The Real Stack: What Firms Actually Run in 2025

Ask ten defenders what software their firm runs and you'll get ten different answers — but the pattern underneath is clear. The modern stack is a mix of heavyweight legal AI, old-guard research platforms, and a growing dose of generic chatbots that were never built for evidence work.

The named tools in active legal use right now include Harvey, CoCounsel by Thomson Reuters, Lexis+ AI, Westlaw with AI-Assisted Research, Ask Practical Law AI, and Microsoft Copilot, which is often recommended over consumer tools for confidentiality reasons. CoCounsel in particular has spread through the defense side: all federal public defenders, a growing number of state defender offices, and even federal courts use it. On the evidence side, JusticeText — built by ex-Google engineers — is used by public defenders across 20 states for bodycam and interrogation review, though those adoption figures are self-reported by the vendor.

The adoption numbers look strong at first glance. According to the 2025 Clio Legal Trends Report, 79% of legal professionals now use AI, and 82% plan to increase usage over the next 12 months. Large firms lead at 87% adoption; solo firms trail at 71%.

But the tool mix is shifting in a direction that should concern anyone doing evidence work. Only 40% of legal professionals use legal-specific AI solutions — down from 58% in 2024 — as many drift toward free generic tools like ChatGPT, Gemini, and Claude. The same report notes that legal-specific AI offers better case law knowledge, stronger data privacy, and more accurate results, yet the free tier is hard to beat.

Here's why that drift matters for criminal defense. Generic chat tools can't diarize a chaotic bodycam scene, redact a minor's face from video, sync dashcam with interview-room footage, or cite the exact second of a jail call. And the accuracy problems are documented: an academic study of 176 bodycam videos found AI transcription identified an average of 3.3 participants per transcript versus 4.8 in human-edited versions — missing speakers and roughly 25,000 words across the sample. A Stanford study found legal research tools hallucinate at alarming rates too, with Westlaw's AI-Assisted Research producing incorrect information more than 34% of the time.

The practical stack in 2025 looks like this:

• Research and drafting: Westlaw, Lexis+ AI, CoCounsel, or Harvey — with mandatory verification of every citation.
• General productivity: Microsoft Copilot or ChatGPT for summaries and admin work, ideally under a firm AI policy (53% of firms have none, or attorneys don't know about it).
• Evidence review: dedicated tools like JusticeText for transcription, or a grounded workspace like Evidence Copilot that pairs timestamped, diarized transcripts with clickable citations into the exact video moment or PDF page — every AI output marked pending human review.

The tension is real: adoption is broad, but the tools most defenders actually reach for weren't built for the discovery dump on their desk. As the NACDL argues, catching "the contradiction buried in 10 hours of bodycam footage" takes more than a chat window.

Why Generic AI Breaks on Evidence: The Documented Failure Modes

If you're trusting a generic AI to review your discovery, the independent research says you shouldn't — not because the tools are bad, but because the failure modes land exactly where criminal defense is most exposed.

Start with the legal research tools defenders already know. A Stanford RegLab/HAI study found that Westlaw's CoCounsel-powered AI-Assisted Research hallucinated more than 34% of the time, while Lexis+ AI and Ask Practical Law AI produced incorrect information over 17% of the time, according to Courthouse News. These aren't consumer chatbots — they're the enterprise products most firms would consider the "safe" choice.

Then look at transcription. Researchers at Northeastern analyzed 176 bodycam videos — roughly 23 hours across 73 incidents — and found AI transcription got most of the text right but failed badly at speaker attribution. The AI identified an average of 3.3 participants per transcript versus 4.8 in human-edited versions, maxed out at 7 speakers where humans identified up to 16, and dropped about 25,000 words across the sample, per the Northeastern study. The cautionary example: a Heber City, Utah police report drafted by AI transcription incorporated dialogue from "The Princess and the Frog" playing at the scene. Kansas City PD subsequently stopped using Rev.ai for bodycam analysis.

For a defender, that 2% isn't trivia. As Calli Schroeder of EPIC put it, a transcript that's 98% accurate can still "completely change who is at fault." When your cross-exam hinges on who said what during a custodial interrogation, a machine that can't keep speakers straight is a liability, not an assistant.

This is why "verify everything" has moved from preference to ethics. The NACDL's "Parity in Practice" report proposes a risk-tiered framework: summaries and admin documents are low-risk; draft motions are medium-risk with verification; case-specific factual analysis and strategy are high-risk, where AI should support — never supplant — human judgment. A proposed Federal Rule of Evidence 707 would subject AI-generated evidence lacking a human expert to a reliability test.

• Low-risk: summaries, administrative documents
• Medium-risk: draft motions, with mandatory verification
• High-risk: privileged information, case-specific factual analysis, strategy — human judgment only

The workflow implication is concrete. Any AI output touching your evidence needs two things: a citation you can check against the source, and a human review before it becomes part of your record. That's the design premise behind Evidence Copilot — every AI answer carries a clickable citation into the exact second of video or page of the PDF, and every output is marked AI-generated and pending human review. The tools that survived the research above are the ones built to be checked, not trusted.

The Evidence-Review Layer: What Your Stack Is Missing

Ask ten defenders what software they use and you'll hear Westlaw, Clio, maybe CoCounsel. Then ask how they review forty hours of bodycam, three jail calls, and a scanned police report, and the answer goes quiet — because the research and memo tools in most stacks simply don't touch that work.

The gap is real and measurable. Public defenders receive hundreds of hours of digital discovery each year, and the vast majority of the millions of hours of police bodycam footage generated nationwide is never watched by a human, according to NPR reporting. Meanwhile, the AI tools defenders are adopting for research are built for case law, not evidence media.

What does an evidence workspace need that a generic chat window can't do?

• Diarized, word-level transcripts — not just text. Researchers at Northeastern found AI transcription of bodycam identified an average of 3.3 participants per transcript versus 4.8 in human-edited versions, with about 25,000 fewer words across the sample. Speaker attribution, not raw accuracy, is where transcription breaks.
• OCR on scanned PDFs — most police reports arrive as image-only scans with no text layer. Without OCR, they're invisible to search and to any AI tool.
• Citations to the exact video second — a Stanford study found Westlaw's AI-Assisted Research hallucinated more than 34% of the time. An answer you can't verify against the source moment is worse than no answer.
• Redaction with immutable originals — discovery turnover requires blurring bystander faces and muting PII without destroying the underlying file or the audit trail.
• Multi-camera sync — bodycam, dashcam, and interview-room footage of one event need to play against a single clock, or you're reconstructing the timeline by hand.

This is where Evidence Copilot sits in the stack: a grounded evidence-review workspace built for criminal defense. Upload the discovery dump, get timestamped diarized transcripts and OCR'd document text, and ask questions — every answer cites the exact second of video or page of PDF, marked AI-generated and pending human review. Redactions create derived files; originals are never touched.

Pricing is published, not behind a sales call: $39/user/mo Solo, $99/user/mo Firm, and $750/attorney/yr for PD offices — roughly 38% under JusticeText's ~$1,200/attorney/yr. The NACDL argues defenders have an ethical obligation to adopt AI responsibly as prosecutors adopt enterprise tools, and that case-specific factual analysis must remain human judgment. An evidence layer built around verifiable citations is how you do that work faster without outsourcing judgment to a model.

Building Your Stack: A Practical Workflow for Discovery Review

The discovery dump lands in your inbox as a zip file with names like X13qrt9__x3455, and the question is always the same: how do you get through all of it before the preliminary hearing? A structured workflow beats heroic all-nighters — and it's why the NACDL now frames responsible AI adoption as an ethical duty of competence for defenders facing prosecutors with enterprise tools.

Step 1: Intake everything. Get the dump out of email and into one case workspace. Evidence Copilot gives each case its own inbound address, so forwarded discovery — bodycam files, interrogation footage, jail calls, 911 audio, scanned police reports — lands as evidence automatically. Nothing gets lost in a forwarded thread three weeks before trial.

Step 2: Transcribe and OCR on ingest. Audio and video become diarized transcripts with word-level timestamps; scanned reports and photo evidence become searchable text. Then do the speaker QC the machine can't. Northeastern research on 176 bodycam videos found AI identified an average of 3.3 participants per transcript versus 4.8 in human-edited versions — so rename speakers early, and verify who said what before anything downstream relies on it.

Step 3: Run the playbooks. With a clean transcript base, work the case:

• Suppression Motion Prep — a consent/search-scope/stop-duration/Miranda sweep that drafts a motion outline with both sides quoted and cited.
• Miranda Timeline — a chronological custodial-interrogation timeline built from the actual footage.
• Review signals — auto-flagged report-vs-bodycam and statement-vs-call-audio inconsistencies, surfaced before anyone asks.
• Cross-exam drafts — suggested questions only, each tied to a quoted source, saved as attorney-review notes.

Step 4: Verify before you file. Treat every AI output as a draft pending attorney review. That's not hedging — it matches the NACDL's risk-tiered framework, which places case-specific factual analysis in the high-risk tier where AI supports rather than supplants human judgment. The Stanford-backed finding that legal research tools hallucinate over 34% of the time is the reason to demand clickable citations into the exact video second or PDF page, and to click through them yourself.

Step 5: Export. Build your exhibit pack — clips, still frames, quoted excerpts, and a citation manifest — so what you argued in the motion traces back to the source file. If a tool can't show you where every claim came from, it isn't evidence review; it's a summary with a confidence problem.