The case is often decided before a lawyer sees it
Most legal AI works on documents that already exist: research, drafting, review. The investigation is where the evidence is built and secured. If it is lost there, nothing downstream gets it back.
caseload is case management for licensed pre-trial investigation in personal-injury and insurance-defense work. It is built first for George A. Szermer, Inc., a licensed investigation agency, to replace a record that today lives in email and fax.
One post, start to finish
An illustrative matter. The people and details are invented; the pattern is real.
Day 0
A claimant with a reported back injury posts publicly about a weekend of paid moving work. The post says the bookings are in voicemails on a friend’s phone.
Day 1
The post reaches the file as a fax from the carrier’s document channel, control-stamped with a received time.
Day 3
The post is gone. The fax is now the only copy.
Later
The fax turns out to cut a comment thread off mid-sentence. Nothing in the file says so.
Meanwhile
The voicemails, the actual proof, sit on a private phone. They were always more perishable than the post.
Four design decisions
Each one answers a moment in that story, and each one lives in the database schema, not in a feature list.
Evidence decays on a clock
The post lasted three days.
Every evidence item carries a decay class: volatile (live social posts, days), short (911 and dispatch audio, roughly 30 to 90 days), medium, durable (filed police reports) or unknown. Protection is tracked separately, and a refused preservation request does not count as protection.
caseload continuously answers one question: what is decaying, with nothing protecting it? It never assumes an industry-default retention window. If a custodian’s retention is not established, it stays unknown.
How a copy was obtained decides its weight
The fax beat a screenshot.
Control stamps and a received time authenticate the channel the copy came through. A fresh screenshot authenticates nothing.
Each item records where it came from, how, when, and its authenticating marks. Truncation is a recorded fact, not a note someone might read, so a cut-off thread is visible to everyone downstream.
A pointer to evidence is not the evidence
The post pointed at the voicemails.
Treating the post as the evidence preserves the durable artifact and loses the one that decides the case.
Each item records what it was learned from. The voicemails become their own item, with their own custodian and their own, much shorter clock.
“Could not confirm” is not “not there”
Nobody has heard the voicemails yet.
A confidently wrong negative is the defining risk of this work. It is how an investigation report becomes a liability.
Every assertion carries a basis and one of six confidence levels, from documented to contradicted. Unconfirmed means we looked and could not confirm. The system never silently merges two people because they share a name.
The third attempt
caseload is the third try at this problem. The first two are worth being plain about.
casetrace
Dec 2024
A cases table with a client, a carrier and a claim number. The right join key, nothing around it.
Redwell
Dec 2024 to Jun 2025
AI document chat for case management. Archived.
caseload
Sep 2026 to now
Built from the investigation side out.
Redwell gave up on scanned documents. Its last substantive change marked image-only PDFs as skipped. In this work, police reports, medical records and faxes are mostly scans, so its AI worked on the minority of its own evidence. Reading scans is the core of the job, not a later phase.
Redwell claimed more than it built. It advertised privacy-law compliance that amounted to display text and a feature flag, and evidence tracking it never built. caseload ships no claim it has not implemented and tested.
Some of it was right. Privilege rules enforced in the database, an append-only activity log, and the trade’s own vocabulary all carried forward.
Proven
- Tenant isolation proven by attack. A security review found two ways one account could reach another’s records. Both are fixed, and each has a passing attack test.
- Cross-account sharing covered by 24 adversarial tests. A guest on a matter sees an item only after it has been sent to them.
- An append-only chain of custody. No role can rewrite it, including the database’s own service role.
- Local OCR, because scans are the evidence. On the first import, 305 of 325 pages had no text layer. Two unrelated models transcribe each page, and an extracted field that cannot be traced to the transcript is stored as unverified.
Not yet
- caseload has not run a real matter end to end. Migrating the agency’s own caseload is that test, and it will find gaps no review can.
- Whether the two OCR models catch each other’s omissions on our own pages is still being measured.
- Whether claims managers will log in to a shared view of their matter, instead of waiting on a fax, is still untested.
Technology
Launching at caseload.us
Building for legal work where the evidence comes from?
Related Case Studies
Explore more digital transformation success stories across industries