What small firms actually need from AI
Solo and small-firm practice is not short of judgment. It is short of hours. A two-lawyer office handling employment matters, contract review, and a scattering of landlord disputes produces the same categories of document hundreds of times a year: engagement letters, demand letters, discovery summaries, client status updates.
AI is not useful for deciding what the letter should say. It is useful for the mechanical layer around that decision: condensing a 180-page deposition transcript into a three-page index of who said what and where, turning rough intake notes into a clean client email instead of spending forty minutes on it, or cutting a motion from 4,800 words to the 3,500 the page limit allows without losing the argument.
Every one of those is a text transformation. None require the tool to know who your client is, what the matter concerns, or what your strategy is. That distinction is the foundation of a safe setup, because text transformation is exactly the class of work that runs well inside a browser on your own machine.
The privilege problem, stated plainly
Attorney-client privilege and the duty of confidentiality are related but not identical, and the difference matters here. Privilege concerns what can be compelled from you in a proceeding. The confidentiality duty is broader and generally covers information relating to a representation regardless of its source. Check how your own jurisdiction's professional conduct rules frame both, because details vary.
For tool selection, the practical risk is not dramatic. It is mundane: your text leaves your device.
When a tool requires a signup, treat that as your first signal. Accounts exist to associate activity with a person, and associated activity is stored activity. When a tool runs in the cloud, your input crosses a network and is processed on hardware neither you nor your client controls. Neither fact is automatically disqualifying, but both should prompt two questions: does my engagement letter permit this, and would I be comfortable explaining the choice to a judge?
Three more resolve most cases. Does the text leave my device? If it does, is it retained, and for how long? Is it used to train or improve the provider's model? If you cannot get plain answers from the vendor's own documentation, the tool does not belong near client material. Vagueness is itself an answer.
This is where the browser-based, no-account model earns its place. If processing happens on your machine, there is no retention window to audit, no training clause to negotiate, and no third-party disclosure to draft. For a small firm with no IT department and no vendor review process, that removes a category of work rather than adding one.
One precision note. "Runs in the browser" is not automatically the same as "private," since some tools present a browser interface and quietly send your text to a server. The reliable test is behavior, not marketing copy: if the tool works with your connection switched off, or the vendor states plainly that processing is on-device, you have a fact. If the page only says "secure" and "encrypted," you have a claim.
A quick scenario
Take a realistic Thursday. You are a solo practitioner with an employment matter heading to mediation. Your client was terminated three weeks after raising a safety complaint internally. You have a 190-page deposition transcript of the plant manager, eleven internal emails the client forwarded, five pages of handwritten intake notes, and a mediation brief due Monday with a 12-page limit.
Start with the transcript. You read it once already, but you need the four exchanges where the manager contradicts the written safety log, and you cannot reread 190 pages this week. You run the document through a PDF summarizer that processes locally, get a structured digest with page anchors, and go back to the source for the two passages you will quote. The method in our guide to summarizing PDFs covers how to set length and scope so the output stays faithful.
Next, the emails. The client forwarded them as one thread with subject lines stripped, and you need a chronology. You condense each message into three or four lines with our free AI text summarizer, paste the results into one document in date order, and the sequence becomes obvious: complaint on the 4th, schedule change on the 6th, termination on the 25th. The step-by-step approach in how to summarize long articles transfers cleanly to email threads.
Then the drafting. Your demand letter template reads too much like the last matter. You use a paraphraser to rework the factual sections so they track this client's timeline, sentence by sentence, checking each one against your notes. Mechanical rewording of your own drafting is safe. Letting a tool restate a date, a dollar figure, or a quotation is not, and the distinction is spelled out in rewriting without a signup.
Finally, the compression pass. The brief is 18 pages and the limit is 12. You check the count with a word counter, cut the two sections that repeat the chronology, then run a grammar check before the draft goes to the client. The file never left your laptop, so if anyone later asks which vendors had access to privileged material, the answer is one sentence.
What to look for in a legal-safe tool
Five checks, in the order that narrows the field fastest. Does it require an account, and if so, what for? Plenty of tools gate genuinely local functionality behind a login purely for analytics. Does the vendor state where processing happens, in plain language, on the tool page? Is there a stated retention period, and if text does reach a server, does your plan say in writing that it is excluded from training? And does the tool work offline? That last check is the quickest available and settles the question.
Two things that matter less than vendors suggest: whether the tool is marketed to lawyers, and how long its feature list is. A general-purpose text tool that runs locally protects confidentiality better than a legal-specific platform that stores everything in the cloud and calls it a secure vault.
Where the tools fit
The mapping below is deliberately narrow. It covers work that is mechanical, reversible, and checkable in a few minutes.
| Task | Tool | Why it fits legal work |
|---|---|---|
| Digesting a transcript or exhibit | PDF summarizer | No upload, so privileged documents stay on your machine |
| Building a chronology from client emails | AI text summarizer | Condenses each message without sending the thread anywhere |
| Reworking template language for new facts | Paraphraser | Changes wording you already wrote, not the underlying facts |
| Final proofread before filing or sending | Grammar checker | Surface errors only, no rewrite of your analysis |
| Hitting a page or word limit | Word counter | Instant counts against the limit, no document upload |
Paid tools still earn their place on research breadth and volume: reading across dozens of authorities at once, or processing a discovery set of five thousand documents. Those jobs are hard to do fully on-device today. What they rarely buy is better summarizing of one transcript, better grammar correction, or better paraphrasing, which have largely plateaued and need no subscription.
Common mistakes
- Pasting client-identifiable facts into a free chat tool because the output looked good, and reading the retention terms afterwards.
- Assuming the paid tier of a well-known tool excludes your input from training, without confirming it in writing for the plan you actually hold.
- Letting a paraphraser change a date, a dollar amount, or a quoted statement. Those are the three places where a small edit becomes a misstatement in a filed document.
- Treating an AI summary as a source rather than a map back to the source, then quoting from it instead of opening the transcript.
- Uploading an entire client file to save an hour of reading when one document would have answered the question.
- Forgetting that drafts, working notes, and internal memos carry the same confidentiality duty as filed documents.
- Believing that deleting a chat thread removes the copy held on the provider's side.
- Chasing a newer model when a single-purpose local tool would have finished the task in a quarter of the time.
A defensible default workflow
One rule survives contact with real matters: anything containing client information is processed locally, and everything else is negotiable.
In practice, summarizing, paraphrasing, grammar checking, and word counting happen in your browser with no account. Research on public sources, marketing copy for the firm website, and continuing-education reading can use whatever is fastest. Keep a one-line note in the matter file recording which tools touched client material. If the local-only rule holds, that note stays one line for the life of the matter.
Two habits reinforce it. Read every AI-assisted sentence against the source document before it goes out, especially any sentence containing a number, a date, or a quotation. And when a tool asks you to sign in before it will work, treat that as a prompt to ask whether this task needs the account at all.
One caution belongs in every firm's own policy. AI output is not legal advice and is not a substitute for your analysis. It is a drafting aid that can be wrong in confident, well-formed sentences, so a human with the matter file open reviews everything before it reaches a client or a court. If you want a concrete check on whether a document contains machine-generated passages, our private AI detector runs the same local way.
Plain legal writing still does most of the work underneath the tooling. Short sentences, defined terms, and a structure a reader can follow will do more for a brief than any tool. The tools buy back hours, not the judgment those hours get spent on.
Frequently asked questions
Can lawyers use free AI tools without waiving privilege? The safer framing is confidentiality rather than waiver. Privileged text sent to a third-party server may remain protected from compelled disclosure in many situations, but the confidentiality duty can be breached regardless. Local, no-account tools avoid the question entirely because nothing leaves your device.
Does a no-signup tool guarantee my text stays private? No signup is a strong signal but not proof. Test it: run the tool with your connection disabled, or look for an explicit on-device processing statement. Vague claims about security and encryption are not evidence about where your text goes.
Which legal tasks should never go through an AI tool? Anything producing legal conclusions, strategy, or citations you have not personally verified, and anything requiring the tool to hold your whole matter file. Mechanical work on text you already wrote is the safe category.
How accurate are AI summaries of depositions and transcripts? Good enough to locate passages, not good enough to quote. Treat a summary as an index with page anchors, then open the source and read the actual exchange before it appears in anything you file.
Do I need to tell clients I use AI assistance? Requirements vary by jurisdiction and many are still developing. A short disclosure in your engagement letter saying you use assistive software, paired with a statement that client material is not sent to third-party AI services, is inexpensive and answers the question before it is asked.