Most conversations about AI in law focus on specific tools: ChatGPT, contract review software, legal research platforms.
But understanding which type of AI approach fits which type of task is far more valuable.
That knowledge helps you evaluate any tool, ask better questions of vendors, and make smarter decisions for your practice.
The Core Question
Before asking "Can AI do this?" ask "What kind of AI task is this?"
There are five main approaches. Each suits a different type of problem.
1. Plain LLM (Direct AI Generation)
You give the AI a task, it responds from its general training. Best for one-shot tasks where you don't need the AI to reference specific documents.
Legal examples: drafting a clause, summarising a judgment, generating negotiation talking points.
2. RAG (Retrieval-Augmented Generation)
The AI is connected to a specific set of documents and retrieves relevant content before responding. It works from your material, not general knowledge.
Legal examples: querying your firm's contract repository, searching past case files, finding precedents within your own database.
3. Pipelines
A fixed sequence of automated steps, each feeding into the next. Best when the process is predictable and repeatable.
Legal examples: client intake → conflict check → NDA generation → review flagging, all as one automated flow.
4. Agents
The AI can make decisions mid-task: choosing what to look at next, using different tools, adapting its approach. Best for complex tasks that require judgment, not just execution.
Legal examples: due diligence research where the AI decides which documents to pull based on what it finds, or a matter summary that draws from multiple sources.
5. Fine-Tuning
The AI model is trained further on firm-specific data so it behaves in a particular way consistently: your style, your jurisdiction, your standards.
Legal examples: a model that always drafts in your firm's preferred language, or one trained on your jurisdiction's specific regulatory framework.
What to Watch Out For
AI can get things wrong in ways that are hard to detect. A few things to keep in mind across any approach.
Hallucination. All AI can fabricate facts, citations, or clauses with confidence. Always verify outputs against source material. RAG reduces hallucination risk but does not eliminate it.
Agent risk. Agents carry higher risk in high-stakes tasks because they take sequences of actions, and errors can compound.
Fine-tuning quality. Fine-tuning requires quality training data. Poor inputs produce poor, consistently poor outputs.
Temperature. Controls how predictable or creative the AI's output is. Low temperature for a contract clause gives you consistent, conservative language. High temperature for brainstorming gives you more varied, unexpected ideas.
Context window. How much text the AI can read at once. A model with a small context window will "forget" the beginning of a long agreement by the time it reaches the end.
Prompting. How you phrase your instruction shapes the output. "Summarise this contract" gives you something generic. "Summarise this contract, flagging any clauses that deviate from market standard, in bullet points" gives you something usable.
Chunking. How long documents are split before being fed into RAG. If a clause spans two chunks, the AI may miss its full meaning — like reading a sentence with the middle cut out.
Embeddings. How the AI converts text into something it can search. Think of it as the index at the back of a textbook — it's what lets RAG find the right passage quickly.
Grounding. Tying the AI's output to specific source material. A grounded answer cites the clause it's drawing from. An ungrounded one may sound equally confident but be invented.
AI Blueprint is a free, open source AI platform for lawyers — with local document intelligence, multi-agent councils, purpose-built personas, and email automation, all running privately on your machine. Download it at github.com/rohasnagpal/AI-Blueprint.