Best Legal AI for Criminal Law: A Small Defence Firm's Shortlist

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Best Legal AI for Criminal Law: A Small Defence Firm's Shortlist.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Best Legal AI for Criminal Law: A Small Defence Firm's Shortlist.

A small defence practice needs four tools for four jobs: a defence-built evidence transcriber such as JusticeText for bodycam and jail-call review, a research assistant such as CoCounsel or Lexis+ with Protégé, a business-plan assistant for drafting, and an AI intake line such as Smith.ai. Get each vendor's data answers in writing before uploading discovery.

Criminal work raises the stakes on every one of those choices. Your client may be in custody, the material may name victims and witnesses, and the discovery may arrive under restrictions that no vendor's privacy page knows about. Choose by job, test on closed files, and let the data answers decide what stays on the shortlist.

Follow me on Instagram@sagnikteaches

Which tool does which job in a defence practice

Prices below are USD list prices from the vendors' own pages where they publish them. Where a vendor sells only by quote, the table says so, because an unpublished price is a fact worth knowing when you budget.

Connect on LinkedInSagnik Bhattacharya
JobTool category and examplesPrice evidenceTest it on
Bodycam, jail calls, interviewsDefence-built evidence review: JusticeText. General transcribers: OtterJusticeText: demo or access request, no public price. Otter Pro $8.33 a month annual, $16.99 monthlyYour noisiest clip
Document discovery and case filePractice management with OCR: MyCase Advanced$130 a user annual, $150 monthly; 5,000 OCR pages a user each month, shared; extra 5,000 pages $30A scanned disclosure bundle
Legal researchCoCounsel (Thomson Reuters), Lexis+ with ProtégéCoCounsel Legal, with Westlaw research, $784 a lawyer a month on a one-year term (widest coverage); CoCounsel Essentials, without research, $360. Lexis by quoteFive questions you already know the answers to
DraftingChatGPT Business, Claude Team, or the assistant in your practice system$20 a seat annual, $25 monthly; two seats minimumA closed-case motion
IntakeAI answering: Smith.ai. Forms and CRM: MyCase ProSmith.ai AI receptionist: free for 25 calls, Pro $150 a month for 75 calls. MyCase Pro $100 annual, $120 monthlyTen recorded calls from your past month

Notice that the two jobs that handle the most sensitive material, evidence review and research, are also the ones whose prices sit behind a demo, a questionnaire or a sales call. That is the pattern across legal AI: the vendors who see your most sensitive files want a conversation first. Use that conversation to collect written answers to the data questions every vendor should face.

Subscribe on YouTube@sagnikteaches

Transcripts: bodycam, jail calls and client interviews

The problem in a defence office is volume. A single case can carry hours of bodycam, dozens of jail calls and recorded interviews, and most of it matters for a few minutes. JusticeText describes its product as turning footage into searchable, time-stamped transcripts in more than 100 languages, building case timelines, flagging inconsistencies, and letting you redact, clip and sync video for trial. It is the only tool on this shortlist built around that exact workflow. It publishes no price, so the first step is an access request and a demo.

General transcribers are cheaper and fine for client meetings and dictated notes. Otter Pro is $8.33 a month on annual billing ($16.99 monthly) with 1,200 minutes a month and a 90-minute limit per meeting, which suits an interview recorded with consent but not a two-hour traffic-stop video. Whichever you choose, the rule is the same: a transcript is an index, not evidence. You verify every quotation against the audio before it appears in a motion or a cross-examination outline.

A ten-minute test that exposes weak transcription

Pick the worst ten minutes you have: wind, overlapping voices, radio traffic, a frightened speaker. Transcribe it with each candidate, then listen once with the transcript in front of you and log every error that changes meaning.

TimestampWhat the tool wrote (illustrative)What was saidEffect
04:12"I did have anything on me""I didn't have anything on me"Dropped negative flips the meaning
04:31"Speaker 1: step out of the car"Two officers speaking at onceWrong speaker; matters for who gave which command
06:50[inaudible]Quiet remark by a passengerMissed potential witness comment
07:15"ten fifty""ten-fifteen" (a radio code)Code misheard; verify against the radio log

Score each tool by the number of meaning-changing errors per ten minutes, not by overall percentage accuracy. A tool with 97 per cent accuracy that drops one negative has done more damage than a tool at 92 per cent that flags its uncertain passages. Prefer the tool that marks low-confidence text and lets you click to the audio.

Illustrative saving: two hours of bodycam across three clips would take about five hours by hand, with note-taking at two and a half times real time. With a searchable transcript, you read and search in about an hour, then check fifteen flagged moments at three minutes each, roughly an hour and forty-five minutes in all. The assumption is yours to replace with your own timings, but the shape holds: the saving comes from checking fewer moments, not from skipping the check. For the evidence-review process end to end, AI discovery review for defence firms covers indexing, search and cross-checking, and summarising bundles and transcripts safely covers the summary step.

The Complete Excel Course with AI Integration thumbnail

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Documents and the case file: OCR before AI

Paper disclosure still arrives as scanned bundles. MyCase lists an OCR assistant for discovery on its Advanced tier at 5,000 pages per user a month, shared across the team, with extra blocks of 5,000 pages at $30 for the firm. For a two-lawyer practice that is 10,000 pages a month in the base allowance. Before you buy it, count your typical disclosure: if one serious case brings 3,000 pages, a pair of cases a month fits comfortably.

The same tier lists a case assistant, and lower tiers list writing and document assistants that extract and summarise key details. The point of OCR is searchability: once a 1,200-page bundle is text, you can search it for a witness name in a second. Use the summarising features for orientation and the search for precision. Never rely on a summary for an inconsistency you plan to put to a witness; read the page.

Research: what to test before paying a research vendor

Research is where hallucination does the most harm in court, because a plausible citation looks like diligence. Both leading vendors present grounded answers with links to primary sources. CoCounsel is Thomson Reuters' assistant; Lexis+ with Protégé is the new name for Lexis+ AI. Thomson Reuters shows online pricing for firms of up to ten lawyers through a short questionnaire on its sales site. For one lawyer, CoCounsel Legal, which draws on Westlaw and Practical Law, showed $784 a month on a one-year term with the widest coverage selected ($637 a month on three years), while CoCounsel Essentials, which covers document analysis, review and drafting without the research database, lists at $360. Lexis quotes on request. Both are sold per lawyer and priced by firm size, so get the figure for your own headcount and coverage before you compare.

Test them with five research questions you have already answered by hand in the last year, covering suppression, disclosure, sentencing and a procedural point. Score each answer on four lines: did it find the authority you found, did it find anything you missed, did any citation fail to resolve, and did it state the rule for the right jurisdiction. The comparison in CoCounsel or Lexis with Protégé goes through the product differences.

A research prompt, a sample answer and the fix

I act for a defendant stopped for a broken tail light. The officer extended the stop to run a dog sniff after the licence check was complete. In [jurisdiction], what is the test for whether a stop was unlawfully prolonged, and which recent appellate decisions apply it? List each authority with a pinpoint and say how confident you are.

Illustrative answer shape: a short statement of the test, three authorities with pinpoint references, and a caution that facts about the length of the extension are decisive. What you fix: pull each authority and read the passage, check that the "recent" decision has not been reversed or distinguished, and add the local rule or practice direction the answer left out. If any authority fails to resolve, delete it and note which proposition it supported. A single unresolved citation is a warning about the rest. The routine in how to stop AI inventing case law turns this into a checklist.

Drafting: motions, letters and client explanations

Drafting is the safest first job for AI in a defence practice, provided it never makes a decision. A general assistant on a business plan handles the structure of a motion, the tone of a letter, and the plain-language explanation a nervous client needs. ChatGPT Business and Claude Team both list at $20 a seat on annual billing ($25 monthly) with a two-seat minimum, and both state that they do not train on your content by default. Use redacted facts anyway: replace names with Client, Witness A and Officer B.

Before and after: a client letter at a readable level

Before: "Pursuant to the scheduling order, the matter is listed for a pre-trial review at which the prosecution will indicate its position on disclosure and any outstanding applications." After, drafted with an assistant and checked by you: "Your next court date is a short hearing on 18 March. You do not need to speak. We will find out whether the prosecution has sent everything it should, and I will tell you straight afterwards what it means for your case." The second version is better because it tells the client what to expect and what to do. What you fix in the draft: the date, the statement that the client need not speak (confirm it applies), and any line that predicts an outcome.

A drafting prompt that keeps the lawyer in charge

Draft the facts section of a bail application. Use only the facts below and mark any gap with [CONFIRM]. Do not argue the law and do not add facts. Facts: Client A is 34, has lived at the same address for six years, works full-time, has two prior convictions, both more than eight years old, and has surrendered a passport.

The marker approach matters. The assistant will otherwise fill gaps with plausible detail, such as a reference from an employer you never had. Search the output for every [CONFIRM] before filing, and read the rest as though a junior wrote it and wants to impress you. For longer filings, which legal tasks should never go to AI draws the line between drafting help and judgement calls.

Intake: an AI on the phone when a defendant or family calls

Intake matters more in criminal work than in most practices, because the first call is often urgent. Smith.ai's AI receptionist answers 24 hours a day, records and transcribes calls, qualifies leads, routes calls, books through scheduling tools and connects to Clio and Zapier. Its free plan covers 25 calls a month, with $3.00 a call after that; Pro is $150 a month for 75 calls ($2.50 for each call beyond them), with 150-call and 300-call tiers at $270 and $500. Its human-first plan, with live receptionists, starts at $300 a month for 30 calls. MyCase Pro adds intake forms and a CRM for $100 a user on annual billing if you want forms rather than calls.

The risks are specific. A caller may be in custody, a family member may be calling on someone's behalf, and an arraignment may be hours away. Set the rules before the line goes live.

  • Never advise. The assistant takes details and routes. It does not say whether to plead, whether to speak to police or whether a charge is serious.
  • Escalate on trigger words. "Arrested", "in custody", "court tomorrow", "police at the door" and "bail" go straight to a phone number you answer.
  • Check conflicts before substance. Capture names of parties and co-defendants first, so you can screen a conflict before anyone describes the facts.
  • Handle recording consent. Say at the start that the call is recorded and by what kind of assistant, and check what your jurisdiction requires.
  • Store with care. Intake recordings may be privileged. Treat them as client files, with the same retention and deletion rules.

An illustrative intake summary from the assistant: "Caller: sister of the accused. Accused: [name], charged yesterday, bail hearing tomorrow 9:30. Charge: told 'assault'. In custody: yes. Conflicts: co-accused named, [name]. Next step: callback requested urgently." What you fix: confirm the charge from the court list rather than the caller's words, and note that the sister's call is not a retainer. For the cost and safe handling of voice agents in more depth, AI voice agents for lawyers goes through scripts and limits.

What to ask each vendor about data

Ask every vendor the same questions, in writing, and keep the answers with the contract. In a defence practice a few of these matter more than they do elsewhere.

QuestionWhy it matters for defence workA good answer
Is my content used to train any model, by you or your model providers?Discovery includes other people's names, addresses and statementsNo, stated in the contract for my plan
Where is it stored and for how long, including backups and logs?You may owe deletion at the end of a caseNamed storage and a retention period I control
Which sub-processors and model providers see my content?Each one is another party holding client materialA published list with change notice
Can your staff read my files, and when?Privilege and protective ordersOnly for support, with my approval and a log
If a court or law enforcement asks you for my data, what do you do?A defence firm's files are exactly what the other side might wantChallenge where possible and tell me first
Can I delete a matter completely and export it?Case closure and file transfersSelf-service deletion with a confirmation
What security evidence do you hold, and for which product?Marketing claims differ from auditsA current report or certificate, scoped to my product

The fifth question is the one most buyers never ask. Ask it plainly, and write down the answer. If the vendor cannot say what it does when asked for your data, that tells you something about the contract.

Check the terms the discovery came under

Before any vendor question matters, read the order or agreement the material arrived under. Many restrict copying, sharing with third parties and use outside the case. If the wording is unclear, ask the court or the other side, in writing, whether uploading to a named vendor under a business-plan contract is permitted. If the answer is no or maybe, keep that material in your own systems and use AI only on your own notes and redacted summaries.

A priced shortlist for three practice sizes

These totals add only the tools with public prices, on annual billing unless stated. Evidence review and research are quoted separately because the vendors do not publish figures.

PracticeIncluded in the sumPer month, annualPer month, monthly billing
Solo defence lawyerChatGPT Business, 2 seats ($40); MyCase Advanced, 1 user ($130); Smith.ai free tier$170$200
Two lawyersChatGPT Business, 2 seats ($40); MyCase Advanced, 2 users ($260); Smith.ai Pro, 75 calls ($150)$450$500
Five people (lawyers, paralegal, investigator)ChatGPT Business, 5 seats ($100); MyCase Advanced, 5 users ($650); Smith.ai Pro, 150 calls ($270)$1,020$1,145

Add a quote for evidence review and one for research, and treat those two as the real budget decisions. A practice that handles video-heavy cases may find the evidence tool is the first purchase; a practice that handles complex appeals may start with research.

A thirty-day pilot with a score sheet

Run one pilot per tool on a closed matter, with a written start and end. Week one: gather the vendor's data answers and the discovery terms, and stop if either fails. Week two: run the test on closed material and log minutes and errors. Week three: involve one other person and ask what they would not trust. Week four: score the tool and decide.

Score each tool out of 20: five points for accuracy on the test, five for minutes saved against your hand timing, five for the quality of the vendor's data answers, five for how well it fits your workflow without extra steps. Buy at 15 or above, trial again at 10 to 14, and drop below 10. Record the decision date. If you would prefer to run that process with another pair of eyes, that is what my AI implementation consultation covers; get in touch when you are ready.

Defence-practice questions before you buy

Can I upload discovery into ChatGPT or Claude?

Only after you have read how the material reached you. Discovery often arrives under a protective order, an undertaking or a confidentiality agreement that limits who may see it. A business plan does not train on your content by default, but it does not override those limits. Ask the court or the other side if the wording is unclear.

Should an AI tool ever advise a client on a plea?

No. Plea and strategy advice is a decision you make from the evidence, the client's instructions and your experience. Use AI to explain a decision you have already made in plain language, to organise evidence and to draft documents you will check line by line.

Do I have to tell the court I used AI?

Some courts now require a statement about AI use or a certificate that citations were verified, and rules differ by court and by division. Check the current practice directions or local rules for each court you appear in, and keep a note of how every citation was checked.

Is a general transcriber like Otter good enough for bodycam footage?

Test it on your worst clip before trusting it. Crosstalk, wind noise, radio chatter and overlapping speech are where general tools fail and where defence work lives. A tool built for evidence review adds clipping, redaction and exhibit features, which general note-takers lack.

Further reads

Sources: JusticeText, MyCase, Thomson Reuters CoCounsel, LexisNexis, Smith.ai, Otter, ChatGPT Business and Claude Team vendor pages (checked 1 November 2026).

Want a defence-firm AI set-up you can defend?

On a 1:1 call we'll map your discovery, drafting and intake steps, choose the tools worth testing first, and write the vendor questions and handling rules around client material.

Book a 1:1 call with me