Back to blogBack on the Bench: Echosaw vs Evidence Copilot
Company

Back on the Bench: Echosaw vs Evidence Copilot

We gave Echosaw and Evidence Copilot the same street-camera clip and the same speech clip, asked the same questions word for word, and checked every answer against the footage. Echosaw answered all six street-clip questions; Evidence Copilot declined to answer all six. Our misses are on the scorecard too.

Echosaw TeamUpdated 5 min read

Two months ago we put Echosaw on the bench against general AI media tools. This time the bench is case footage, and the other entrant is a product built for lawyers: Evidence Copilot. Same two files, same questions word for word, and every answer checked against the footage frame by frame. The full scorecard is on Echosaw vs Evidence Copilot.

Why this bench

A personal injury file fills up with video long before it reaches a deposition: dashcam clips, doorbell and street cameras, bodycam, recorded calls, the depositions themselves. Someone at the firm has to watch all of it to find the few seconds that matter. That is the job we built Echosaw for video evidence to do, so that is the job we tested.

The rules

Both products got the same two files on November 12, 2026. Clip 1 is a 111-second silent street-camera clip: a parked Nissan pickup, a Toyota Tacoma pulling out of a driveway, a silver Toyota Sienna, a silver Lexus sedan, a white minivan, and trash bins marked PROPERTY OF MIDWAY CITY SANITARY DISTRICT. Clip 2 is 110 seconds of a NASA public-domain call with four speakers.

We asked each question separately, word for word, in each product's built-in assistant: Evidence Copilot's evidence agent on a 7-day trial, and Emma on an Echosaw account. Echosaw was re-run on November 13 (UTC) on its current release, on the same uploads, each question in a fresh chat with no earlier conversation; that run is the one scored here. Evidence Copilot could not process the street clip's original 4K HEVC file, so both products were given the same 1080p H.264 copy for the questions.

Two clips and one run is not enough to generalise from. It is enough to show real differences, and the scorecard shows our misses next to our wins.

The street camera

We asked six questions about Clip 1: the make of each vehicle, the model, the license plates, every appearance with timestamps, direction of travel, and all visible on-screen text.

Evidence Copilot's assistant captured frames but said their content was not visible to it, and declined to answer all six.

Echosaw answered all six. It confirmed the parked Nissan and named Toyota and Lexus as the likely makes of the moving pickup and the sedan, listed when each of the five nearby vehicles appears and which way each moving one travels, read every piece of on-screen text including the clock overlay and the bin web address, and said no plate was legible. That last one matters: no plate in the clip is legible, and neither product invented one.

Street clipEchosawEvidence Copilot
Make of each vehicle✓ Nissan identified; Toyota and Lexus named as the likely makes✗ declined to answer
All visible on-screen text✓ clock overlay, IPCAM, bin text, phone number, website✗ declined to answer
Every appearance with timestamps✓ all five nearby vehicles, within about a second✗ declined to answer
License plates✓ none legible✗ declined to answer
Model of each vehicle◐ likely models named for four of five vehicles, none stated as certain✗ declined to answer
Direction of travel✓ all four moving vehicles right✗ declined to answer

The speech clip

Echosaw's transcript separated all four speakers, with each change of speaker within a second of the footage. Evidence Copilot's transcript used three speaker labels.

Four speakers: Butch Wilmore (0:00 and 1:33), Suni Williams (0:09), Bill Nelson (0:49), Pam Melroy (0:56).

Asked who said what, with timestamps, Emma listed seven timestamped turns under Speaker 1 to 4, each under the right speaker, and named Butch, Suni (heard as "Sonny") and Pam from hand-offs and greetings in the call, flagged as inferred from the audio. Bill Nelson's name is never said in the clip, and Emma left him unnamed. Evidence Copilot listed lines with timestamps for three speakers, with names inferred from the conversation, and gave Bill Nelson's line to Butch Wilmore. That one goes to Echosaw.

Back to work

The bench gave us a list, in order:

  • Models. Emma named likely models but stated none as certain, and did not identify the silver Sienna at 18s.
  • Only what was asked. The make answer also gives each vehicle's model, and the plate answer also gives direction of travel.

What it means for a firm

The question a paralegal brings to a street-camera clip is rarely "summarise this." It is "when does the Toyota pull out," "what does the sign say," "is the plate readable." Echosaw watches the picture as well as the sound, and every answer opens at its timestamp so you can check it against the footage yourself.

The same applies to the rest of the file. Turn a recorded deposition into a page-line deposition summary, or a recorded call into a timestamped 911 call transcript. See how it fits a practice on Echosaw for personal injury.

Echosaw is analysis software, not a law firm. Nothing here is legal advice; for all advice, please seek a lawyer.

Count for yourself

The fairest test is your own footage. Your first upload to Echosaw is free, and Evidence Copilot offers a 7-day trial. Ask both the same questions and compare the answers to what the camera shows. The questions, both products' verbatim answers, the ground truth and the scripts are in the test harness.


Tested November 12, 2026: one run of Evidence Copilot, and Echosaw re-run on November 13 (UTC) on its current release. Results may change as both products update. Echosaw: Encrypted in transit (TLS 1.2+) and at rest (AES-256), never used to train AI models, backed by a 99.9% uptime SLA.

Stop watching hours of footage to find one moment.

Echosaw transcribes every word and identifies what's on screen (vehicles, people and visible text), and every answer links to the exact moment. Start with our Starter plan for $9/month — usage-based pricing with no per-seat fees, so you only pay for what you analyze.