Key takeaways
- Compare tools by where their answers come from: general training data, a curated legal database, your own documents or your precedent bank.
- General chatbots are useful for wording and structure but unreliable for law; research platforms with retrieval are better but still need checking.
- Test any tool on twenty questions from your own practice with known answers, including traps, before you buy.
- Every category needs the same core safeguards: contracted terms, approved tasks, human checking and citation verification.
Compare categories, not brands
The market for AI legal tools changes every month, with new products, new features inside existing products and frequent rebranding. Comparing brand names quickly goes out of date. A more durable approach is to ask one question of every tool: where do its answers come from? The answer tells you most of what you need to know about its strengths, its risks and the safeguards it needs.
Six categories cover most of what UK firms are offered. Many products combine more than one, so use the categories to frame questions rather than to pigeonhole a supplier.
The six categories
| Category | Where answers come from | Strengths | Main risk | Key safeguard |
|---|---|---|---|---|
| General-purpose chatbot | Patterns in broad training data | Wording, structure, plain English | Invented law and citations | Never used as a source of law |
| Assistant inside your office suite | Training data plus your own emails and files | Summaries and drafts from material you hold | Oversharing across matters; permissions | Tidy permissions before switching on |
| Legal research platform with AI | Retrieval from a curated legal database | Answers linked to real sources | Misdescribed holdings; out-of-date law | Read every cited source |
| Drafting add-in with clause library | Your firm's or the vendor's precedents | Consistent clauses, faster drafting | Wrong clause for the context | Precedent ownership and review |
| Contract review and extraction | The documents you upload | Finding clauses across many documents | Missed or misread clauses | Sample review by a person |
| Firm-built tool over own knowledge | Your precedents, notes and know-how | Firm-specific answers | Stale or unvetted internal content | Curated source set and named owner |
Category notes
General-purpose chatbots are the most capable writers and the least reliable lawyers. Use them, under a firm contract, to restructure a letter, simplify an explanation or suggest headings. Do not ask them what the law is. The Divisional Court has said that freely available generative AI tools are not capable of conducting reliable legal research[1].
Assistants built into an office suite work over your own emails and documents, which makes them good at summarising what you already have. Their risk is permissions: if a fee earner can technically open files from another team's matters, the assistant may surface them. Review sharing and access settings before switching such features on.
Legal research platforms with AI retrieve real materials from a curated database and then write an answer with citations. This greatly reduces invented authorities, but the summary can still misstate what a case decided or miss a recent change. The citations are a reading list, not a conclusion.
Drafting add-ins with clause libraries are closest to traditional document automation. They are as good as the precedents behind them, so they need the same ownership and review routine. Contract review and extraction tools are strong at finding clauses across large sets of documents, for example in due diligence, but a person should review a sample of both flagged and unflagged documents.
Firm-built tools that answer questions from your own know-how can be powerful in a specialist practice. Their quality depends on which internal documents they draw from. An old precedent or an unvetted research note can be repeated confidently. Curate the source set and give it an owner.
What clients expect from AI
Legal service users are broadly optimistic. In the Legal Services Board's 2026 public panel, 70% expected AI to improve ease of use of legal services, 66% better accessibility and 64% better affordability[2].
Source: Legal Services Board, 2026[2]
Show the data
| Item | Value |
|---|---|
| Ease of use | 70% |
| Accessibility | 66% |
| Affordability | 64% |
On accuracy, views are more mixed: 58% expected AI to lead to better accuracy for legal service users, while 28% expected worse accuracy[3]. Which way that goes for your clients depends largely on which category of tool you use and how you check it.
How professional firms tend to buy
Professional, scientific and technical businesses are more likely to access AI by buying external software or ready-to-use services than through free tools[4]. Large language models are the most widely used AI technology among UK businesses with ten or more employees, used by 18% in June 2026[4]. For law firms, buying makes sense: a contract gives you terms on confidentiality, data location and reuse of inputs that free services rarely offer. It also gives you someone to hold to account.
Buying also changes who does the evaluation. Free tools tend to arrive one enthusiastic user at a time; a purchase goes through the approval route in your AI policy, with the COLP and data protection lead involved. That slower path is a feature, not a cost, because it is where the questions about confidentiality and checking get asked before any client data is involved.
Testing a tool before you buy
Build a test set
Write twenty questions from your own practice area for which you know the answer and the authority, drawn from recent files.
Include traps
Add questions involving an overruled case, a recent statutory amendment, a point where the law differs between England and Wales and Scotland, and a question with no clear answer.
Run every candidate tool
Use the same questions and the same wording for each tool, and save the outputs.
Score blind
Have a fee earner who did not run the tests score each answer for accuracy, citation validity and whether the tool admitted uncertainty.
Check the contract
Compare data terms, retention, training on your inputs, sub-processors and exit, alongside the scores.
Decide on tasks, not just tools
Approve each tool for the tasks where it scored well, and record the rest as not approved.
Questions to ask every supplier
- Where do the tool's answers come from, and can it show the source for every statement it makes?
- How current is the legal content, and how quickly are new judgments and legislation added?
- Does the tool cover the law of England and Wales specifically, and does it flag when a source is from another jurisdiction?
- Are our prompts, uploads and outputs used to train or improve models for anyone else?
- Where is our data processed and stored, for how long, and by which sub-processors?
- Can administrators see who used the tool, when and for which matter?
- What does the tool do when it does not know the answer?
The last question is revealing. A tool that says it cannot find an authority is far more useful to a fee earner than one that always produces an answer. Ask the supplier to demonstrate it on a question with no good answer, and watch what happens.
Safeguards for every category
- A firm contract with clear terms on confidentiality, storage location and use of inputs.
- An entry in the firm's approved tools register listing the tasks it may be used for.
- Training before access, covering the tool's known weaknesses.
- A named person checks every output that leaves the firm.
- Every citation verified in an authoritative source and read in context.
- An incident log for errors and near misses, reviewed by the AI lead.
Confidentiality applies whichever category you choose. The Law Society's generative AI primer highlights the risk of misuse or disclosure of confidential information, especially for smaller firms[5]. Before uploading client documents, confirm what the tool does with them and for how long.
Example: a commercial team compares three tools
Example (fictional): a commercial property team tests a general business chatbot, a research platform with AI and a contract review tool on twenty questions and two sets of leases. The chatbot writes the clearest summaries but cites a non-existent case and misstates a statutory notice period. The research platform gives accurate, linked answers on fifteen questions, but misses a recent amendment on one. The contract review tool finds break clauses across forty leases quickly, but misreads two conditional breaks.
The team approves the chatbot for plain-English drafting only, the research platform for research with full verification, and the contract review tool for first-pass extraction followed by fee-earner review of every break clause. Each decision goes into the register with its conditions.
Choosing between them
- If your main need is clearer client writing, a contracted general tool may be enough.
- If you research law regularly, a research platform with retrieval is the safer choice, with verification.
- If you review many similar documents, contract review tools save the most time.
- If your value lies in specialist know-how, consider a firm-built tool, but only with a curated, owned source set.
For the rules that should sit around any of these, see writing a firm AI policy and AI hallucinations and fake citations. For lower-risk uses outside legal work, read AI for practice management and admin tasks, and for client-facing wording, telling clients when you use AI.
Frequently asked questions
Which AI tools are safe for legal research?
Research platforms that retrieve answers from a curated legal database are safer than general chatbots, but every citation still needs to be found and read in an authoritative source before use.
Can we use a general AI chatbot for drafting?
For wording and structure, under a firm contract and with checking, yes. Do not rely on it for statements of law or citations.
How should a firm test an AI legal tool?
Run twenty questions from your own practice with known answers, including traps such as overruled cases and recent amendments, score the answers blind and compare data terms alongside the results.
Do AI tools inside our office software need special care?
Yes. They work over your own emails and files, so they can surface documents a user has technical access to. Review permissions and sharing before switching them on.
References
The National Archives (Find Case Law) (2025). R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank QPSC [2025] EWHC 1383 (Admin) (opens in a new tab) ↑
Legal Services Board (2026). AI powered legal services: LSB Public Panel expectations (infographic) (opens in a new tab) ↑
Legal Services Board (2026). AI in Legal Services: Research Report (LSB Public Panel) (opens in a new tab) ↑
Office for National Statistics (2026). Artificial intelligence in UK businesses: 2023 to 2026 (opens in a new tab) ↑
The Law Society (2026). Generative AI: the essentials (opens in a new tab) ↑