Who this is for
This is a practical guide for working lawyers — solo practitioners, small-firm partners, in-house counsel, and associates at larger firms — on using AI for document-heavy tasks without getting sanctioned, disbarred, or embarrassed. We’ll cover where the real AI document review wins are, how to handle privilege and confidentiality, and where AI is still more trouble than it’s worth.
We are not going to tell you AI will replace the profession. It won’t, at least not soon. We are going to tell you that the lawyers who learn to use it well will be doing substantially more thinking and less formatting in five years, and the ones who don’t will lose on price. Both of those things can be true at the same time.
The tasks where AI is already earning its keep
Contract review — first pass
The place AI document tools shine is the first read-through of a long agreement. Drop a 60-page merger agreement into Claude and ask it to identify every defined term, flag clauses that deviate from your standard playbook, and produce a plain-English summary of the key commercial deal points. What would have taken a first-year associate an afternoon takes the AI ten minutes — and the associate’s afternoon can go to negotiating the five things that actually need negotiating.
The catch: you still need to read the document. AI’s first-pass review misses things humans catch, especially cross-references buried across sections. The AI is a smart highlighter, not a substitute for actually reading the clauses your client is about to sign. Most effective use: AI produces an issue list and draft markup, you verify and refine.
Deposition and transcript summaries
Modern transcripts are machine-readable out of the box, which makes them excellent candidates for AI processing. A typical workflow: feed a 400-page depo into an AI with a 200K+ context window (Claude, Gemini Pro, or an internal tool built on similar foundations), ask for a chronological summary organized by topic, and get back something that would have taken a paralegal two days.
Tables of contents, exhibit indices, and "every time the witness mentioned the June 3rd meeting" searches are all fair game. For the mechanics, the blog post on converting depositions and transcripts walks through the specifics. The key preprocessing step: convert PDFs to markdown first. Transcript PDFs often have line numbers and headers that confuse AI tools unless cleaned up. See also the legal document converter for the tooling side.
Document production and privilege review
Large-scale e-discovery has been using machine learning for over a decade (technology assisted review, predictive coding). What’s changed in the last two years is that general-purpose LLMs can now do first-pass privilege screening on documents they’ve never seen before, catching the obvious "this is from the client to the attorney about legal strategy" cases. The false positive rate is high enough that humans still do the final pass, but the volume reduction is real.
Legal research drafting
Use with discipline. A typical pattern that works: use Westlaw or Lexis+ AI features (which are trained on licensed case law and return proper citations) for the actual research, then use ChatGPT or Claude to help draft the memo around those findings. The general-purpose tools are better at prose. The vendor tools are better at citations. Don’t swap the roles. See legal research with markdown for a deeper workflow.
Client-facing simplification
Undervalued use case. A plain-English explanation of a lease assignment, a non-compete, or a settlement offer is something AI writes remarkably well — often better than what the lawyer would write, because the lawyer is too close to the formal language. Draft with AI, review for accuracy, send to client. The client experience improves; your billable minute on the "what does this mean?" phone call drops.
Privilege, confidentiality, and the "where does the file go" question
This is the part that matters. Everything above assumes you’ve already answered correctly: where does this document go when I hit enter?
| Deployment | Privilege-safe? | Use for |
|---|---|---|
| ChatGPT Free / Plus | No | Legal research on public cases, CLE-style learning, drafting from publicly available models |
| Claude Pro | No (consumer TOS) | Same as above — default-no-training helps, but not a substitute for contractual protection |
| ChatGPT Team / Enterprise, Claude for Work | Generally yes with DPA | Most firm work, subject to engagement letter disclosures |
| Lexis+ AI, Westlaw Precision AI, CoCounsel | Yes, designed for the profession | Legal research and document review; built-in citation grounding |
| On-prem / VPC deployments | Yes | High-stakes matters, regulated clients, sensitive investigations |
Two tests we recommend for each file before you upload:
- Would you forward this document to a vendor for a routine service (copy shop, office services, IT) without a signed NDA? If no, don’t paste it into a free consumer chat.
- If this conversation came up in discovery, would you be comfortable explaining your data-handling to the court? If you’d have to say "I assumed ChatGPT didn’t keep it," you haven’t done the diligence.
The deeper version of this discussion is in secure document conversion and privilege. For the file-handling specifics, see our broader legal industry materials.
Preprocessing: why markdown matters in legal work
A detail most lawyers don’t know: legal PDFs are among the worst file types for direct AI upload. They often contain:
- Multi-column briefs where AI extractors read left-to-right across both columns
- Line-numbered pleadings where line numbers get glued to the text
- Footnotes that migrate into the middle of paragraphs
- Scanned exhibits that OCR silently fails on
- Repeating headers and footers on every page
Each of these degrades the AI’s answer quality without producing a visible error. Converting to markdown first strips out the PDF scaffolding and keeps only the content — with tables preserved, headings intact, and a much smaller token footprint. A 400-page deposition as markdown often fits in a context window that the original PDF would blow past. The PDF converter handles the conversion.
Realistic workflow: contract attorney reviewing a 400-page merger agreement
A concrete example of how this actually plays out in practice. Firm attorney representing the buyer in a mid-market M&A deal. Received the target’s 400-page merger agreement the night before a partner call.
- Preprocess. Convert the PDF to markdown. See the output to confirm nothing was lost in conversion (the scanned exhibit in section 4.3 gets flagged for a re-run through vision tooling).
- Issue spotting. Paste the markdown into Claude (firm’s Enterprise-contracted deployment) and ask for an issue list organized by section, flagging clauses that deviate from a standard buyer-friendly playbook the attorney pasted in as reference.
- Cross-reference check. Ask "identify every defined term used in section 7 and tell me whether each is defined in the definitions section." This is the kind of mechanical check associates quietly dread; AI nails it.
- Timeline extraction. Ask "pull every date, deadline, and time commitment mentioned in the document and put them in chronological order with the section reference." Produces a timeline that goes straight into the partner memo.
- Draft memo. Ask AI to draft a 3-page summary memo for the partner, organized by commercial point. Edit for accuracy and house style.
- Verify. Attorney reads the actual agreement for the 12 flagged clauses in detail. Everything else is spot-checked against the AI’s summary.
Total time: ~3 hours. Without AI, that’s a Saturday. Notice that the attorney still read the document; AI just made the read faster and more focused.
Where AI is still more trouble than it’s worth
Anything with real citations
Consumer AI chats invent case citations. This has cost real lawyers real money and real bar complaints. Never use ChatGPT or Claude for case citations unless you verify each one in a real research database. For research, use purpose-built legal AI tools that ground on licensed case law.
Jurisdictionally sensitive interpretation
"Does this contract violate the California Consumer Privacy Act?" is a question general AI will confidently answer and often get wrong — because the training data pools together many state regimes and the model blurs them. Use AI to summarize facts and issues; use a human attorney (or a jurisdiction-specific research tool) for legal conclusions.
Anything a judge might ask about your work product
If the answer to "how did you generate this?" would embarrass you, either you shouldn’t have used AI for it or you shouldn’t be producing the work product. The bar is not "was AI involved" — the bar is "is this work product I can stand behind." AI doesn’t change that standard; it just gives you more ways to fall short of it.
Client communication and engagement letters
A growing number of firms explicitly address AI use in their engagement letters. Common provisions:
- A general statement that the firm may use AI tools in the course of the representation.
- A commitment that client data will only be processed by enterprise-grade tools with contractual protection.
- A clause about pricing — some firms pass AI cost savings to clients, some keep them, some charge flat fees that implicitly account for the productivity gain.
Talk to your malpractice carrier about their current expectations. Policy positions on AI use have shifted meaningfully in the last 18 months.
A minimum viable setup
If you’re a solo or small-firm lawyer starting from scratch, the minimum effective stack in 2026 is:
- Claude for Work or ChatGPT Team as your primary AI, with an actual Data Processing Agreement in place.
- Lexis+ AI or Westlaw Precision AI for any task that involves case citations.
- A document converter for turning PDFs into clean markdown before upload. Our legal document converter is built for this. See also the lawyers using AI to analyze court cases post for more on the preprocessing side.
- A documented AI policy for your firm, even if you’re a solo. Two pages. Who can use what, for which tasks, with what disclosure.
What to watch next
The interesting frontier for legal AI in 2026 is not capability — the models are plenty capable. It’s integration: AI tools that read directly from your document management system, that draft in Word without roundtrips, that produce reviewable output with citations to specific pages of specific exhibits. The firms that deploy those well, with proper governance, will pull noticeably ahead on price-per-matter. If you want to read ahead, the post on small law firms and AI document analysis covers the operational side, and converting legal documents to markdown for AI review covers the preprocessing specifics.