NotebookLM vs Claude Projects vs MCP for Case Files

By Jude Lee · · Comparison

Two lawyers reviewing case documents on a laptop next to stacked file boxes in a firm conference room

Spend any time in law firm AI threads and you see the same arc: someone uploads a deposition transcript to a chatbot, gets a startlingly good summary, then hits a wall trying to do it across an entire file room. The wall is real, and it is not a model-quality problem. It is a plumbing problem. Which tool you need depends on whether the AI has to read documents, produce documents, or find documents across matters it has never seen.

What a grounded notebook actually does for a lawyer

NotebookLM is Google’s grounded research tool. You upload sources — PDFs, Docs, pasted text, transcripts — and the assistant answers using only those sources, with inline citations pointing back to the passage it relied on. Per Google’s NotebookLM documentation, source grounding and citation-linked responses are the core design choice; it is built to stay inside the file rather than roam.

For legal work, that constraint is the feature. A general chatbot asked “what did the plaintiff say about the stairwell” may blend your document with its training data. A grounded notebook is far more likely to say the sources don’t address it. You still verify every citation — grounding reduces fabrication, it does not eliminate it — but the loop is much shorter because every answer links to a page.

Where it fits: a new file lands with a few hundred pages of medical records, a policy, and three transcripts, and you need to orient fast. Load it, ask for a treatment timeline with source citations, ask where the defendant admits notice, and spend your reading time on the twelve pages that matter.

Where it stops: notebooks are read-oriented. They don’t draft your motion, don’t write back to Clio or NetDocuments, and each notebook is an island. Source limits per notebook vary by plan, so check Google’s current documentation. More importantly, every one of these tools has a total context ceiling, and large document sets degrade quietly rather than failing loudly — recall across a 6,000-page production thins out before anything visibly breaks. The fallback is chunking: one notebook per custodian, per date range, or per deposition, instead of one notebook for the whole case.

Claude Projects: the same file, plus the ability to produce work

A Claude Project is a persistent container holding shared project knowledge (your uploaded documents) plus project instructions. Two mechanisms get conflated here and shouldn’t be: instructions are per-project prompt context that shapes every conversation inside that one project, while skills are packaged, reusable procedures you can invoke across projects and share with colleagues. Availability of both varies by plan tier, so confirm against Anthropic’s current product documentation for the edition you pay for.

That distinction matters for recurring work. A “Deposition Summary — Firm Standard” skill can specify your page-line format, the topics you always index, and the rule that every assertion carries a cite — written once, behaving the same for every associate, in every matter. Project instructions can’t travel that way.

NotebookLM-style notebook
Best for: fast, cited orientation to an unfamiliar file. Strong grounding discipline. Minimal setup. Weak at drafting in your firm’s voice, no write-back to your systems, one notebook per matter with no cross-matter search.
Claude Projects
Best for: an active matter you’ll work in for weeks. Shared file plus firm-specific instructions and reusable skills, so Q&A and drafting happen in one place. Still requires manual uploads, still siloed per project.

Both share a ceiling: a human moves files in, a human moves work product out. Fine for one matter. Miserable across fifty.

The fourth category is vertical legal AI — CoCounsel, Lexis+ AI, Harvey and similar — plus firm-wide enterprise chat deployments. A general-purpose assistant has no licensed case law corpus and no citator. If the job is researching authority, validating that a case is still good law, or producing work where the vendor’s own documentation commits to grounding in primary law, a vertical tool is the better buy and the general assistants aren’t close. Enterprise deployments win on a different axis: SSO, admin controls, firm-wide retention policy and logging in one place rather than per-lawyer accounts. The tradeoffs are cost per seat and a workflow you can’t reshape. Verify capability claims against the vendor’s current documentation, not a demo.

The connected option: stop uploading entirely

MCP — the Model Context Protocol — is an open standard, created and open-sourced by Anthropic and since adopted by other AI vendors and tool builders, for giving an assistant governed access to a system you already run. Instead of exporting PDFs out of NetDocuments into a chat window, you expose your document store or matter database as tools the assistant can call, with permissions you define.

The shift is scope, not convenience. A connected assistant can answer across matters — “which open files have a deadline in the next 21 days and no responsive pleading in the folder” — because it queries live data. It can also act: open a matter, file a draft, log time. Whether it should be allowed to act is a separate decision, laid out in read-only vs write access for law firm AI agents.

Notebooks and projects change how one lawyer reads one file. MCP changes what the firm can ask about every file at once. Those are different purchases.

MCP is also the expensive option. Check whether your DMS or practice management vendor already ships a connector before scoping a custom build. A firm-specific server means authentication, permission mapping to your matter security model, and maintenance when the underlying API changes. Don’t build one to summarize a deposition.

Confidentiality decides more than features do

The ABA’s Formal Opinion 512 (July 2024) addresses generative AI under the Model Rules — competence, confidentiality, client communication, and fees. Read it in full. Several state bars have issued guidance that goes further or differs; confirm your jurisdiction’s position with your state bar directly.

Operationally the question is narrower than “is AI safe”: which account tier am I on, and what do its terms say? Consumer and business tiers of the same product frequently differ on exactly that.

Working the economics yourself

Ignore any dollar figure a vendor puts on this, including ours. Use your own inputs.

H × R
Value of recovered orientation time: hours saved per matter × your realization-adjusted hourly rate
Illustrative formula — supply your own H and R
N × H × R
Annualized: multiply by matters per year where the workflow actually applies
Illustrative formula
1 matter
Reasonable pilot size before buying anything beyond a seat license
Editorial recommendation

Recovered hours only convert to revenue if they land on billable or business-development work rather than evaporating. A custom MCP build has to clear a much higher bar: build cost plus maintenance, against hours recovered firm-wide. The long form is in our automation ROI framework.

A sensible sequence

  1. Pilot on one closed matter

    Use a file you already know the answers to. Ask the same ten questions of each candidate tool and check every cite against the source. A workable threshold: if more than one in ten cites points to a passage that doesn’t support the claim, the tool is not ready for live-matter use on that document type. You are testing verification cost, not eloquence.

  2. Settle the terms question in writing

    Ask the vendor four questions and keep the answers: Is our input retained, and for how long? Is it used to train models, and can that be disabled for our account? Who inside the vendor can access it? What is your process when you receive a subpoena or government request for customer data? Then confirm the account tier you actually pay for matches those answers.

  3. Pick one job and package it as a skill

    Deposition summaries, chronologies, or production triage. Encode the format once so output is consistent across everyone who uses it.

  4. Keep the human checkpoint explicit

    Nothing leaves the firm — no filing, no client letter, no privilege call — without a lawyer confirming the underlying cites.

  5. Only then ask whether to connect systems

    Re-uploading the same documents weekly, or needing cross-matter answers, is the signal to price a connection — vendor connector first, custom build second.

When the answer is none of the above

If nobody can find the current version of a document, the fix is document management hygiene, not an AI layer over the chaos. If deadlines get missed, a rules-based calendaring system with court-rule sets beats a probabilistic assistant. And if you handle three matters a month, the manual read is still cheapest. Plenty of law firm AI shopping ends with the correct conclusion that the workflow was never the bottleneck.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides