AI tools for solo attorneys: reduce the admin load without outsourcing judgment

A solo practice has a special pressure: the lawyer is often the intake person, researcher, writer, billing reviewer, project manager, and client communicator. AI can reduce the administrative load between those roles. It can turn an intake into a fact checklist, a pile of documents into a chronology, a research question into a plan, and a verified status into a clearer client update.

The temptation is to treat fluent output as legal work. That is where the risk begins. A model can invent a citation, flatten a factual conflict, misread a jurisdiction, expose confidential information, or give a client a confident answer that no lawyer has approved. The ABA's Formal Opinion 512 makes the core point practical: existing duties around competence, confidentiality, communication, supervision, candor, meritorious claims, and fees still apply when a lawyer uses generative AI.

My operating rule would be simple: AI may organize, compare, draft, and ask questions; the attorney must verify, decide, advise, and take responsibility. The smaller the practice, the more important it is to make that boundary explicit because there may be no separate technology or compliance team to catch a mistake.

I would classify every proposed use into three buckets. Green work uses public, synthetic, or nonconfidential information. It includes public-authority research organization, generic office templates, hypothetical examples, website drafts, and checklists. It still needs legal verification when it contains a legal proposition, but it has a smaller confidentiality surface.

Amber work touches an actual matter but can be narrowed, redacted, or handled in an approved legal environment. It includes document classification, chronologies, discovery maps, contract issue lists, and client-status drafts. The lawyer should know what data the tool receives, what it retains, who can access it, and how the output is reviewed and corrected.

Red work is the judgment itself: whether a client has a claim, whether a filing is accurate, whether a privilege applies, whether a settlement is acceptable, whether a fact is credible, whether a deadline has changed, or whether a client should follow a course of action. AI may help prepare the material, but it does not get the final authority.

New-client intake is often where small practices lose time. Notes are incomplete, names are inconsistent, deadlines are buried in a call, and the initial question is not the legal problem the client ultimately needs addressed. AI can turn approved intake notes into a structured handoff: facts, people, documents, timeline, missing information, potential issues for review, conflict-check prompts, and next administrative step.

The model should preserve the difference between a fact and an inference. The client saying that a notice arrived Tuesday is not the same as the deadline being Tuesday. A good intake assistant flags the question and points the attorney to verify it. It should not decide whether to accept the matter, estimate a result, or provide advice before the lawyer has completed the professional review.

Use a standard output and save it in the matter system. Ask the person reviewing the intake whether the checklist found missing dates, names, or documents and whether it created false certainty. Measure time from first contact to a complete review, follow-up cycles, missed conflict questions, and the percentage of matters that had a clear next step.

A chronology is one of the most useful AI-assisted artifacts for litigation, investigations, employment matters, real estate disputes, estate work, and many other practices. Give the tool approved document excerpts or metadata and ask for date, event, people, source reference, confidence, and unresolved conflict. Then compare the result with the original materials.

The instruction to preserve conflict matters. If two documents give different dates, the output should show both and ask for resolution. If a witness account is disputed, the chronology should label the account rather than choose a version. A summary should never become the only record when the original document contains the nuance that could change the case.

Use the chronology to prepare questions, not to replace the lawyer's analysis. Ask which dates affect a limitation period, which documents need authentication, which facts are missing, and which assumptions drive the current theory. The attorney decides what is material and how the evidence should be characterized.

AI can help a lawyer formulate a research plan. Ask it to break a question into jurisdiction, authority level, date range, elements, factual assumptions, counterarguments, and primary sources to retrieve. It can suggest search terms and organize authorities that the lawyer supplies. It cannot be trusted as the final source of law without verification.

A safe research workflow has a hard stop: retrieve the actual opinion, statute, regulation, rule, or official guidance; read the relevant text; confirm the citation; check whether it remains current and applicable; and record the source behind each proposition. Treat every model-generated citation as a lead until that process is complete.

For a solo attorney, the verification record can be lightweight. Keep a research table with proposition, authority, support, jurisdiction, current-status check, and reviewer. That table helps when a brief changes, a client asks where a conclusion came from, or a court requires the lawyer to explain the basis for an argument.

AI can classify documents by date, custodian, topic, responsiveness, issue, and follow-up. It can identify duplicates, summarize a defined excerpt, or help create a review queue. Those are useful accelerators when the original file remains available and the lawyer understands the tool's limits.

Privilege is not a label a model should apply invisibly. Ask it to flag possible privilege questions, identify the people and communications involved, and point to the source. A lawyer makes the privilege determination under the applicable rules and facts. The same applies to relevance, waiver, authenticity, credibility, and production decisions.

Use sampling and quality checks. Review documents marked responsive, nonresponsive, privileged, and uncertain. Track missed documents, false positives, inconsistent classifications, and the time needed to correct them. If the tool cannot expose the basis for a classification or preserve an audit trail, limit it to a lower-risk sorting task.

A contract assistant becomes more useful when it has a client-approved playbook. Instead of asking to review a contract, define the client's commercial objective, risk tolerance, fallback positions, required clauses, approval owner, and prohibited changes. AI can compare a draft to that playbook and produce clause, issue, business question, fallback, and owner.

The lawyer still decides the legal and commercial position. A clause that looks unusual may reflect a negotiated business tradeoff. A model may miss an interaction between provisions, a defined term, an exhibit, or a jurisdiction-specific issue. Ask it to identify questions and cross-references rather than to declare a contract safe.

For client communications, separate the legal explanation from the business choice. The client may accept a risk for speed or price. AI can create a decision table, but the lawyer must explain the consequence clearly, confirm the client's instruction, and preserve the approved version in the matter record.

A client usually wants to know what happened, what is next, what the lawyer needs, what decisions are open, and whether anything changed. AI can draft that structure from verified matter facts. It can also produce versions for a client email, an internal task list, and a status note.

The review needs to be more than grammar. Check every date, party, procedural posture, commitment, and characterization. Remove any sentence that implies an outcome, advice, or approval that the lawyer has not provided. If the client needs legal advice, the update should direct the attorney to give it rather than allowing a summary to stand in its place.

Use a clear source and approval process. The client update should come from the matter record, be reviewed by the responsible attorney, and be sent through the approved channel. A tool should not send a message automatically because a date appeared in a document.

In discovery, AI can help create a custodian map, topic list, search questions, production checklist, and review summary. It can identify a missing date range or show that a document set contains several spellings of a key name. These tasks can reduce the mechanical work around a review.

Traceability is nonnegotiable. Preserve the original files, collection information, search parameters, reviewer decisions, privilege process, and production record. Ask the model to cite the document or excerpt behind a conclusion. Do not let a generated summary become the only evidence of what was reviewed or why a document was withheld.

The attorney must understand the method well enough to explain it and supervise any nonlawyer assistance. Test the workflow on known documents and edge cases before relying on it. If the output is not reproducible or the tool cannot show what it saw, use it for brainstorming or organization rather than a final discovery determination.

Solo attorneys do not need to begin with case analysis. AI can help with engagement-letter checklists, deadline-review prompts, billing narratives, website FAQs, meeting action items, internal procedures, referral follow-up, and client onboarding. These workflows still need an approved source and a human review, but they keep sensitive legal judgment out of the first experiment.

Billing deserves care. Formal Opinion 512 discusses reasonable fees and the difference between using a tool and charging for time that was not actually spent. Keep records truthful. Do not turn an AI-generated draft into an imaginary block of manual work, and do not pass on general technology-learning costs as if they were client-specific legal work without a defensible basis.

Measure practice improvements that a solo lawyer can feel: time from intake to engagement, missed administrative tasks, invoice clarity, client questions caused by unclear updates, deadline-review completion, and time spent searching for a document. The goal is more reliable service, not simply a larger volume of generated text.

Ask what the tool receives, stores, indexes, and sends to other services. Does it use customer data to train a model? What are retention and deletion settings? Who can access prompts, files, and outputs? Where is the data processed? Are there audit logs, role permissions, encryption, backups, subprocessors, and incident-notification terms? Can the firm export a matter and remove it completely?

Ask how the tool handles uncertainty and errors. Can it show the document behind an answer? Can the lawyer correct an output? Does a correction create a record? Can an administrator disable an integration quickly? What happens when a staff member leaves? If the vendor cannot explain the failure path, keep the workflow to public or synthetic information until the firm has a defensible answer.

Client consent may be relevant depending on the matter, jurisdiction, agreement, tool, and use. Do not assume that a generic notice or a vendor's enterprise plan resolves professional obligations. The lawyer remains responsible for the representation. Review applicable bar guidance, document the firm's decision, and revisit it as the product changes.

Days one through five are baseline and policy. Pick one low-risk workflow such as client-status drafting, intake checklists, or public-authority research organization. Record the current steps, time, corrections, data involved, and approval owner. Write down prohibited uses before anyone pastes a matter into a tool.

Days six through fifteen are shadow mode. Generate drafts without sending or filing them. Classify errors: wrong fact, invented authority, lost uncertainty, confidentiality exposure, missed deadline, unsupported advice, formatting issue, or useful gap found. Save examples as a test set. A wrong citation is a severe error, not a cosmetic one.

Days sixteen through thirty are supervised use. Review every output, record the source and correction, and ask whether the workflow improved client service. Measure turnaround, rework, research verification, missed tasks, client clarification, and staff confidence. Expand only when the lawyer can explain the tool's limits and remains accountable for the final work.

A solo practice should have one approved place for matter records, one written rule for what may enter an AI tool, a small set of tested workflows, and a review checklist. The lawyer does not need a sprawling automation stack to get value. The first win is a reliable handoff and a visible source of truth.

A two- or three-lawyer firm can add a shared prompt and evaluation library, matter-specific permissions, vendor review, and periodic sampling. Define who owns research verification, client communication approval, document review quality, and incident response. Supervision should be a real process rather than an assumption that everyone is using tools in the same way.

The final test is professional judgment. Does the tool help the firm understand the facts and law more carefully, communicate more clearly, and serve the client more reliably? If it makes the firm faster but less able to explain its work, it is the wrong implementation.