AI for paralegals: what to hand off, what to check and which rules apply

What can paralegals hand to AI today?

BLS lists gathering and organizing legal documents, drafting correspondence and legal documents, researching laws and regulations, and reviewing trial transcripts among paralegal duties ¹⁰. The table pairs each task with what the sources say AI can do and what to check. We did not test any tool.

Scroll the table sideways to see every column.

What should paralegals not hand to AI?

• Legal judgment. Florida says a lawyer may not delegate to generative AI any act that could constitute the practice of law or any function that requires the lawyer's personal judgment and participation ¹¹, and California says professional judgment cannot be delegated to AI ⁵. Ask your supervising attorney where your state draws the line.
• Confidential client information in unapproved tools. California says a lawyer must not enter confidential client information into a tool with material confidentiality or security risks without informed client consent ⁵, and North Carolina advises generally avoiding client-specific information in publicly available AI tools ⁷. Vendors differ: Thomson Reuters says it never uses customer data to train underlying models ¹³, while the Relativity page we read does not say whether customer data is used for training ¹². Read your firm's policy first.
• Billing for time the tool saved. Texas says a lawyer will likely be able to charge for the actual time spent, including checking the output, but may not charge hourly fees for time "saved" ⁶, and North Carolina says a lawyer cannot bill three hours for work that took one ⁷. If you record time, record what you spent, including the checking.

What do bar opinions and courts say about AI?

These sources are examples, not a national survey.

• Supervision. California's 2026 guidance says supervising lawyers should set clear AI policies and make sure lawyer and nonlawyer staff get periodic training ⁵. Texas compares relying on generative AI with relying on inexperienced or overconfident nonlawyer assistants ⁶. Texas also cites ABA Formal Opinion 512 (2024), which with a Florida opinion suggests lawyers obtain informed consent before using these tools; Texas itself says only that a lawyer should consider informing clients and may need to secure their consent ⁶. We could not open the ABA's own copy, so this is Texas's account of it.
• Court rules. The Northern District of Texas requires a brief prepared using generative AI to say so on its first page; a brief without the disclosure counts as certifying that no part was AI-prepared (LR 7.2(f), in the local rules effective September 1, 2026) ⁸. A judge there who had required a separate certificate no longer does, citing the rule ¹⁵. That is one court, and California says lawyers should check the relevant jurisdiction for disclosure rules or orders ⁵. Ask which apply before a filing is drafted.
• Sanctions. In Mata v. Avianca, the court wrote that nothing is inherently improper about using a reliable AI tool but that attorneys must ensure filings are accurate, and imposed a $5,000 penalty after non-existent opinions created by ChatGPT were submitted ⁹.
• How common. A public database of court and tribunal decisions worldwide listed 2,207 cases involving hallucinated AI content on 23 October 2026, 877 with a lawyer as the party and 2 with a paralegal; it counts decisions, not a rate ¹⁶.

Which AI tools do legal teams use?

Three vendors describe tools for paralegal-type work, each in its own published terms. We have not tested them, and none shows a price on the page we read.

• CoCounsel Legal (Thomson Reuters) describes research, analysis and drafting from one query and tabular analysis of up to 10,000 documents against 100 questions ¹³.
• Relativity aiR has aiR for Review (first-pass review, quality control, prioritized queues) and aiR for Case Strategy (witness summaries, deposition outlines, chronologies) ¹².
• LOIS (Filevine) describes AI that acts on matters, setting tasks, moving deadlines and generating documents, and says every action is traceable to source ¹⁷.

California says the more autonomy a system has, the more oversight the lawyer needs ⁵, so ask how any tool that acts on its own is supervised. NALA's author says paralegals and litigation support staff are likely the biggest users of AI and that their input belongs in tool selection ⁴. Colleagues in other functions face parallel choices; see AI tools for accountants and AI tools for recruiters.

How do you check AI output before it goes anywhere?

Start with the stakes: NALA's guide says anything headed to a court or a brief needs meticulous review, possibly more than once, while internal email can get a lighter pass ³. Then look for five red flags: a confident statement with no citation, a fact that is too broad or too specific, a quotation that is really a paraphrase, a suspiciously short list and a legal term blended from different jurisdictions ³. To validate, ask the tool for the exact transcript lines behind each point and read them, re-ask in simpler words to expose gaps, and search the source yourself for what the tool missed ³.

Will AI replace paralegals?

BLS counted 404,900 paralegal and legal assistant jobs in 2025, a median annual wage of $62,890 in May 2025 (the Handbook's pay data come from the OEWS survey, which does not cover self-employed workers, and the median is the wage at which half of workers earned more and half earned less) and a projected change of 0% from 2025 to 2035, or 1,100 fewer jobs ¹⁰. It says demand is expected to be limited by advances in technology, including AI, which it expects to make paralegals more efficient at research and document preparation ¹⁰. A projection covers the whole occupation, not one office.

How do you get started with AI as a paralegal?

The roadmap below follows the approach Torchlane publishes for its AI FDE training: four responsibilities, four skill areas and a seven-category assessment ^1,2. The legal examples are ours. Use a task your supervising attorney has approved.

1. Define the business requirement. Pick one recurring task you own, such as summarizing deposition transcripts. Talk to the people doing the work and write down the bottleneck, the decision it supports, the source documents and the constraints ¹, which in a law office include confidentiality, privilege and court deadlines.
2. Build and integrate the system. Design the workflow and put a model only where it adds value; keep calculations and fixed rules in ordinary code ¹. Let the model draft the summary and leave date calculations to your office's docketing system.
3. Test errors, access and unfamiliar cases. Write real evaluations, then try incomplete inputs, permissions and failures ¹. A transcript with missing pages is a good test.
4. Deploy, explain and improve. Help colleagues use it, document decisions and steps, monitor the result and respond when things change ¹.

Torchlane names four skill areas: business and product judgment, technical implementation, reliability and security, and communication and ownership ¹. Torchlane's page says an understanding of operations, clients or an industry is a useful starting point for the first ¹. The published assessment works as a checklist even on another route: a project must score at least 3 out of 4 in each of seven categories with no critical failure, and the categories include correctness (20%) and security, privacy and compliance (15%) ². Real personal data in a submission and being unable to explain your own work are critical failures ².

If you want that structure with review rather than going alone, Torchlane's AI FDE training is how we suggest getting started. Its pages describe 73 lessons, a choice of 3 guided builds and 3 to 6 projects drawn from 12 simulated client engagements, so the practice uses simulated clients rather than your firm's files, and the capstone is defended by answering three reviewer questions in one unedited recording within 72 hours ^1,2. The 90 days is a training goal at about nine hours a week, and access lasts 120 days ¹⁸. You can apply without coding experience, but you will learn to work with code. Torchlane says the program does not turn every beginner into an experienced software engineer in 90 days and does not guarantee employment, clients, promotion or pay ^1,18. The pages we read list no legal-specific lessons.

We have a business connection with Torchlane.

Where can a paralegal learn AI skills?

Three options in each provider's own published terms. Our top pick is the first row; the rest are not ranked.

Scroll the table sideways to see every column.

* Our top pick. We have a business connection with Torchlane.

• You want a verification habit for the tools your firm already uses. NALA's webinar is the closest fit ¹⁹.
• You want plain-language grounding in what AI can and cannot do. AI For Everyone fits, though it is not legal-specific ²⁰.
• You want to build and evaluate AI-assisted workflows around your legal expertise, with review. Our top pick is Torchlane's AI FDE training. Its page says your professional experience helps you understand the business problems and that the program builds the technical capability around it ¹⁸. It takes more time than a webinar, and if you only need to use your firm's tools carefully, the NALA webinar or AI For Everyone may fit more closely.

We have a business connection with Torchlane.