The 3 Levels of AI in Legal Work

Ask ten lawyers whether they use AI and you will get ten different answers describing ten different things. One means asking ChatGPT to tighten up an email. Another means running a research question through CoCounsel. A third means scanning four hundred thousand documents for privilege before a production deadline. All three are "using AI," and the phrase has become almost useless because of it. The tools involved differ so much in purpose, in safeguards, and in where they sit relative to your evidence that lumping them together causes real confusion - in budgeting conversations, in client disclosures, and in risk assessments.

A simpler mental model helps. Think of legal AI as operating on three levels. Each level answers a different question, carries a different risk profile, and belongs in a different part of your practice. Once you see the levels clearly, most of the hard questions about adoption become easier to answer.

Level 1: General purpose chat assistants

The first level is the one everyone knows. ChatGPT, Claude, and Gemini are general purpose assistants trained on broad public data. They draft, summarize, brainstorm, translate, and explain. They are inexpensive, immediately available, and remarkably capable at everyday language tasks.

They are also the level with the least connection to your matter. A general purpose chat assistant knows nothing about your case unless you paste material into it, and pasting client material into a consumer chat tool is exactly where professional obligations start to bite. Confidentiality terms vary by product and plan. Retention policies vary. Training-use policies vary. Most law society and bar guidance on AI risk was written with this level in mind, because this is where lawyers most often improvise.

Level 1 tools have a legitimate place. Use them for first drafts of non-confidential text, for explaining unfamiliar concepts, and for thinking out loud. Treat them as a very fast, very well-read stranger. You would not hand a stranger your client file, and the same instinct serves you well here.

Level 2: Legal-specific platforms

The second level is purpose-built for legal work. CoCounsel, Harvey, and Spellbook are the familiar names. These platforms wrap strong underlying models in legal workflows: research memos with citations, contract review against playbooks, deposition preparation, clause drafting. They add the guardrails the consumer tools lack - enterprise confidentiality terms, citation grounding, and interfaces designed around how lawyers actually work.

Level 2 is where strategic thinking happens. You ask what the elements of misrepresentation are under Ontario law and get a structured answer. You ask for a summary of the allegations in a statement of claim and get a clean draft. The work product is analysis and language, and the input is usually a modest set of documents you selected yourself.

What Level 2 tools are not built for is the evidence itself at scale. Uploading a curated folder of fifty contracts is one thing. A live litigation workspace holding hundreds of thousands of emails, with coding panels, family relationships, privilege logs, and production deadlines, is another. We wrote about the practical boundary between these tools and full-scale review in how to get a useful Claira prompt from CoCounsel or Harvey, and the short version is that they work upstream of review, not inside it.

Level 3: Evidence review tools

The third level lives where the evidence lives. Evidence review tools operate inside the eDiscovery platform that already holds your documents, and they work the full population rather than a sample you uploaded. This is the level Claira was built for.

Claira runs inside Nuix Discover. It reads each document the way a human reviewer would, applies your review criteria, and writes its decisions back into the coding panels your team and your QC process already trust. It handles first-pass relevance, privilege flagging, objective coding, summarization, and chronology building across document sets that would take a review team months. The documents never leave the platform, which means the security posture your firm already vetted for Nuix Discover keeps applying.

The distinguishing feature of Level 3 is not just scale. It is that the tool participates in a defensible workflow. Decisions land in coding fields where they can be sampled, audited, overturned, and reported on. You can tell the court, or opposing counsel, exactly what was reviewed, by what criteria, and with what quality control. That audit trail is the difference between AI as a convenience and AI as part of the record.

Context matters at this level too. A reviewer on day one gets a case briefing before touching a document, and Claira works the same way through Case Context, which gives the AI background on the parties, the issues, and the terminology of your matter so every scan is informed by the case rather than guessing from a single document.

Why the levels should not be confused

Most AI missteps in legal practice come from using a tool one level below the job. Pasting deposition excerpts into a consumer chatbot is a Level 1 tool doing Level 3 work, with none of the safeguards. Uploading a fraction of a document population into a research platform and treating the output as a review is a Level 2 tool doing Level 3 work, with no audit trail and no defensible link back to the full evidence set.

The reverse confusion is cheaper but still wasteful. You do not need an evidence review platform to draft a client update, and you do not need a legal research subscription to translate a paragraph.

The levels also answer the budgeting question. Level 1 is a utility, priced like one. Level 2 is a professional subscription that pays for itself across the practice. Level 3 is matter-driven - it scales with the case in front of you, which is exactly how litigation spend behaves.

Putting the model to work

Next time AI comes up in a partner meeting, a client call, or a procurement conversation, start by asking which level is on the table. The question dissolves most of the talking-past-each-other that plagues these discussions. A firm can reasonably run all three levels at once: general assistants for everyday language work, a legal platform for research and drafting, and an evidence review tool inside Nuix Discover for the matters where documents number in the hundreds of thousands.

If Level 3 is the gap in your stack, that is the one we can help with directly. Book a short demo and we will show you what Claira does with a real document set inside Nuix Discover, from first scan to coded results your team can verify.