Document Automation
Gavel Workflows
By practice area
By company size
AI Contract Review
Resources
LearN
Case Studies
Pricing
%20(1).avif)
A plain-English guide to the AI concepts lawyers need to understand: LLMs, RAG, fine-tuning, hallucinations, zero data retention, and more, with insight into how Gavel Exec applies them.
Easy intake and document automation to auto-populate your templates.
If you're a lawyer evaluating AI tools, you'd probably love a resource on the key terms and how the legal AI companies are handling them. This article is a resource for you, with a primer on LLMs, RAG v. fine-tuning (and why nobody fine tunes anymore), zero data retention, hallucinations, and security.
Whatever product you choose, you should walk away from this understanding enough to ask the right questions of any legal AI vendor.
Gavel Exec is an AI-powered contract review and redlining assistant directly inside Microsoft Word. It analyzes your contract or set of contracts, reviews the document against a set of legal standards (either market standards or your own defined standards and repository of documents), identifies issues, explains the reasoning, and suggests changes. It can also draft entire documents, portions of documents, identify risk, and run a playbook based on your own rules.
Each flagged issue follows a structured format: what the requirement or market standard is, how the AI identified the issue in the document, and what redline language it recommends. You can accept, reject, or modify every suggestion. The AI provides the analysis; the lawyer makes the call.
Now let's get into how it all works.
A large language model (LLM) is the core AI technology behind tools like ChatGPT, Claude, and most legal AI products on the market today. At a high level, an LLM is a system that's been trained on massive amounts of text data to predict and generate language.
The major LLMs right now include OpenAI's GPT models (the technology behind ChatGPT), Anthropic's Claude, Google's Gemini, and Meta's LLaMA. Each has different strengths, different architectures, and different performance characteristics. They're also all evolving rapidly; a model that's best-in-class today might be leapfrogged in three months.
When we say Gavel Exec is LLM-agnostic, we mean we're not locked into any one model (for example OpenAI's GPT models or Claude). The ay we're built lets us plug in any LLM and swap between them based on performance.
So why is this important?
First, model performance varies by task. We actively benchmark, meaning we systematically test and compare, different models against each other on real legal work, with practicing lawyers evaluating the outputs. What we've found is that no single model is best at everything. For example, right now, OpenAI tends to outperform on redlining tasks, generating precise inline edits with clean formatting. Claude tends to be stronger on drafting and analysis, producing more nuanced legal reasoning and better-organized output. So we use both, routing each task to whichever model our benchmarking shows is most accurate.
Second, the AI landscape is moving fast. New models and major updates ship every few months. Being agnostic means we can adopt improvements immediately. If a new model comes out tomorrow that outperforms everything else on contract analysis, we can integrate it without rebuilding our product.
If you're evaluating legal AI tools, this is a good question to ask any vendor: which LLM are you using, and what happens when a better one comes along?
When you take a general-purpose LLM and want to make it useful for a specific domain like law, there are two main approaches. Understanding the difference will help you evaluate any AI product, not just ours.
Fine-tuning means retraining the model itself on domain-specific data. You feed it thousands of examples of legal analysis (contracts, memos, redlines) and the model's internal parameters (the mathematical weights that determine its outputs) adjust to reflect that training. After fine-tuning, the model has legal knowledge baked into its core.
The upsides: the model responds faster (no need to retrieve external data) and can develop a more intuitive feel for domain-specific patterns. The downsides are that fine-tuning is expensive and slow to update. And most importantly, fine-tuning is tied to a specific model. So, f you spend months and significant resources fine-tuning, say, GPT-5 for legal work, that investment dies with GPT-5 when a newer, better model comes out. You either start the fine-tuning process over from scratch on the new model (more time, more money, more data risk) or you keep running on an increasingly outdated foundation.
Retrieval-augmented generation (RAG), which Gavel Exec uses, takes a different approach. Instead of changing the model, you give it the right reference material at the right time. When you ask the AI to review a contract, the system first retrieves relevant information (documents, market standards, playbook rules, deal context, your preferences) and includes that alongside the document as context for the model. The model uses that information to generate its analysis, but none of it is absorbed into the model's weights. The base model stays untouched.
We use RAG at Gavel for several reasons. It's more transparent because the AI is working from specific, identifiable sources rather than opaque training data you can't inspect. It's easier to keep current, because when legal standards change or we improve a playbook, those updates are reflected immediately without retraining. It's more auditable, since you can trace the AI's reasoning back to the specific rules it applied.
A RAG system is only as good as the information it retrieves. Here's how we've structured ours.
Layer 1: Market-Standard Playbooks. We've hired practicing attorneys to build detailed review playbooks for a wide range of corporate transactional and standard contract types. These are structured rules covering market standards across industries, company sizes, and jurisdictions. Each rule specifies the standard, how to identify deviations in a document, and what redline language to recommend. This is the baseline knowledge the AI works from on every review. When people ask how we combat hallucinations, this is the core answer: the AI isn't making things up, it's applying specific rules written by lawyers.
Layer 2: Deal-Specific Context. A contract doesn't exist in isolation: the same indemnification clause reads differently depending on whether you're looking at a $500K asset purchase or a $50M merger. Our Projects feature lets you upload related transaction documents, like term sheets, LOIs, prior drafts, board resolutions, side letters. Gavel Exec incorporates that context into its review, so it can flag inconsistencies between documents, verify that negotiated terms are accurately reflected, and analyze provisions in light of the actual deal rather than reviewing in a vacuum.
Layer 3: User-Specific Learning. Over time, the system learns your preferences. If you consistently adjust certain types of suggestions, prefer particular drafting conventions, or have specific risk thresholds, Gavel Exec adapts to reflect that. In AI terms, this is sometimes called personalization or preference learning — the system is tuning its outputs to match your patterns without retraining the underlying model.
Here's the important distinction: it learns for you, and your data is never shared or used to train the system. Your preferences and patterns improve your experience only. They're never fed back into the general model, never used to train the AI for other customers, and never shared across accounts (unless you give access to a colleague).
Hallucinations are one of the most talked-about risks in legal AI. A hallucination occurs when an AI model generates information that sounds plausible and confident but is factually wrong, like citing cases that don't exist, inventing contract provisions that aren't in the document, or misrepresenting legal standards.
This happens because of how LLMs work, and because when LLMs don't have information, they may make it up. This is something we've been very focused on solving at Gavel, and here's how we do it:
Hallucinations are far more likely when an AI model is working without grounding. If you paste a contract into ChatGPT and ask "is this indemnification clause market standard?" — the model is pulling from its general training data and essentially guessing. There's no reference material, no defined standard, no deal context.
The more context you give an AI model, like specific rules to apply, documents to reference, standards to measure against, the less room there is for the model to fill gaps with fabricated information. This is one of the core reasons we built Gavel Exec around structured playbooks and the ability for users to upload their own deal documents. The AI isn't generating analysis from scratch. It's applying specific, lawyer-written rules to your document while referencing the actual materials from your transaction. When the model can point to a specific playbook rule or a specific provision in your term sheet, it doesn't need to make things up.
It's also worth noting that hallucination risk in a contract review tool is fundamentally different from hallucination risk in, say, a legal research tool.
When a research tool hallucinates a case citation, that's a serious problem because you might rely on it in a brief without catching it. But contract review operates differently. Gavel Exec is a thought partner. Every suggestion it makes goes into your Word document as a tracked change or comment that you review before accepting. You're the one reading the contract. The AI is flagging issues and proposing language, but nothing gets into the final document without you approving it.
On top of that, the AI can reference back to your actual documents. If it flags an inconsistency between the APA and the term sheet, you can look at both documents and verify. If it recommends a redline based on a playbook standard, you can evaluate whether that standard applies to your deal. The AI's work is checkable against the source material in a way that a hallucinated case citation in a research memo isn't.
None of this means hallucinations don't matter, but the risk profile is different when the AI is working alongside you on a document you're already reading, versus generating standalone work product you might not independently verify.
Security in AI is a set of specific practices. Here are the ones that matter most for legal work, and what we do at Gavel.
Zero data retention (ZDR). When any AI tool processes your document, it sends data to an LLM provider's servers (OpenAI, Anthropic, etc.) for the model to analyze. The question is: what happens to that data after the analysis is done? With zero data retention, the answer is nothing. The provider doesn't store your prompts, doesn't store the model's responses, and doesn't log the interaction. Your data is processed in memory and discarded. We have ZDR agreements with every LLM provider we use. If you're evaluating other AI tools, ask whether they have ZDR — and whether it applies to all data or just some of it.
Encryption. Encryption in transit means your data is protected while it's being sent between your computer and our servers (using TLS, the same protocol that secures online banking). Encryption at rest means your data is protected while it's stored on our servers. We do both. This applies to documents you upload, the AI's analysis, and any stored preferences or project data.
No training on client data. This is distinct from ZDR. ZDR means the LLM provider doesn't store or train on your data. But what about the legal AI company itself? Some vendors use client documents and interactions to improve their own models or systems. We don't. We do not use client documents, edits, or interaction data to train any AI model, meaning not the underlying LLMs, and not any internal models. Our playbooks and standards are built independently by our legal team and lawyers we've engaged and hired.
Data siloing. Siloed data means each client's information is architecturally isolated, not just protected by access controls (like a password), but stored separately so there's no shared data layer between accounts. Your documents, your projects, your preferences, and your review history are completely inaccessible to other users (unless you actively give someone access).
There's a lot of terminology in the AI space, and it can be hard to separate substance from marketing. Here's the quick reference version of the concepts covered above:
If you're evaluating AI tools for legal work, whether it's Gavel Exec or another tool, these are the concepts worth understanding and the questions worth asking. If you want to see how this works in practice, we're always happy to walk through it.
See how legal AI can give you superpowers. Try Gavel Exec free here, or book a demo with our sales team.
LegalOn's pre-built playbooks get you started fast, but the redlines often need heavy cleanup. This guide compares 5 LegalOn alternatives for 2026 — including Gavel Exec, CoCounsel, Ivo, StrongSuit, and Spellbook — by features, AI architecture, pricing, and redline quality.
Not getting the contract review you want from Ivo? We review 5 alternatives to their AI contract review, including features, pricing, training data, and more.
Frustrated with Spellbook? We review the top Spellbook alternatives legal professionals are using, as we compare their accuracy, features, pricing, and how to switch easily.
