AI for paralegals: four legal tools priced against $20 ChatGPT, and the consent rule that comes first

Can paralegals use ChatGPT for client work?

Yes for work that carries no client information, and for client work only after the supervising attorney has cleared the tool and, where the tool learns from what you type, the client has agreed. That is the line ABA Formal Opinion 512 drew on 29 July 2024: "a client's informed consent is required prior to inputting information relating to the representation into such a GAI tool." The opinion speaks to lawyers, and the lawyer answers for a paralegal's AI use under Model Rule 5.3(b), which the opinion reads as a duty to make "reasonable efforts to ensure that" the nonlawyer's conduct meets the lawyer's obligations. So the first question about any tool is its training setting, and the second is its price. We checked both against the vendors' own pages on 27 September 2026, in US dollars.

One of the four legal tools publishes a number. The rest sell to the firm, which is the honest signal about who they are for: a solo or a five-lawyer office buys Clio or Paxton, a national firm negotiates CoCounsel or Harvey by the seat. The three general rows, Whizi included, connect to no Westlaw or Lexis content, manage no matter, run no conflicts check and keep no audit log a managing partner can pull. They are reading and drafting tools that sit beside the research platform, never in its chair.

The training column is where consumer accounts fail Opinion 512 before anyone reads a word of output. OpenAI's help center, read 27 September 2026, says of ChatGPT for individuals that "we may use your content to train our models", with the opt-out under Settings, Data controls; the same page says Business, Enterprise and Edu are not used by default. The opinion also says that "merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient." The opinion places that consent with the lawyer, so the paralegal's job is to ask which tools are cleared and on which setting before the first upload. What never to paste into an AI chat is the list for the days when no answer has come back yet.

Can AI be trusted to check citations?

No: every citation an AI produces is a lead until you have opened the case, and the courts have been pricing that rule since 2023. Damien Charlotin's public database of decisions in which a court caught invented or misquoted authority listed 2,086 entries on 27 September 2026, 1,430 of them from US courts, and 828 of them filed by lawyers rather than self-represented litigants.

Three of those decisions are the receipts a paralegal should be able to name. Mata v. Avianca (S.D.N.Y., 22 June 2023): Judge Castel fined two lawyers and their firm $5,000 jointly for a brief built on six invented decisions from ChatGPT, and wrote that "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings." Johnson v. Dunn (N.D. Ala., 23 July 2025): Judge Manasco disqualified three Butler Snow attorneys from the case and referred them to the Alabama State Bar, after one used ChatGPT for citations "without verifying their accuracy", at a firm whose 2023 policy already required written permission and "full checks of the accuracy of any results". Noland v. Land of the Free (Cal. Ct. App., 12 September 2025): a $10,000 sanction, California's first published opinion on the subject, where "21 of the 23 quotations" in the opening brief were fabricated and counsel, in the court's words, "did not read the cases the AI tools cited".

The paid research platforms narrow the problem without closing it. Stanford's preregistered evaluation of 202 queries, released June 2024 and published in the Journal of Empirical Legal Studies in 2025, found that Lexis+ AI and Thomson Reuters's Westlaw AI-Assisted Research and Ask Practical Law AI "each hallucinate between 17% and 33% of the time", against a higher rate for plain GPT-4. Opinion 512 cites the same study in its footnotes. So the cite-check is the same four steps whatever produced the cite: open the case in Westlaw, Lexis, PACER or CourtListener; find the quoted words at the pin cite; confirm the case says what the brief says it says; run KeyCite or Shepard's for later history. How to fact check AI answers covers the general version, including the ten second test for an invented citation.

How do paralegals summarize depositions and discovery with AI?

Upload the transcript, ask for a page-line summary in a fixed shape, then spot-check a sample against the transcript before the attorney sees it. Opinion 512 blesses exactly that method: a lawyer relying on a tool to summarize many contracts "would not necessarily have to manually review the entire set of documents to verify the results" after testing the tool on a smaller subset and finding the summaries accurate. The prompt we use, for a cleared tool and a text-searchable PDF:

"You are preparing a deposition summary for a supervising attorney. Use only the attached transcript. Produce a page-line summary as a table with four columns: Page:Line range, What the witness said in neutral words, Exhibits referenced, Follow-up flag. Quote the witness verbatim only where the exact wording matters, and give the page and line for every quote. If something is not in the transcript, write not in transcript. Do not add legal conclusions, credibility opinions or facts from outside the document."

Two things break this in practice. Scanned exhibits with no text layer come back as guesses unless the tool transcribes the images first; a text-searchable PDF is the input, and a 300 page transcript fits: OpenAI's own pricing page puts the reasoning input on ChatGPT Plus at about 320 pages of text, and the current GPT, Claude and Gemini flagships all carry a 1M token window. And a summary that drops a page range is worse than one that admits a gap, which is why the prompt forces "not in transcript". For discovery sets, run the same prompt per document with a fifth column, Bates range, and mark every AI-made summary as such in the file, the practice Opinion 512 suggests in its training footnote so a later reader knows what to trust. Summarize a PDF with AI has the length and format limits by model.

Timelines and correspondence: the two drafts an attorney accepts from a paralegal

A chronology with a source pin on every row and a first draft of routine correspondence are the two AI outputs an attorney can check line by line against something in the file. The chronology prompt: "From the attached documents only, build a chronology as a table: Date (ISO), Event in one sentence, Source document and page, Confidence (stated, inferred, or unclear). Order by date. Where two documents conflict, list both rows and flag the conflict. Do not fill gaps from general knowledge." The Confidence column is the whole trick: it makes the model show where it is inferring, which is where a paralegal reads first.

Correspondence is safer still: client status letters, records requests, scheduling emails and deadline reminders drafted from the matter facts you paste in, then edited to the attorney's voice. Keep the billing entry honest. Opinion 512 gives the worked example: "If a lawyer uses a GAI tool to draft a pleading and expends 15 minutes to input the relevant information into the GAI program, the lawyer may charge for the 15 minutes as well as for the time the lawyer expends to review the resulting draft for accuracy and completeness." The hour the draft used to take is not billable once it took fifteen minutes.

Where this runs matters as much as the prompt. In Whizi, which publishes this page, a Project pins up to 10 files of up to 10 MB each (PDF, .docx, .txt, .csv) and up to 32,000 characters of standing instructions, so the complaint, the scheduling order and the firm's style sheet ride along on every prompt in that matter, and the same thread can be handed from GPT to Claude to Gemini for a second read; ordinary chat attachments expire after up to 30 days. Whizi has no matter management and none of Westlaw's or Lexis's content, so it sits beside the research platform, never in place of it.

Can AI format Bluebook citations?

Yes for the form of a citation you have already verified, and no for the citation itself. Given a real case with its reporter, volume, page and year, current models apply the 22nd edition's rules (reporter abbreviations, pincites, Id. and short forms, Bluepages court-document style) well enough to save the typing, and the $49 print book or the Bluebook Online settles any disagreement between the model and the rule. Never let the model supply the authority: that is how Noland's 21 fabricated quotations happened.

The formatting prompt: "Format the following citations in Bluebook 22nd edition form for a court document (Bluepages). Do not change any party name, reporter, volume, page, pincite or year. If an element is missing, write [MISSING] in its place instead of guessing. Return one citation per line, then a second list of every change you made." The second list is the audit trail; a model that silently corrected a volume number has created a new cite to check. State courts with their own style manual and a judge's standing order both override the Bluebook, so the local rules are the last read before filing.

Will AI replace paralegals?

Not on the government's numbers: the Bureau of Labor Statistics projects "little or no change" in paralegal and legal assistant employment from 2025 to 2035, from 404,900 jobs in 2025, with about 40,800 openings a year as people retire or move, and a median wage of $62,890 in May 2025. BLS names AI as the reason demand will not grow, because it makes paralegals faster at research and document preparation, which is a description of the work moving, not ending.

The tools are already in the building. Clio's Legal Trends Report, released 7 October 2024, found 79% of legal professionals using AI in their work, up from 19% in 2023, and the same report put 69% of hourly billable paralegal work in the range AI could automate. The 31% it cannot is the part every case in the chart above turned on: someone who opens the cited case and notices that two documents tell different stories. The paralegal AI cannot replace is the one who reads the case.