The tool categories nobody explains clearly

A senior partner corners you after a demo and asks: "Doesn't this do the same thing as Harvey, our CLM, and that review tool we bought last year?" You have thirty seconds to answer before she moves on. If you cannot draw the line between drafting, review, research, discovery, and knowledge retrieval tools, you will either kill a useful purchase or greenlight a redundant one. This lesson gives you that line.

Why the confusion is structural, not your fault

Legal AI vendors market horizontally. A single product demo might show drafting, a chat interface over your documents, and a research citation, all in one flow. That is a sales choice, not a technical reality. Underneath, these are different jobs, built on different data, with different failure modes.

The taxonomy that actually matters is about what the tool is optimized to produce, not the interface it wears.

The five categories, honestly described

Drafting tools

These generate first-draft text: contracts, memos, pleadings. Examples include Harvey (built with OpenAI models, used by firms like Allen & Overy) and Spellbook, which plugs into Microsoft Word for contract drafting.

What they are honestly good at: producing a plausible first pass fast, working from templates and clause libraries, and adapting tone across document types.

What they are not good at: getting the deal-specific commercial terms right without a lawyer supplying them, and knowing your firm's precedent bank unless it has been explicitly integrated. A drafting tool with no connection to your document management system (DMS, the repository like iManage or NetDocuments where firm documents live) is drafting from generic training data, not your firm's actual precedents.

Review tools

These read existing documents and flag issues: risk clauses, missing terms, deviations from a playbook. Examples include Kira Systems (now part of Litera) and Luminance, both built originally for due diligence contract review.

Honestly good at: high-volume consistency checks across hundreds of contracts in M&A due diligence or lease portfolio audits, catching what a tired associate misses at 11pm.

Not good at: judgment calls on commercial risk appetite. It flags a limitation-of-liability clause; it does not tell you whether your client should accept it.

Research tools

These answer legal questions with citations to case law, statutes, and secondary sources. Examples include CoCounsel (Thomson Reuters, formerly Casetext), Lexis+ AI, and Westlaw Precision.

Honestly good at: surfacing relevant authority faster than manual Boolean search, summarizing case holdings.

Not good at: being trusted blind. Every major provider still requires citation verification because hallucinationhallucinationEine Hallucination liegt vor, wenn ein KI-Modell Output erzeugt, der flüssig und selbstbewusst klingt, aber faktisch falsch, erfunden oder nicht durch die Quelldaten gedeckt ist.Vollständige Definition ansehen → (a model generating a fabricated or nonexistent citation with high confidence) remains a real risk. The infamous *Mata v. Avianca* (2023) sanctions case, where lawyers filed a brief with fake ChatGPT-generated citations, is the standing cautionary tale cited in nearly every legal AI ethics CLE (continuing legal education) since. See the court's own record via CourtListener for the underlying opinion.

eDiscovery tools

These process large document sets for litigation disclosure obligations: identifying relevant, privileged, or responsive documents. Examples include Relativity and DISCO.

Honestly good at: technology-assisted review (TAR), a form of supervised classification that has been court-approved since *Da Silva Moore v. Publicis* (2012, S.D.N.Y.), for cutting document review volume by large margins on million-document matters.

Not good at: substituting for privilege judgment calls on close-call documents, which still need lawyer sign-off for defensibility.

Knowledge retrieval tools

These let fee earners ask natural-language questions of the firm's own internal knowledge: past matters, precedent clauses, internal memos. This is usually built on retrieval-augmented generation (RAG), a technique where the model retrieves relevant firm documents at query time and generates an answer grounded in them, rather than relying only on what it learned in training.

Honestly good at: turning "has anyone at the firm done a similar earn-out structure" into a five-second answer instead of a Slack message to eight partners.

Not good at: working well if the underlying DMS metadata is a mess. RAG quality is bounded by retrieval quality: garbage indexing in, garbage answers out.

A simple worked distinction

Here is the test partners should actually ask, framed as a decision tree:

Is the tool primarily generating new text?        -> Drafting
Is it primarily flagging issues in existing text?  -> Review
Is it primarily answering questions from external
  legal sources (case law, statutes)?               -> Research
Is it primarily classifying documents for
  litigation disclosure?                            -> Discovery
Is it primarily answering questions from the
  firm's OWN internal documents?                    -> Knowledge retrieval

Most "does this overlap" confusion resolves the moment you ask which column the new tool sits in. Harvey and CoCounsel look similar in a demo (both are chat interfaces) but Harvey is optimized for drafting workflows and CoCounsel for research citation quality. They are not substitutes.

The DMS integration question that decides everything

None of these five categories deliver firm-specific value without integration into the DMS. A drafting tool with APIAPIApplication Programming Interface: eine standardisierte Schnittstelle, über die Anwendungen kommunizieren und Daten austauschen, ohne die interne Funktionsweise der jeweils anderen zu kennen.Vollständige Definition ansehen → access to iManage can pull the firm's actual precedent for a similar deal. Without that connection, you are paying for a generic layer on top of a general-purpose model, which associates could get elsewhere.

This is the single most useful diagnostic question for evaluating any vendor pitch in 2026: does this read and write to our DMS, or does it operate next to it? "Next to it" tools create a second silo and additional copy-paste risk, which matters under confidentiality obligations and, for firms with EU-linked matters, GDPR (General Data Protection Regulation) data handling rules.

Pricing and licensing, briefly

Legal AI vendors mostly price per-seat annually (estimate: several hundred to low thousands of USD per seat per year as of 2025, varying widely by vendor and firm size) or per-document/per-matter for discovery tools, where document volume drives cost directly. Some drafting and research tools are moving toward consumption-based pricing tied to query volume, similar to APIAPIApplication Programming Interface: eine standardisierte Schnittstelle, über die Anwendungen kommunizieren und Daten austauschen, ohne die interne Funktionsweise der jeweils anderen zu kennen.Vollständige Definition ansehen → pricing models elsewhere in AI. Get the pricing model matched to actual usage pattern: a research tool priced per-seat is wasteful if only litigation partners use it three times a month; a discovery tool priced per-document is dangerous if document volumes spike unpredictably on a large matter.

Why this taxonomy is also a change management tool

Partners resist AI not because they misunderstand the technology, they resist because nobody has told them clearly what a tool actually replaces versus what it assists. "This drafts your first pass, it doesn't replace your judgment on the earn-out terms" is a sentence a skeptical partner can trust, because it is specific and honest about the boundary. "This uses advanced AI to transform your practice" is a sentence that gets the vendor thrown out of the room, and rightly so.

🎬 [VIDEO: "How Law Firms Are Actually Using AI Right Now" — youtube.com — search for recent legal tech conference panels (e.g. Legalweek or ILTA sessions) featuring practicing lawyers describing real deployments, not vendor demos]

Key Takeaways

• Categorize by output, not interface: drafting generates text, review flags issues in existing text, research answers questions from external legal sources, discovery classifies documents for litigation, knowledge retrieval answers questions from internal firm documents.
• A tool's value depends heavily on DMS integration. Ask "does it read/write to our DMS or sit next to it" before any pilot.
• HallucinationHallucinationEine Hallucination liegt vor, wenn ein KI-Modell Output erzeugt, der flüssig und selbstbewusst klingt, aber faktisch falsch, erfunden oder nicht durch die Quelldaten gedeckt ist.Vollständige Definition ansehen → risk is real and specific to research tools; citation verification is non-negotiable, as *Mata v. Avianca* demonstrated.
• Match pricing model (per-seat, per-document, consumption-based) to actual usage pattern, not vendor default.
• Honest, specific descriptions of what a tool does and does not replace are the actual change management lever with skeptical partners, more than any feature list.