See how AI medical-record review links every fact to the exact Bates page that proves it — click any citation and jump straight to the record.
See the 60-second demo →I've been on the expert-witness side of depositions for thirteen years. The best depositions I've watched were the ones where the examining attorney noticed something in minute thirty that didn't land until minute ninety — a nonresponsive answer, a phrase that contradicted what was said forty pages earlier, a claim that couldn't have come from the record. The attorneys who could do that in real time were the ones who won motions in limine and closing arguments.
Most attorneys can't. Not because they aren't skilled, but because the human brain can't hold a 120-minute stream of testimony in working memory while simultaneously reading a chart, drafting follow-up questions, and watching the witness for tells.
That's what AI is genuinely good at — not replacing attorney judgment, but surfacing the details the attorney's conscious mind misses while focusing on the next question.
We built Courtroom AI to do exactly this, and last night I ran it against a real 2-hour cardiothoracic surgeon deposition (publicly available on Miller & Zois's channel) to stress-test what it actually catches. Here's what I learned.
Courtroom AI does three things simultaneously during a live deposition:
The whole round-trip (transcribe → batch → analyze → render) takes 5-15 seconds per chunk. Fast enough that by the time you've finished one question, the analysis of the prior answer is on your screen.
Segment from the session:
"Now like I said, we have a team effort. We have a surgeon and cardiologist perform that procedure, and that's relatively common for that to happen."
That's a witness answering a question about whether he participated in a specific patient's surgery. Most lawyers would keep moving. Here's what the four lenses caught:
"Testimony lacks procedural specificity and contains vague generalizations about surgical practice that cannot be fully evaluated without procedure identification. Procedure specification unclear — commonality of two-specialty involvement depends on specific procedure type; some procedures are typically single-specialty."
Translation: the witness made a claim ("relatively common") without naming which procedure. Depending on the procedure, "common" means different things. The AI flagged this as a medium-severity inaccuracy — technically true in some contexts, misleading in others.
Three vulnerabilities flagged:
That second one is sophisticated. An expert who disclaims access to the very data they're making population claims about is vulnerable to exclusion under Daubert v. Merrell Dow Pharmaceuticals and its progeny. A junior associate might not catch the juxtaposition.
These are trial-quality questions. The second one in particular — asking for specific data, literature, or statistical analysis — is a direct Daubert reliability attack that forces the witness to either produce sources or concede they don't have any.
No AI tool is perfect. Here's what we found and fixed during validation:
Early tests on an orthopedic knee deposition showed Deepgram inserting cardiology terms — "HAS-BLED" (a bleeding-risk score) in a discussion of knee dislocation reduction. Why? The keyword boost list we passed to Deepgram was cardiology-focused. The fix: per-specialty keyword sets for nine medical specialties (cardiology, orthopedic, emergency medicine, OB/GYN, neurology, general surgery, oncology, pulmonology, internal medicine), dynamically selected based on the case context.
When a batched segment contained both an attorney's question and the witness's partial answer, Claude sometimes called the combined text "syntactically incoherent" and scored medical accuracy at 0. It wasn't — it was just Q&A mashed together by the transcription engine. The fix: explicit prompt instructions telling the model that the input is live deposition transcript with multiple speakers.
One early cross-exam question contained a literal [SPECIFIC PROCEDURE] placeholder — the model didn't fill in the blank. The fix: explicit prompt rule against bracketed placeholders plus a post-processing regex filter that drops any question containing one.
These are the kinds of bugs that only show up when you run the tool on real depositions across multiple specialties. We found them. We fixed them. We deployed the fixes.
A 3-hour deposition costs the tool roughly $5-6 in API calls (Deepgram streaming + Claude Haiku for the four analysis lenses). That's the COGS.
The value to the attorney:
At $300-500/hour in billable time, that's $900-2,500 in value saved per deposition. A tool that costs us $6 and delivers $1,000-2,500 in value has obvious unit economics. Full firm-level modeling lives on the Courtroom AI ROI page, and the Courtroom AI product overview walks through the feature set in depth.
COURTROOM50 at checkout. Max 10 redemptions.
Courtroom AI isn't a replacement for attorney judgment. It doesn't decide strategy, choose which witness weaknesses to exploit, read a jury, or conduct the deposition. It's a real-time research assistant that catches the details your conscious mind misses while you're focused on the next question.
Nor is it a transcript replacement. The court reporter's certified transcript is still the record. Courtroom AI is a working-notes layer on top.
And it's not yet a tool that can read tone or body language. It analyzes words.
Over the next 30 days we're adding:
We're also working with several plaintiff med-mal firms on specialized add-ons. If you want to be one of them, the COURTROOM50 pilot is the door in.
Questions or feedback on this post? Email [email protected]. If you'd like to see a live demo on one of your own deposition recordings, we do that too.