Customer story · Legal AI · OpenAI compatible API

From first call to production-scale legal AI in three weeks.

Robot, Robot & Human routed billions of tokens in its first month through TrustedRouter while processing privileged litigation documents at production scale.

See the results Verify the gateway

Billionsof tokens routed in the first month

170,974documents across 34 active matters

3 weeksfrom first call to production batch scale

616,682 pagesextracted and analyzed

~1.49B charactersof extracted litigation text

More than 1B tokensduring the peak week

One base URLto start routing production calls

The customer

Legal reasoning over documents that cannot be treated casually.

Robot, Robot & Human is a legal AI platform. It ingests litigation productions, performs OCR and text extraction, embeds the resulting material, and runs multi-stage reasoning to produce searchable evidence, timelines, contradiction analysis, and attorney work product.

The inputs include opposing-party productions, privileged communications, and matters under active litigation. The team needed production-scale inference without building its own GPU fleet or maintaining a separate integration for every model provider.

The implementation

One API, explicit route policy, and an attested gateway.

RRH kept its OpenAI-shaped application calls and moved the base URL to TrustedRouter. It could use lower-cost models for extraction and stronger models for synthesis and verification while keeping routing, provider fallback, billing, and metadata accounting behind one interface.

Every request entered through TrustedRouter's attested confidential-compute gateway. TrustedRouter does not log or store prompt or output content. RRH can verify the running gateway image against the published source and choose ZDR or independently verified end-to-end confidential provider routes where the workload requires them.

Production results as of August 5, 2026

170,974

documents processed

across 34 active matters

616,682

pages analyzed

~126 GB

source files ingested

609,090

retrieval chunks embedded

190,054

case insights generated

25,610

evidence highlights surfaced

4,528

contradiction pairs

2,381 verified against source

2,254

timeline events reconstructed

In their words

Privacy had to be evidence, not another promise.

“We handle other people's privileged litigation documents, the kind of material that gets subpoenaed. ‘Trust us, we won't peek’ was never an acceptable answer to give a client. [...] And it was a base-URL change. We went from the first call to hundreds of thousands of documents in a few weeks, running cheap models for extraction and frontier models for the reasoning that actually matters [...] without standing up a single GPU.”
Robot, Robot & Human

Trust boundary

Gateway attestation and provider guarantees answer different questions.

The live attestation proves which TrustedRouter gateway workload is receiving the request and that it matches the published image. Downstream model handling remains route-specific and is published on each model and provider page. Use ZDR routes for contractual zero retention and independently verified E2E routes when the provider's model-serving boundary must also be confidential.

Speed

Production in weeks

RRH onboarded on July 3 and reached production batch scale within three weeks. The public snapshot covers the first month through August 5.

Scale

Batch inference without a GPU platform

The same API handled extraction, retrieval preparation, synthesis, source verification, and timeline reconstruction across entire litigation productions.

Control

Model tier matched to the task

Lower-cost routes handled volume work. Stronger routes handled the reasoning stages where answer quality mattered most.

Measurement note

Corpus and reasoning counts came from RRH's production database. Token-volume claims were cross-checked against TrustedRouter's metadata-only usage records. TrustedRouter did not inspect prompt or output content. Public figures are aggregated and do not identify a matter, party, document, user, or API key.

Route sensitive document workloads through an API your team can inspect.

Keep the SDK, choose the privacy tier, and verify the running gateway.

Legal AI guide