toca.io/legal
Book a demo

SPECTRAL · DOCUMENT INTELLIGENCE, RIGHT-SIZED

Explainable AI for the back office, at volume.

Spectral is a sovereign-first platform. It reads every document on a case file and pulls out the facts that decide a claim: hire period, impecuniosity, prognosis, when the limitation clock started. It runs on right-sized models inside your own environment, so reading a hundred thousand pages costs what you expect and no client data leaves your control.

Book a demo See how it works ↓ Assurance & SRA position

THE PROBLEM · WHAT AI COSTS AT CASELOAD SCALE

The right-sized model protects your margin.

Spectral uses open-weight AI models to extract data and automate the heavy administration around your CMS.

HOW IT WORKS · ONE FILE, MANY FINDINGS

White light in. A spectrum of findings out.

Spectral extracts each fact as a separate finding with its source page attached. Clear-cut findings go straight to the file. Anything ambiguous goes to a lawyer in the loop.

How Spectral reads a case file A claimant bundle enters a prism and is separated into seven labelled findings, each with its source page. Where the model jury is unanimous the finding goes to the file automatically; where it is split, it is routed to a lawyer. CLAIMANT BUNDLE · 88 PAGES 1 Claim notification form p.1 2 Credit hire agreement p.3 3 Bank statements, Jan–Apr 2026 p.7 4 GP records p.22 5 Medical report, Dr A Rahman p.41 6 Engineer’s report p.47 7 Correspondence p.53 “…prognosis 6 to 9 months from the date of the accident. The GP records note an episode of low back pain in 2023 which may be of relevance.” p.43 ONE CASE FILE Prose and scans. Invisible to every system the firm runs. SPECTRAL Hire period 42 days · hire agreement p.3 Unanimous TO THE FILE Daily hire rate £187.50 + VAT · hire agreement p.4 Unanimous TO THE FILE Impecuniosity 27 of 31 days under £500 · statements p.9 Unanimous TO THE FILE Repair duration 31 days · engineer’s report p.49 Unanimous TO THE FILE Prognosis Soft tissue, 6 to 9 months · report p.43 Unanimous TO THE FILE Need for hire School run cited, no 2nd car · p.5, p.61 Split 2–1 TO A LAWYER Pre-existing condition Back pain 2023, relevance unclear · p.31 Split 2–1 TO A LAWYER SEVEN FINDINGS, EACH ASSESSED ON ITS OWN, EACH SOURCED TO A PAGE Jury unanimous, to the file automatically Jury split, a lawyer decides, every time

Scroll the diagram →

Worked example on a synthetic credit hire claim. The claimant, documents and figures are invented; the verdicts shown are illustrative, not measured results.

CONSENSUS VALIDATION · WHAT LAWYER IN THE LOOP MEANS HERE

A jury of models. One accountable lawyer.

Every extraction is put to more than one model. Unanimous findings proceed. If any model dissents, the finding stops and goes to a lawyer with the disagreement shown, and every verdict is logged against the passage it came from.

  1. Read

    Parsed in place, inside your environment.

  2. Separate

    Each fact extracted on its own, with its page attached.

  3. Put to the jury

    More than one model assesses every extraction.

  4. Route

    Unanimous findings go to the file above the threshold you set. Anything else goes to a qualified person.

Secondsfor an impecuniosity analysis that took a handler 3–4 hours, UK PI firm, 70-strong team
Up to 80%lower cost of basic hire rate rebuttal reports, Bond Turner
7,000 a daydocuments read by the same engine inside one firm’s network
1 in 9routed to a person where the jury isn’t unanimous, measured at NHS scale

Legal figures from Toca deployments in UK volume PI. NHS figures from live Spectral processing at Royal Berkshire NHS Foundation Trust to August 2026, measured over a 213,000-document sample. Results vary by document type and the thresholds each firm sets.

Spectral deployments

Bond TurnerLegal · Anexo Group Winn GroupLegal · accident management & PI eCapitalCommercial finance · invoice finance NHSRoyal Berkshire & Morecambe Bay · same engine

WHAT IT DOES · NAMED FUNCTIONS, NOT GENERAL AI

Purpose-built for the jobs on your file.

Named functions with defined inputs, defined outputs and their own tests on real documents. Use ours, tune them, or write your own.

Impecuniosity analysisCredit hire · bank statements

Reads claimant bank statements and builds the impecuniosity position, each transaction classified with a confidence score. Previously three to four hours per claim.

In live use

Basic hire rate rebuttalCredit hire · with Bond Turner

Drafts the rebuttal report from hire details, comparables and authorities, at up to 80% below the cost of commissioning one. Integrated with Proclaim at Bond Turner.

In development

Bundle classification & extractionAny matter · OCR and structure

Identifies each page of a bundle at 95%+ accuracy (around 60% for legacy OCR) and extracts dates, parties and figures with page references.

In live use

Chronology & case narrativePI & litigation

Builds a dated chronology across the whole bundle and keeps the narrative current as documents land.

In live use

Schedule of LossPI · quantum

Turns invoices, payslips, receipts and care records into a court-ready Schedule of Loss, every line traced to its source.

In build, not yet live

Form E & financial disclosureFamily · matrimonial finance

Cross-checks Form E, bank and pension statements into a comparable asset schedule, flagging gaps for the lawyer.

Evaluation stage

SOVEREIGNTY · WHERE THE PROCESSING HAPPENS

Runs inside your four walls.

Client files, including privileged material, are processed inside your environment and stay there. Models are fixed, versioned and never trained on your documents, so the data boundary and the firm boundary are the same line.

On-premise or air-gappedYour data centre

Your own hardware, no external calls. The same model we run inside NHS Trust networks.

Your cloud tenancyUK region

Azure UK or equivalent. Your keys, your access policy, your retention rules.

HybridSplit by sensitivity

Processing on-premise, orchestration and reporting in a secure cloud.

ASSURANCE · EVERY LAYER, NOT ONE BADGE

Evidence at every layer, not one badge on the box.

Professional obligations, privilege, data protection, information security and human oversight each carry their own requirements. Our position on each is set out in full, and where a question is still open we say so.

The full assurance position →

FOR THE IT DIRECTOR · THE CONSOLE BEHIND THE FUNCTIONS

Engineered like software. Governed like legal work.

Spectral dashboard, document functions across active matters
StudioOne workspace for every document function, versioned and published like software.
Authoring a function in plain English
AuthorWrite the instruction in plain English and define the output you expect.
Tests pinning real documents to expected structured output
TestEach test pairs a real document with the output it should produce.
A case run with the narrative rebuilding live
RunDocuments are processed in case order, and the narrative is rebuilt as each finding lands.
The audit log, every prompt, document and output retained
AuditPrompt, document, model versions, verdict and reviewer kept together. Exportable for SRA review.

Swipe the screens →

FAQ · CONFIDENTIALITY, DATA AND PROCUREMENT

Common questions.

What about privileged material?

Processed in place under your access controls. Nothing leaves the firm.

Do you train models on our documents?

No. Models are fixed, versioned artefacts.

Does anything reach the file without a lawyer?

Only where the model jury is unanimous and your firm has set a threshold allowing it. You control the thresholds, per function.

Does this replace our CMS?

No. Spectral reads the documents you hold and writes findings back into the system you run.

What do we need to start?

Access to your documents and somewhere to write findings. If your CMS has an API we use it; if not, Toca Connect works through the application’s own screens.

See it run on your own documents.

Every deployment starts the same way: a scoped evaluation on your files, in your environment, with your lawyers reviewing the output.

Book a demo Cost calculator