AI Workflows · Workflow playbook · Updated June 2026
AI Paralegal: What to Delegate and What to Keep
The vendor pitch is "replace your paralegal." The senior move is the opposite: treat AI as a paralegal you supervise. It does the structured legwork at speed; you keep the judgment, the strategy, and the client. Here is the exact line between the two, and the supervision duty that holds it.
Key takeaways
- Supervise, do not replace. The durable framing is an AI paralegal you direct and check, not a tool that removes the human from the loop. That framing is also what Rule 5.1 and 5.3 require: the lawyer stays responsible for the output.
- Delegate the bounded, checkable work. Intake summaries, document indexing, deposition digests, and timeline building are structured tasks with a source you can verify against in minutes. That is exactly where an AI paralegal earns its keep.
- Keep the unbounded, consequential work. Legal judgment, case strategy, and client counsel cannot be delegated. They require weighing risk, reading people, and owning advice, and no model does that for you.
- Confidentiality is the hard line. Privileged and client confidential material does not belong in a public consumer model. Use an approved enterprise deployment that does not train on your inputs, or redact, or do not use AI on that matter at all.
The professional's real problem
Walk into any practice and you will find a senior lawyer doing two very different kinds of work in the same day, billed at the same rate, valued very differently by the client. One kind is judgment: deciding whether to settle, how to frame a motion, what to tell a nervous client at nine at night. The other kind is structured legwork: reading three hundred pages of intake to write a one page summary, building a chronology from a box of emails, indexing a document production so the team can find what it needs, turning a four hour deposition transcript into a usable digest. The first kind is why clients hire you. The second kind is necessary, unglamorous, and quietly eats the hours that should go to the first.
That second category is exactly what the market wants to sell you a product to erase. A wave of legal AI vendors arrived with the same pitch, sometimes said out loud, more often implied: replace your paralegal, automate the support function, cut the headcount. For a senior lawyer that framing is both wrong and a little reckless. It is wrong because the support work is not waste, it is foundation, and a thin or sloppy foundation poisons the judgment built on top of it. It is reckless because it invites you to stop supervising work you remain responsible for. The lawyers getting real leverage from AI are not replacing anyone. They are delegating the structured legwork to a model and supervising it, which frees their own hours for the judgment only they can supply.
The question is not whether AI can do paralegal work. It is whether you will treat the model as a replacement you stop watching, or as a paralegal you supervise and stay responsible for.
Replace your paralegal vs the AI paralegal you supervise
Before the delegation line itself, it is worth being precise about the two mental models, because they lead to very different practices and very different exposure. This is not anti software. It is about where responsibility sits and who is watching the work.
| Dimension | "Replace your paralegal" (vendor frame) | The AI paralegal you supervise |
|---|---|---|
| Who is responsible | Implied to be the tool. The output arrives finished, and the human is invited to trust it. | You are. The model drafts; you review, correct, and certify, exactly as you would a junior's work. |
| What you delegate | Everything the tool claims to handle, often including judgment dressed up as a feature. | Only the bounded, checkable tasks. Judgment, strategy, and client counsel stay with you. |
| Verification | Treated as optional once you trust the tool, which is the moment the risk starts. | Built in. Every delegated output is checked against its source before it is used. |
| Ethics posture | Encourages the lawyer to step out of the loop, which Rule 5.1 and 5.3 do not permit. | Maps to the supervision duty: you remain accountable for assisted work product. |
| What it frees you for | Headcount cuts, which is the wrong prize and the wrong measure. | The judgment, strategy, and counsel that clients actually pay a senior lawyer for. |
Read that as the whole argument in miniature. The vendor frame optimizes for removing people. The supervised frame optimizes for raising your output while keeping you responsible, which is the only version that survives a bar complaint or a malpractice question.
What to delegate and what to keep
The practical heart of this is one table. On the left, the tasks an AI paralegal can take off your plate, because each has a source you can check against and a bounded right answer. On the right, the tasks that stay human, because each requires weighing things a model cannot weigh and owning a result a model cannot own. The test that sorts them is whether the work is checkable in minutes or consequential beyond checking.
| Safe to delegate (you supervise) | Stays human (you own) |
|---|---|
| Intake summaries. Condense a long intake file or client questionnaire into a structured one page summary you read against the source. | Legal judgment. Deciding what a fact pattern means, what theory of the case to run, what risk is acceptable. Not delegable. |
| Document indexing. Catalog and tag a document set on a single matter so the team can find things, then spot check the index. | Case strategy. Sequencing, leverage, when to push and when to settle. This is the work clients hire a senior lawyer for. |
| Deposition digests. Turn a transcript into a topic indexed digest with page and line cites you verify against the transcript. | Client counsel. Reading the person, framing hard news, advising on a decision they have to live with. A human relationship, not an output. |
| Timeline building. Assemble a chronology from documents and correspondence, with each event sourced so you can confirm it. | The final certification. Signing your name to a filing, an opinion, or advice. You vouch for it; the model cannot vouch for anything. |
Notice the shape of the left column. Every delegated task produces an artifact you can lay next to a source and confirm in a few minutes: the summary against the file, the index against the documents, the digest against the transcript, the timeline against the record. That checkability is the whole permission slip. The moment a task stops being checkable against a source and starts requiring you to weigh, decide, or reassure, it has crossed into the right column and it stays there.
How to run supervised delegation, step by step
Here is the method as a workflow you can run today on an approved AI deployment. Each step is a delegated task paired with the supervision you do alongside it. The point is not speed for its own sake. It is reaching the same standard of work product with far less of your own toil, while staying squarely inside the supervision duty.
Step 1: Decide what is delegable before you delegate
Run every task through one filter first: can I check this output against a source in minutes, and does it require no judgment a model cannot make. If yes, delegate it. If no, it stays on your desk. This sorting step is itself a supervisory act and it takes seconds.
In practice: intake summary, yes. Document index, yes. "Should we file the motion," no. "What is our exposure here," no. When you are unsure, treat the task as human until you can articulate exactly how you would verify the output.
Step 2: Delegate the intake summary
Hand the model the intake file and ask for structure, not conclusions. You want a summary you can scan against the source, not an opinion about the matter.
Example prompt: "Summarize this client intake into a structured one page brief: parties, key dates, the client's stated goal, the documents referenced, and any open questions a lawyer would need answered. Do not assess the merits or recommend a course of action. Flag anything internally inconsistent."
Supervise: read the summary against the intake. A missed party or a wrong date tells you how carefully to check the rest, and the summary is now a verified starting point rather than an unread output.
Step 3: Delegate document indexing and the timeline
For a single matter's document set, ask for an index and a sourced chronology. This is patient, comprehensive work where a model genuinely outperforms a tired human reader, provided you confirm the result.
Example prompt: "Build an index of these documents: for each, give a short description, the date, the author and recipient, and a topic tag. Then build a chronology of events drawn only from these documents, and for every event cite the specific document it came from. Do not infer events that are not documented."
Supervise: spot check the index against the documents, and confirm each timeline event against its cited source. An AI paralegal doing this is indexing one matter for retrieval, which is a different task from e-discovery review across a production. For that distinct discipline, see our comparison of generative AI vs TAR for document review.
Step 4: Delegate the deposition digest
Turn a long transcript into a usable, cited digest. This is the kind of task that swallows an afternoon and produces something a model can draft in minutes for you to verify.
Example prompt: "Digest this deposition transcript into a topic indexed summary. For each topic, give a short synopsis of the testimony and the page and line citations that support it. Quote directly where the exact wording matters. Do not characterize the witness's credibility or draw legal conclusions."
Supervise: pull the cited page and line ranges and confirm the digest reflects the actual testimony. A digest with a misquoted line or a wrong cite is worse than no digest, so this verification is not optional.
Step 5: Keep the judgment, strategy, and counsel
With the legwork done and verified, your hours now go where they belong. Read the verified materials and make the calls only you can make: the theory of the case, the negotiation posture, the advice to the client. Do not ask the model to make these for you, and treat any strategic suggestion it volunteers as a prompt for your own thinking, never a decision.
Supervise yourself here too: the temptation, once the model has been reliable on the legwork, is to let it drift into judgment. Hold the line. The output may be fluent; the responsibility is still yours and is not delegable.
Honest usage notes
A few things become clear once you run this on real matters rather than a demo.
The indexing and timeline tasks are where an AI paralegal is most reliable and most freeing. Cataloging documents and building a sourced chronology are exactly the patient, comprehensive jobs where a model rarely tires and a human always does. Start your delegation here and the rest of the matter moves faster because the record is organized and verified.
The summary and digest tasks are strong but demand citation discipline. A model can produce a clean, confident deposition digest with a citation that points to the wrong page, or an intake summary that quietly drops the one inconsistent fact that matters. The output reads well regardless of whether it is right, which is the trap. Useful as a fast first draft, never as a finished work product you have not checked against the source.
The judgment tasks are where you stay most disciplined about not delegating at all. This connects to the larger picture: how a senior lawyer should structure AI across an entire practice, which our senior lawyer AI operating model lays out, and which sits inside the broader map of how law firms run on AI. Delegation is one room in that house, and it is the room where the supervision duty is most concrete.
The supervision duty, confidentiality, and verification
This is the part of the method you do not get to skip, and the part the "replace your paralegal" pitch quietly skips for you.
Rule 5.1 and 5.3 make this your responsibility
The ABA Model Rules on supervision of lawyers (5.1) and of nonlawyer assistance (5.3) are the right lens for AI. When you direct an AI paralegal, you are supervising assistance, and the work product is yours. Many bar authorities have read these duties to cover AI tools: you must understand the tool well enough to supervise it, and you remain responsible for the result. That is why the supervised frame is not a style choice. It is the rule. Treat every delegated output the way the rules require you to treat a nonlawyer's work: reviewed, corrected, and owned.
Never put privileged or confidential data in a public model
A client's intake, documents, transcripts, and matter file are confidential, and often privileged. Do not paste them into a free or consumer grade AI tool whose terms may use your inputs to train models or do not contractually protect the data. Use only an approved enterprise deployment that contractually commits not to train on your inputs and meets your firm's security requirements, or redact identifying and sensitive material first, or do not use AI on that matter at all. This maps directly to Rule 1.6 on confidentiality and Rule 1.1 on competence. Our confidentiality guide for attorneys walks through the deployment questions to settle before any client matter touches a model.
Verify every delegated output against its source
A delegated output is a draft, not a finding. Before you rely on a summary, an index, a digest, or a timeline, confirm it against the source it was built from. Models can misstate a date, miscite a transcript line, or drop an inconsistent fact while sounding completely sure. The duty to be right is yours and is not delegable, which is the whole reason the supervised frame exists.
How we built this method
This playbook reflects hands on use of leading general purpose models on the kind of structured legal support work senior lawyers actually delegate: intake summaries, document indexing, deposition digests, and timeline building. The delegate versus keep line is a practitioner judgment grounded in the ABA Model Rules, not a product and not a survey. The Leveraged Years just launched, and we do not publish invented statistics, testimonials, or enrollment counts we do not have. Where we describe what AI is good and bad at, we mean what holds up in repeated practical use as of June 2026, on materials that contained no real client confidential information. AI capabilities and vendor policies change quickly, so we date this guide and refresh it. None of this is legal advice, and none of it changes your professional duties of competence, confidentiality, supervision, and candor. Confirm any approach against your jurisdiction's rules and your firm's policy before using it on a live matter.
What this means for your week
You do not need to replace anyone, and you do not need a black box that invites you to stop watching. You need a delegation line you trust and apply every time. Delegate the bounded, checkable legwork to an AI paralegal you supervise: intake summaries, document indexing, deposition digests, timeline building. Keep the judgment, the strategy, and the client counsel, because those are not delegable and never will be. Verify every delegated output against its source, keep client confidential material out of public models, and remember that the work product is yours under the supervision rules no matter what helped produce it.
That is the whole premise of how we train senior lawyers to work with AI: not a thinner team and sloppier work, but the same standard of work product reached with far less of the toil, and the judgment hours returned to where clients actually value them. The Leveraged Attorney course is built to install this delegation discipline, and the supervision duty behind it, as a habit you can defend to a partner and a client alike.
Part of TLY's AI Workflows → workflow playbooks for senior professionals.
Frequently asked questions
What is an AI paralegal?
An AI paralegal is not a product category so much as a way of using a capable general purpose AI model: as a fast, tireless assistant you delegate structured legal support work to and then supervise. It can summarize intake, index documents on a matter, digest deposition transcripts, and build sourced timelines far faster than a person can. What it is not is a replacement for a lawyer's judgment or a license to stop supervising. The work product remains yours, and under the supervision rules you remain responsible for it.
What legal tasks can I safely delegate to AI?
Delegate the bounded, checkable tasks: intake summaries, document indexing on a single matter, deposition digests with page and line cites, and timeline building from the record. The common thread is that each produces an artifact you can lay next to a source and verify in minutes. Keep the unbounded, consequential tasks human: legal judgment, case strategy, client counsel, and the final certification you sign your name to. The test is simple. If you can verify it against a source quickly and it needs no judgment, delegate and check it. If it requires weighing risk or owning advice, it stays with you.
Does AI replace paralegals or lawyers?
No, and the framing of replacement is the mistake. AI can take over a large share of structured support work, which changes how that work gets done and frees senior time. It does not take over judgment, strategy, or client relationships, and it does not remove the lawyer from responsibility for the output. The supervised frame, an AI paralegal you direct and verify, is both the more useful model and the one the professional rules require, because work produced with AI assistance is still your work product.
What do ABA Rules 5.1 and 5.3 mean for using AI?
Model Rules 5.1 and 5.3 govern a lawyer's responsibility for the work of other lawyers and of nonlawyer assistants. The widely adopted reading is that directing an AI tool is a form of supervised assistance: you must understand the tool well enough to use it competently, and you remain responsible for the result. Practically, that means you cannot treat an AI output as finished and unreviewed. You review it, correct it, and own it, exactly as you would a junior's draft. That duty, alongside competence under Rule 1.1 and confidentiality under Rule 1.6, is what keeps AI use defensible.
Can I put client documents into an AI tool?
Only into an approved enterprise deployment that contractually commits not to train on your inputs and meets your firm's security requirements. Never paste privileged or client confidential material into a free or consumer grade AI tool whose terms may use your inputs to train models or do not protect the data. The safer paths are an enterprise deployment with the right contractual terms, redacting sensitive material first, or simply not using AI on that matter. This maps to your confidentiality duty under Rule 1.6. When you are unsure whether a deployment is safe, treat the material as off limits until you have confirmed it.
Build the delegation discipline, not just the opinion
Knowing the line between delegate and keep is the start. Applying it every time, with the supervision, citation, and confidentiality discipline baked in, is the skill that compounds. We teach the full delegation method, the prompts, and the guardrails as one repeatable system a senior lawyer can defend under Rule 5.1 and 5.3.
Start with Leveraged Attorney: the supervised AI delegation system for lawyers Join The Leverage Club for $49 and get the delegation checklists, prompts, and supervision templates Not sure where to start? Take the 2-minute course finderSources: ABA Model Rules of Professional Conduct on responsibilities of a supervisory lawyer (5.1), responsibilities regarding nonlawyer assistance (5.3), competence (1.1), and confidentiality of information (1.6); Anthropic Claude enterprise and commercial data usage policies (Anthropic, 2026); TLY hands on use of leading general purpose models on intake summaries, document indexing, deposition digests, and timeline building containing no real client confidential information (June 2026). Capabilities and vendor policies as published as of June 2026 and subject to change. This guide is not legal advice.