Legal AI glossary 2026: the terms vendors use and what they mean
Definitions as buyers encounter them, not as vendors would prefer them. Where a term is genuinely contested we say so, because a word that means two things across two demos is worse than a word nobody uses.
Contract terms
| Term | What it means |
|---|---|
| CLM | Contract lifecycle management. Software covering creation, approval, signature, storage and obligations. The most over-applied term in this market: vendors doing two of those five describe themselves as CLM. |
| Playbook | Your written positions on standard clauses, and the fallbacks you will accept. The word describes both the document and the software feature that enforces it, which is a frequent source of confusion in demos. |
| Clause library | Approved wording your team reuses. Useful only if somebody maintains it; an unmaintained library is a source of outdated positions rather than consistent ones. |
| Redlining | Marking up a counterparty's draft with tracked changes. Ask whether a tool produces real tracked changes or a comment describing one, because the difference decides whether the other side can accept it. |
| Obligation tracking | Recording what a signed contract commits you to do, and when. Distinct from renewal reminders, which only track dates. |
| Fallback position | The next thing you will accept when your preferred wording is rejected. Encoding these is what separates playbook software from a clause library. |
| Third-party paper | A contract drafted by the counterparty. Most tools are much better on your own templates than on inbound paper, and this is worth testing directly. |
| Pre-signature and post-signature | Before and after execution. Tools are usually strong at one and thin at the other, and this is the single most useful question for placing a vendor. |
AI terms
| Term | What it means |
|---|---|
| Large language model | The general-purpose model underneath most of these products. Several vendors build on the same handful of models, so the model name tells you much less than the workflow around it. |
| Retrieval-augmented generation (RAG) | Fetching your documents and giving them to the model as context, rather than relying on what the model memorised. Almost every tool here does this; it is table stakes rather than a differentiator. |
| Extraction | Pulling structured data (parties, dates, values, clause types) out of unstructured documents. The hardest thing to do well on badly scanned or unusually drafted files, and where analytics tools genuinely differ. |
| Technology-assisted review (TAR) | In eDiscovery, using a model trained on reviewer decisions to prioritise or exclude documents. Established and court-tested, unlike most terms on this page. |
| Hallucination | The model asserting something untrue with confidence. In legal work the dangerous form is not an invented case but a plausible misreading of a real clause, which is much harder to spot. |
| Human in the loop | A person reviews output before it takes effect. A meaningful claim only when you ask which step, because it is often applied to a stage nobody was going to automate anyway. |
| Agentic | The model takes multi-step actions rather than answering once. Currently the least standardised word in this market. Ask what it does without a human, and what it cannot do. |
| Grounding | Tying an answer to a source you can check. Ask whether the tool cites the specific clause it relied on, because an uncited answer cannot be verified. |
Commercial terms
| Term | What it means |
|---|---|
| Spend under management | The value of invoices processed through a legal spend tool in a year. Used as a pricing unit, which means the bill rises with the thing you bought the tool to reduce. |
| Seat or per-user pricing | Priced on people. Predictable, and it penalises giving access to occasional users, which is often where the value is. |
| Volume pricing | Priced on contracts, gigabytes, entities or matters. Cheaper at low volume and unforecastable at high volume, which is the trade. |
| Annual escalator | A contractual yearly increase, commonly reported at 5 to 10% in this market. Negotiable at signature and very hard at renewal: 5% compounding turns $100,000 into $121,000 by year three. |
| Implementation fee | A one-off charge to get you live. Routinely a fixed proportion of the annual subscription, and rarely mentioned first. |
| Total cost of ownership | Licence plus implementation, training and internal time. Reported to add 40 to 60% to the licence on large IP and CLM deployments. |
| Published, reported, unsourced | Our own labels. Published means the vendor states a rate. Reported means a figure exists from marketplace or procurement data. Unsourced means we could find neither and say so rather than guessing. |
Terms to push back on
- "AI-native" and "AI-first". No agreed meaning. Ask which specific step the model performs and what happens when it is wrong.
- "End-to-end". Ask which end. Most tools covering creation through signature do not cover obligations, and most covering obligations do not draft.
- "Enterprise-grade". Usually a claim about access controls and uptime commitments. Ask for the specific controls and the SLA rather than accepting the adjective.
- "Accuracy" without a benchmark. Ask what was measured, on whose documents, against whose judgement, and whether the methodology is published.
- "Integrates with". Ask whether that means a supported native integration, a generic connector somebody has to configure, or an API and a project.