Key takeaways
- Claude is one of the best AI tools for lawyers, and small firms should use it for strategy and case work, not just drafting.
- Connecting Claude straight to Clio, MyCase, Google Drive, Dropbox or SharePoint works for quick lookups. It breaks down on large case files, because big documents never fully reach Claude.
- The data layer between your systems and Claude decides how fast, how accurate and how expensive every answer is.
- Start with your firm's use cases, then choose where each one gets its data. The subscription is the easy part.
Why small firms use AI but don't see the return
AI for small law firms is no longer an experiment. Clio's 2026 Legal Trends Report for Solo and Small Law Firms found that 75% of small firms and 71% of solos use AI, but only about a third have increased their revenue with it. More than half of solo and small firms still have no policy on how AI is used. The 8am 2026 Legal Industry Report puts AI use at 69% among solo practices and firms with two to five attorneys.
Small law firms and AI, 2026
75% of small law firms
31% of small law firms
Source: Clio, 2026 Legal Trends Report for Solo and Small Law Firms.
The gap between using AI and earning from it usually isn't about the model. It comes down to three things:
- What Claude is used for. Most attorneys use it for emails, letters and quick summaries. That saves minutes, not cases.
- What data Claude can actually see. Claude is only as good as the documents and records that reach it. In most firms, those are scattered across a practice management system, a file drive, email and the phones.
- How the firm prices and runs its work. If you bill by the hour and AI makes you faster, your revenue can go down unless something else changes. We cover this in Will AI replace lawyers?
This guide is about the first two, because that's where small firms lose the most time and money with Claude.
What Claude can do for a law firm in 2026
Claude has become one of the most widely used legal AI tools. In May 2026, Anthropic launched Claude for Legal, a set of practice-area plugins and connections to legal software such as Thomson Reuters, DocuSign and Harvey. In October it added prebuilt legal artifacts, such as trackers and dashboards, that attorneys can customize. Clio and 8am have both released official Claude connectors for their practice management systems.
For a small firm, that means Claude can now help with much more than writing:
- Case work: reading productions, building chronologies, comparing deposition testimony, finding the page where a fact appears, and tracking discovery requests.
- Strategy: testing case theories, spotting weaknesses in the other side's position, preparing for mediation, and planning motions.
- Running the firm: looking up matters, deadlines, time and billing, and answering questions such as which matters have stalled or which leads haven't been called back.
- Drafting: letters, pleadings, discovery responses and client updates, built on the facts of the actual matter.
The catch is in the first bullet. Claude can only work a case file it can actually read. That's where most of the mistakes happen.
The 7 mistakes small firms make with Claude
None of these is a reason to stop using Claude. Each one is a reason to set it up differently.
Mistake 1
Subscribing to Claude and hoping the value shows up
The most common pattern we see: the firm buys a few Claude seats, attorneys use them to polish emails, and six months later nobody can say what changed. Claude didn't fail. The firm never decided what it was for.
Do this instead: list the five to ten tasks that cost your firm the most time or money, such as intake follow-up, discovery review, medical record summaries, deposition prep or billing review. For each one, write down what data it needs and where that data lives. That list, not the subscription, is your AI strategy. Our page on AI for law firms walks through how to build it.
Mistake 2
Pointing Claude at Clio or MyCase and calling it case management
The official connectors are a real step forward. The 8am MyCase connector lets attorneys look up matters, contacts, tasks, deadlines, time and billing in plain language. The Clio connector lets Claude hand legal tasks to Clio, which works on them with your matter context and its own legal sources. Use them.
The mistake is assuming a connector turns Claude into a system that knows your cases. Here's what is actually happening when you ask a question through one of these connectors (also called MCP connectors, after the Model Context Protocol they use):
- Every question is a fresh lookup. Claude calls the practice management system, gets back a result, and answers. Nothing it learns is saved for the next question or the next person.
- Results have a size limit. On claude.ai and Claude Desktop, a single connector result can be about 150,000 characters, and each call has a 240-second timeout, according to Anthropic's connector documentation. That's plenty for a matter record. It isn't enough for a 300-page transcript.
- Practice management systems store documents. They don't read them for Claude. A deposition PDF attached to a matter is a file, not searchable evidence.
- Writes are real. 8am's MyCase tool catalog warns that a write takes effect in MyCase as soon as the tool runs and can't be undone from Claude, and that automated tasks may access your data on a schedule. Decide who is allowed to do what before you connect.
Good fit for a direct connector
"Which of my open matters have a deadline in the next 14 days?" "What did we bill on the Martinez matter last month?" "Show me unpaid invoices over 60 days."
Poor fit for a direct connector
"Compare what all three witnesses said about the HOA vote." "Build a medical chronology from 1,400 pages of records." "Which discovery requests are still missing documents?"
Mistake 3
Connecting Claude to Google Drive, Dropbox or SharePoint and expecting it to read the case file
This is the mistake that costs small firms the most, because it looks like it works. You connect your file storage, ask Claude about the case, and get a confident answer. What you can't see is how much of the file Claude never read.
File storage was built to keep files, not to feed them to AI. Its search has limits that matter for litigation. Geistra Legal's comparison of SharePoint, Google Drive and Geistra Legal follows one fictional case through discovery to show what happens. Here's the first test: a 2,000-page production delivered as a single PDF.
A 2,000-page production in one PDF: how much gets searched?
The first 2 million characters. The rest is never searched.
Nothing. PDFs over 80 pages aren't indexed.
Every page, scanned pages included.
Based on Geistra Legal's comparison, drawing on Microsoft's SharePoint search limits and Google's Drive connector indexing limits. The SharePoint bar is our estimate: at about 1,170 characters a page, 2,000 pages is roughly 2.3 million characters.
File size and format add more gaps:
| SharePoint | Google Drive | Geistra Legal | |
|---|---|---|---|
| Large files | Content skipped past 150 MB (512 MB for PDF and Office files) | Most files capped at 10 MB; PDFs over 50 MB or 80 pages not indexed | Reads files up to 2 GB |
| Zip files | Stored as one file | Stored as one file | Opens the zip and reads every document inside |
| Scanned exhibits | Depends on the file | OCR limited to 80 pages | Read, scans included |
Dropbox works on the same principle: it stores and syncs files, and whatever connector hands those files to Claude still runs into the size limit on what Claude can take in at once. That limit is the next problem.
The mail slot problem
Geistra describes it as a mail slot. However large your case file is, a connector can only push so much text through to Claude at a time. On claude.ai that's about 150,000 characters per result. A 300-page deposition transcript is about 350,000 characters. If the answer is on page 188, a connector that sends the transcript from the top never reaches it.
One question: "How did Whitlock describe the HOA vote?"
Through a file storage connector
About 150,000 of the transcript's 350,000 characters fit through. The answer is never reached.
Through Geistra Legal
Every page is already read and indexed, so Claude gets at most 30,000 characters, taken from the right pages. Answer: Whitlock Dep. 188:4–19.
Adapted from Geistra Legal's comparison. The case and witnesses are fictional. Result limits from Anthropic's connector documentation.
It gets worse with harder questions. In the same comparison, asking Claude to compare testimony across three depositions through SharePoint or Drive runs out of room after about six steps, with transcripts cut off. Through Geistra Legal it finishes in five steps, with every transcript compared and each point cited to a witness, page and line.
The danger isn't that Claude refuses to answer. It's that Claude answers from the part of the file it received, and the answer sounds just as confident either way.
Mistake 4
Paying Claude to re-read the case file
AI is priced by tokens, roughly the amount of text it reads and writes. When Claude works through a file storage connector, it re-reads large chunks of the file for every question, because nothing it read last time was kept. Over a long case, that adds up.
One case, 60 days, 300 questions: pay your AI to re-read, or to answer
270 million tokens · about $1,080
14 million tokens · about $56
Estimate by Geistra Legal, not a live run. Assumes a 4,000-page case, 5 questions a day for 60 days, and $4 per million input tokens.
About 19 times the cost, for one case. Multiply that by the active matters in your firm. And the cost isn't only money: every re-read is time the attorney spends waiting, and every cut-off file is a chance the answer is incomplete.
Mistake 5
Letting the work disappear when the chat closes
An attorney spends an afternoon with Claude matching a production to the discovery requests. The answer is excellent. Then the chat is closed, and the work lives in one person's chat history.
When the work is saved as data instead, the whole team can build on it. In Geistra Legal's example, the discovery requests become a tracker that updates as documents arrive:
Discovery request tracker · updated as documents arrive
From Geistra Legal's comparison. The case is fictional. In file storage, the same requests stay a document with nothing to query, and matching results are lost when the chat closes.
The same applies when an attorney leaves. If the case knowledge lives in personal chats, it walks out the door. We wrote about this in One pane of glass: run your law firm with Claude, Copilot or ChatGPT.
Mistake 6
Skipping the AI policy and the verification step
More than half of solo and small firms don't have an AI policy, according to Clio. That's a risk, not a formality. ABA Formal Opinion 512 applies the existing Model Rules to generative AI. It covers competence, confidentiality, communication with clients, candor toward the tribunal, supervision and reasonable fees, and it says lawyers shouldn't rely on AI output without an appropriate degree of independent review.
A workable policy for a small firm answers a few questions:
- Which AI tools and plans are approved? Use commercial or business plans for client work. 8am, for example, requires a commercial Claude account for the MyCase connector.
- Which connectors may be used, by whom, and with read or write access?
- Where may client data go, and where may it not?
- How is output checked before it's relied on or filed? Answers that cite the exact page make this much faster.
- When do you tell clients, and how does AI affect what you bill?
- Where is AI work kept, so it stays with the matter instead of with the person?
Check your state bar's guidance too. Several states have issued their own opinions.
Mistake 7
Starting with the tool instead of the use case
Most firms ask "which AI should we buy?" first. The better first question is "which problem are we solving, and what data does it need?" A document drafting task needs almost no firm data. A deadline check needs your practice management system. A deposition comparison needs every page of every transcript, read and indexed. Each one belongs in a different place, and choosing that place is what makes Claude fast, accurate and affordable.
The data layer: why it decides speed, cost and accuracy
Between the systems your firm runs and the AI your team uses, there's always a layer that decides what Claude gets to see. In most firms, that layer is accidental: whatever a connector happens to return. A data layer makes it deliberate.
The AI you already use
The data layer · Geistra Legal
The systems you run today
The difference shows up in three places:
- Speed. The documents are read and indexed once, ahead of time. A question goes straight to the right passages instead of waiting for files to be fetched and re-read.
- Cost. Claude reads a few thousand characters of the right pages instead of hundreds of thousands of the wrong ones. That's the difference between $56 and $1,080 in the example above.
- Accuracy. Every page is in scope, scans included, and every answer points to its page and line, so an attorney can check it in seconds. That's the verification Opinion 512 asks for, made practical.
Geistra Legal works with Claude, ChatGPT, Microsoft Copilot and Gemini, so your team keeps the assistant it knows. ScaleLegal is a certified Geistra Legal consultant: we set it up, load your active matters and connect it to the systems you already run. You can see Geistra Legal's full benchmark on its comparison page.
Direct connector or data layer?
You don't have to choose one. Most small firms should use both, each for what it does best.
| Question | Direct connector (Clio, MyCase) | Data layer (Geistra Legal) |
|---|---|---|
| Matter status, contacts, tasks, deadlines | Best fit | Works |
| Time entries, invoices, billing totals | Best fit | Works |
| Find a fact in a 300-page transcript | Often cut off | Best fit, cited to page and line |
| Review a 2,000-page production | Not designed for it | Best fit, scans and zips included |
| Compare testimony across witnesses | Runs out of room | Best fit |
| Track discovery requests over weeks | Lost when the chat closes | Best fit, saved for the team |
| Questions across firm data (calls, leads, billing, cases) | One system at a time | Best fit |
Use cases: match each one to the right data source
Here's how we map Claude use cases for small firms during our assessment. The point isn't the list. It's that every use case gets a deliberate answer to "where does Claude get its data?"
| Use case | Who it helps | Where the data should come from |
|---|---|---|
| Draft letters, client updates and first-pass pleadings | Attorneys, paralegals | Claude alone, plus the facts of the matter |
| Morning deadline and task check | Attorneys, firm administrator | Clio or MyCase connector |
| Billing review and unpaid invoices | Firm administrator | Clio or MyCase connector |
| Medical record summaries and chronologies | PI and workers' comp teams | Data layer: every page, scans included |
| Discovery tracking and production review | Litigation teams | Data layer: saved tracker the team shares |
| Deposition prep and testimony comparison | Trial attorneys | Data layer: cited to page and line |
| Picking up a matter from a departing attorney | Whole firm | Data layer: documents plus AI history kept with the matter |
| Which marketing channels bring cases that pay | Managing partner | Data layer joining phones, CRM, billing and cases |
For the firm-wide questions in the last row, see 10 law firm KPIs every managing partner should see weekly and why law firms lose good leads at intake.
A 90-day plan for a small firm
Days 1 to 30: decide what Claude is for
- List the tasks that cost the firm the most time or money, and rank them.
- For each one, note the data it needs and where that data lives today.
- Write a one-page AI policy: approved tools and plans, connector permissions, client data rules, verification and billing.
- Move everyone onto a commercial plan, so client work isn't done in personal accounts.
Days 31 to 60: connect the right layer to the right use case
- Turn on the official Clio or MyCase connector for lookups, read-only to start.
- Set up a data layer for document-heavy matters, and load your active cases first.
- Pilot two use cases with one practice group, such as medical chronologies and discovery tracking.
Days 61 to 90: measure, train and expand
- Measure time saved, cost per matter and how often answers needed correcting.
- Write short SOPs for each use case and train the team on them.
- Expand to the next practice group, and revisit pricing for the work AI now speeds up.
This is the same rhythm as our 30-Day Strategic Assessment and 90-day sprints: find the problems that cost the most, then fix one at a time.
Claude among the best AI tools for lawyers
Search for the best AI tools for lawyers and you'll find long lists. For a small firm, it's simpler to think in three layers:
- A general AI assistant such as Claude, ChatGPT, Microsoft Copilot or Gemini. This is where your team works every day. Claude stands out for long documents, careful reasoning and its growing set of legal plugins and connectors. If your firm lives in Microsoft 365, Copilot may be the natural choice for some staff.
- The AI inside your practice tools, such as Clio's legal AI or the AI features in your research platform. These are excellent within their own data.
- A data layer that connects the two to your case files and firm data, so the assistant you use can work the whole case, not just the part a connector can carry.
The firms getting the most from AI tools for lawyers aren't the ones with the most subscriptions. They're the ones that decided what each tool is for and made sure it sees the right data.
Claude for lawyers: frequently asked questions
Is Claude good for lawyers?
Yes. Claude is strong at reading long documents, drafting, summarizing and reasoning through legal and business questions, and Anthropic now offers Claude for Legal with legal plugins and connectors. How good the results are depends on how your firm feeds Claude its case data, and lawyers must verify output before relying on it.
Should my firm connect Claude directly to Clio or MyCase?
Yes, for lookups: matter status, contacts, tasks, deadlines, time and billing. Start read-only and decide who can make changes. Direct connectors aren't built to read and reason over a full case file, and every answer starts from scratch, so pair them with a data layer for document-heavy work.
Can Claude read my whole case file from Google Drive, Dropbox or SharePoint?
Not reliably. SharePoint search stops parsing an item after 2 million characters, Google's Drive connector doesn't index PDFs over 50 MB or 80 pages, and Claude accepts about 150,000 characters per connector result on claude.ai. A 300-page transcript is about 350,000 characters, so parts of large files never reach Claude, and the answer won't tell you what it missed.
What is a data layer for legal AI?
A secure layer between your systems and your AI. It reads every document, indexes it by matter, sends Claude only the passages it needs, and saves findings for the whole team. That makes answers faster, cheaper and checkable, because each one cites its page. Geistra Legal is an example, and ScaleLegal is a certified Geistra Legal consultant.
Is it ethical for lawyers to use Claude with client data?
It can be. ABA Formal Opinion 512 applies existing duties to generative AI: competence, confidentiality, client communication, candor toward the tribunal, supervision and reasonable fees. Use a commercial plan, set a written AI policy, check your state bar's guidance, and verify output before relying on it. This guide isn't legal ethics advice for your jurisdiction.
How much does it cost to use Claude in a law firm?
There's the subscription, and there's how much text Claude has to read to answer. On a large case, re-reading files through a storage connector can cost many times more than reading indexed passages. Geistra Legal's estimate for one 4,000-page case over 60 days is about $1,080 through SharePoint versus about $56 through a data layer.
Claude, ChatGPT or Copilot: which is best for a small law firm?
Each is capable. Claude is especially strong with long documents and now has dedicated legal features. Copilot fits firms that run on Microsoft 365. What matters more than the model is the data it can reach. With a data layer like Geistra Legal, different people can use different assistants on the same case data.
Where should a small firm start with AI?
Start with use cases, not tools. List the tasks that cost the most, note what data each needs, write a short AI policy, and pilot two use cases for 90 days. Our 30-Day Strategic Assessment does this with your firm's own data.
Sources
- Geistra Legal: Work your case with AI, compared with SharePoint and Google Drive
- Clio: 2026 Legal Trends Report for Solo and Small Law Firms
- 8am: AI adoption for solo and small law firms in 2026
- Anthropic: Build an MCP server for Claude (size and timeout limits)
- 8am MyCase MCP and MyCase MCP tool catalog
- Clio: Connect Clio to Claude
- Microsoft: Search limits for SharePoint
- Google: Drive connector indexing limits
- ABA Business Law Today: Formal Opinion 512 offers a useful framework
- Artificial Lawyer: Claude for Legal launches
Claude, Clio, MyCase, Google Drive, Dropbox, SharePoint, ChatGPT, Microsoft Copilot and Gemini are trademarks of their owners. ScaleLegal is a certified Geistra Legal consultant and isn't affiliated with or endorsed by the other companies named.
