how a tool fails tells you more about its suitability than how it performs on vendor-prepared demonstrations.
- Legal professionals need a more critical AI literacy than most professions because the professional and legal consequences of AI error in legal work are categorically more severe than in other contexts.