NewToca 10 is here. Branch, commit and roll back every workspace, with Git built in.See what’s new

toca.io
Talk to us

Spectral · the AI operating system

AI you can explain. Decisions you can defend.

Named functions, not a chatbot. Every finding traced to its source, every doubt routed to a person, every document staying inside your environment.

Run it on your own documentsHow it works

How it works · start with your documents

Documents in. A spectrum of findings out.

Spectral separates a document into individual findings. Each carries its source passage, is assessed on its own, and either goes straight on or goes to a person. Pick the documents you work with.

Clinical lettersClinical codingCase summariesLegal formsBank statementsInvoices
How Spectral reads a clinical letter One discharge summary enters a prism and is separated into seven labelled findings, each with its source. Where the model jury is unanimous the finding goes on automatically; where it is split, it goes to a person. ONE DISCHARGE SUMMARY Dear Dr Achampong, Mrs Sarah MOGGS (999 123 4567) was admitted on 28 Jul with chest pain. Troponin negative, ECG unremarkable. Reviewed on the post-take ward round and considered medically fit for discharge today. She will go home to her usual residence, although her daughter has raised concerns about managing the stairs and a package of care may need consideration. TTOs dispensed, three items. Please arrange cardiology follow-up in six weeks and onward referral to community rehabilitation. Prose. Invisible to every system the Trust runs. SPECTRAL SEVEN DECISIONS, EACH ASSESSED ON ITS OWN, EACH SOURCED TO A PASSAGE Discharge decision Medically fit for discharge · line 6 Unanimous ACTIONED Discharge date 04 Aug 2026 · line 7 Unanimous ACTIONED Follow-up required Cardiology, 6 weeks · line 12 Unanimous ACTIONED TTO medication 3 items, dispensed · line 11 Unanimous ACTIONED Onward referral Community rehab · line 13 Unanimous ACTIONED Discharge destination Usual residence, ambiguous · line 7 Split 2–1 TO A CLINICIAN Complexity flag Possible package of care · line 10 Split 2–1 TO A CLINICIAN Every model agrees: actioned, with the passage attached The models disagree: a clinician decides, every time

Scroll the diagram →

Worked example. Patient details are synthetic and the NHS number uses the reserved 999 test range; the verdicts are illustrative, not published performance.

Clinical

Clinical letter decision analysis

Live at Royal Berkshire

500,000clinical documents read at Royal Berkshire to date, around 12 seconds a letter

Spectral in the NHS →
Reads
Clinical letters, referrals and discharge summaries, and the rest of consultant correspondence.
Gives you
Every decision in the letter as a structured finding with its passage: fitness for discharge, destination, follow-up, medication and onward referral.
Decided by
A clinician, wherever the models disagree.
How Spectral proposes clinical codes One admission summary enters a prism and is separated into seven labelled findings, each with its source. Where the model jury is unanimous the finding goes on automatically; where it is split, it goes to a person. ONE ADMISSION SUMMARY ADMISSION SUMMARY · WARD 7 Mr John DOE (999 765 4321), 78 Admitted with productive cough, fever and breathlessness. CXR: right lower lobe consolidation. Treated for community- acquired pneumonia with IV antibiotics. Background: COPD, type 2 diabetes, hypertension, atrial fibrillation on apixaban. Creatinine rose from 92 to 168 on day 2 and settled with fluids. Discharged home on day 6. Prose. A coder reads every line of it today. SPECTRAL SEVEN CODES, EACH ASSESSED ON ITS OWN, EACH SOURCED TO A PASSAGE Primary diagnosis J18.9 Pneumonia, unspecified · line 7 Unanimous PROPOSED Type 2 diabetes E11.9 Without complications · line 8 Unanimous PROPOSED Hypertension I10 Essential hypertension · line 9 Unanimous PROPOSED Atrial fibrillation I48.9 AF and flutter, unspecified · line 9 Unanimous PROPOSED Anticoagulant Z92.1 Long-term anticoagulants · line 10 Unanimous PROPOSED COPD J44.0 with acute infection, or J44.9? Split 2–1 TO A CODER Acute kidney injury Creatinine 92 to 168, no diagnosis written Split 2–1 TO A CODER Every model agrees: proposed to the coder, with its passage The models disagree: the coder decides, or asks the clinician

Scroll the diagram →

Worked example. Patient details are synthetic and the NHS number uses the reserved 999 test range. The codes are illustrative.

Clinical

Clinical coding

In build at Royal Berkshire and Morecambe Bay

38 sper code suggestion, with the supporting passage attached

Spectral in the NHS →
Reads
The clinical record: admission and discharge summaries, clinic letters and operation notes.
Gives you
Proposed ICD-10 and OPCS codes, each with the passage that supports it.
Decided by
A clinical coder, wherever the models disagree, within thresholds agreed with the Trust's Clinical Safety Officer.
How Spectral summarises a document on a case file One letter on the file enters a prism and is separated into seven labelled findings, each with its source. Where the model jury is unanimous the finding goes on automatically; where it is split, it goes to a person. ONE LETTER ON THE FILE Northway Insurance · Motor Claims To: Carter & Co Solicitors Re: J Smith v A Patel · RTA 03.02.26 Date: 14 April 2026 We confirm that liability for the accident on 3 February 2026 is admitted, subject to the medical evidence. We offer £2,350 in settlement of general damages. This offer is open for 21 days. Please send the hire agreement and the invoices by 30 April 2026. One of hundreds on the file, each read in full. SPECTRAL SEVEN FINDINGS, EACH ASSESSED ON ITS OWN, EACH SOURCED TO A PASSAGE Document Letter from the defendant's insurer · head Unanimous TO THE FILE Date 14 April 2026 · head Unanimous TO THE FILE Offer £2,350, general damages · line 10 Unanimous TO THE FILE Documents asked for Hire agreement and invoices · line 12 Unanimous TO THE FILE Deadline 30 April 2026 · line 13 Unanimous TO THE FILE Liability Admitted, subject to the medical evidence Split 2–1 TO A LAWYER Offer expiry 21 days from the letter, or from receipt? Split 2–1 TO A LAWYER Every model agrees: summarised to the file, sourced The models disagree: a lawyer decides, every time

Scroll the diagram →

Worked example on a synthetic personal injury file. The parties, the insurer and the figures are invented; the verdicts are illustrative.

Legal

Legal case summarisation

In live use

7,000documents a day read by the same engine inside one firm's network

Toca Legal →
Reads
Medical records and every other document on a case file, in date order: letters, emails, attendance notes, engineer's reports, hire documents and court documents.
Gives you
A neutral summary of each document with its dates, figures, deadlines and actions, and a case narrative that updates as each one lands.
Decided by
A lawyer, wherever the models disagree.
How Spectral reads a legal form One form e enters a prism and is separated into seven labelled findings, each with its source. Where the model jury is unanimous the finding goes on automatically; where it is split, it goes to a person. ONE FORM E FORM E · FINANCIAL STATEMENT Applicant: Jane Doe Family home, 14 Elm Road Value, agent's estimate £485,000 Mortgage outstanding £127,000 Pensions Teachers' Pension (DB) CETV £312,000 Personal pension (DC) £189,000 Net income £3,120 a month [x] Valuation enclosed [ ] To follow Typed, ticked and checked against the bundle. SPECTRAL SEVEN FIELDS, EACH ASSESSED ON ITS OWN, EACH SOURCED TO A PAGE Applicant Jane Doe · page 1 Unanimous TO THE FILE Family home value £485,000, agent's estimate · page 4 Unanimous TO THE FILE Mortgage outstanding £127,000 · page 4 Unanimous TO THE FILE Personal pension (DC) £189,000 · page 7 Unanimous TO THE FILE Net income £3,120 a month · page 9 Unanimous TO THE FILE Teachers' Pension (DB) CETV £312,000, not like-for-like with DC Split 2–1 TO A LAWYER Valuation Ticked enclosed, none in the bundle Split 2–1 TO A LAWYER Every model agrees: to the file, page referenced The models disagree: a lawyer decides, every time

Scroll the diagram →

Worked example on a synthetic Form E. The parties, the property and the figures are invented; the verdicts are illustrative.

Legal

Legal form understanding

Bundles in live use, Form E in evaluation

95%+of bundle pages identified correctly, against around 60% for legacy OCR

Toca Legal →
Reads
Court and claim forms, typed or ticked, and every page of the bundle behind them.
Gives you
Each answer as a field with its page, cross-checked against the statements behind it, with the gaps flagged.
Decided by
A lawyer, wherever the models disagree or a figure does not reconcile.
How Spectral reads a bank statement One bank statement enters a prism and is separated into seven labelled findings, each with its source. Where the model jury is unanimous the finding goes on automatically; where it is split, it goes to a person. ONE BANK STATEMENT CURRENT ACCOUNT · J SMITH Statement 1 to 31 March 2026 01 Mar Balance brought forward 212.40 03 Mar SALARY ACME LTD 1,412.00 04 Mar RENT STANDING ORDER 650.00 06 Mar TESCO STORES 2214 48.20 11 Mar TFR J SMITH 40-11-08 400.00 15 Mar BETFRED ONLINE 20.00 22 Mar COUNCIL TAX DD 118.00 and 21 more transactions 31 Mar Balance carried forward 96.15 A PDF of numbers no system can query. SPECTRAL SEVEN FINDINGS, EACH ASSESSED ON ITS OWN, EACH SOURCED TO A LINE Account holder J Smith · page 1 Unanimous TO THE FILE Period 1 to 31 March 2026 · page 1 Unanimous TO THE FILE Regular income £1,412 salary, monthly · 3 Mar Unanimous TO THE FILE Fixed outgoings £650 rent by standing order · 4 Mar Unanimous TO THE FILE Days under £500 27 of 31 · running balance Unanimous TO THE FILE Transfer out £400 to a second account: disclosed? Split 2–1 TO A PERSON Discretionary spend £20 online betting: material or not? Split 2–1 TO A PERSON Every model agrees: classified, with a confidence score The models disagree: a person decides, every time

Scroll the diagram →

Worked example on a synthetic bank statement. The account holder and the transactions are invented; the verdicts are illustrative.

Finance and back office

Bank statement analysis

In live use

Secondsfor an impecuniosity analysis that took a handler three to four hours

Toca Legal →
Reads
Bank and credit card statements, as PDFs or scans.
Gives you
Every transaction classified with a confidence score, and the position you need built from them, such as an impecuniosity analysis.
Decided by
A person, wherever the models disagree.
How Spectral reads a supplier invoice One supplier invoice enters a prism and is separated into seven labelled findings, each with its source. Where the model jury is unanimous the finding goes on automatically; where it is split, it goes to a person. ONE SUPPLIER INVOICE From: accounts@ashford-facilities.co.uk Subject: Invoice INV-20481, PO 7741 Please find attached our invoice for the September cleaning contract. Terms are 30 days as agreed, though payment on receipt would be appreciated this month. INVOICE INV-20481 Ashford Facilities Ltd PO 7741 Net 4,050.00 VAT 810.00 Total 4,860.00 GBP Pay to: 20-45-77 3310 8821 (new account) Prose and a PDF. Invisible to every system you run. SPECTRAL SEVEN FINDINGS, EACH ASSESSED ON ITS OWN, EACH SOURCED TO A PASSAGE Supplier Ashford Facilities Ltd · invoice header Unanimous TO THE RECORD Invoice number INV-20481 · invoice header Unanimous TO THE RECORD Purchase order PO 7741, matched to the open order Unanimous TO THE RECORD Net amount 4,050.00 · totals Unanimous TO THE RECORD VAT 810.00 at 20% · totals Unanimous TO THE RECORD Bank account Differs from the supplier record · footer Split 2–1 TO A PERSON Payment terms 30 days on the invoice, on receipt in the email Split 2–1 TO A PERSON Every model agrees: filed, no one needs to look The models disagree: a person takes a look, every time

Scroll the diagram →

Worked example on the supplier invoice from the platform page. The supplier, the figures and the account numbers are invented; the verdicts are illustrative.

Finance and back office

Invoice processing

Live at Delta Housing

1,000+repair invoices a month processed automatically on Toca at Winn Group

Follow this invoice through the platform →
Reads
Supplier and repair invoices, with their covering emails, purchase orders and statements.
Gives you
Supplier, number, order, net, VAT and bank details, matched to your records, with anything that differs flagged.
Decided by
Your accounts team, wherever the models disagree or a detail does not match.

The problem

Unlock valuable information locked in prose.

The decision that matters is usually a paragraph: fit for discharge, hire period, payment terms, a changed bank account. A person reads it and keys the result into another system. Spectral turns that prose into structured, auditable findings at source, so the systems you already own can search it, measure it and act on it.

Consensus Validation

A jury of models. One accountable verdict.

Every extraction is put to more than one model. Unanimous findings proceed. If any model dissents, the finding stops and goes to a person with the disagreement shown, not averaged away. Each verdict is logged against the passage it came from.

  1. Read

    Parsed in place, inside your environment. Nothing is uploaded and no external service is called.

  2. Separate

    Each fact and decision is extracted on its own, with the passage it came from attached as evidence.

  3. Put to the jury

    More than one model assesses every extraction. Agreement and disagreement are both recorded.

  4. Route

    Unanimous findings proceed at the threshold you set. Anything short of that goes to the person who owns the decision.

Measured in production

Named organisations. Real numbers.

NHS figures are from 213,000 live documents processed at Royal Berkshire to August 2026, of which 24,000 were routed for human review. Legal figures are from live UK volume personal injury deployments. Results vary by document type and review threshold.

500,000clinical documents read at Royal Berkshire to date
12 sto read a discharge letter and separate the decisions
1 in 9routed to a person where the jury is not unanimous
0records leaving the customer's network
7,000 a daydocuments read by the same engine inside one firm's network
80%lower cost of basic hire rate rebuttal reports, up to, at Bond Turner
Secondsfor an impecuniosity analysis that took a handler three to four hours

Running at Royal Berkshire and Morecambe Bay (NHS), Bond Turner (Anexo Group), Winn Group, eCapital, Delta HousingCertified UKCA Class I, ISO 42001, ISO 27001

Engineered AI

Hardened for your demanding tasks.

Each AI function has defined inputs, defined outputs and is tested and calibrated on real documents. Use the foundational and tested Spectral components to solve your hardest organisation data problems.

Referral cohorting

Classifies incoming referrals by specialty, urgency and pathway suitability, and surfaces the relevant history for the clinician answering an Advice & Guidance request.

Alongside eTriage · evaluation stage

Basic hire rate rebuttal

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

With Bond Turner · in development

Schedule of Loss

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

Personal injury and quantum · in build

The console behind the functions

Gain full visibility of the AI in your organisation.

Each function is versioned like software. Prompt, document, model version, verdict and reviewer stay together, ready for audit.

The Spectral home screen: function executions this year and the list of published and draft functions
StudioOne workspace for every document function, versioned and published like software.
Authoring a function: the prompt for summarising a supplier document and its response schema
AuthorWrite the instruction in plain English and define the output you expect.
Tests for the Document Summary function, with one test result open beside its structured output
TestEach test pairs a real document with the output it should produce.
A deployment run over a supplier file, with a timeline of documents and the live account narrative
RunDocuments are processed in date order, and the account narrative is rebuilt as each finding lands.
The audit log: one entry open, showing the endpoint, the request and the structured response
AuditPrompt, document, model versions, verdict and reviewer kept together. Exportable for audit.

Swipe the screens →

Sovereignty

The data boundary is your boundary.

Spectral runs where the documents already are. No identifiable data is sent to an external AI service. Models are fixed, versioned artefacts and are never trained on your documents.

On-premise or air-gapped

Your hardware, no external calls in normal operation. The model used inside NHS Trust networks and law firms.

Your cloud tenancy

A tenant you control, such as Azure UK or an HSCN-connected environment. Your keys, access policy and retention rules.

Hybrid

Processing on-premise, with orchestration and reporting in a secure environment where that suits your estate.

Assurance & Evidence

The foundation behind our software.

Device regulation, clinical safety, professional obligations, data protection, information security and human oversight each need their own evidence. Position at August 2026; the full evidence pack is available on request.

MHRA device status

Registered as a Class I medical device. New clinical use cases are classified with the MHRA before deployment, not assumed.

DTAC, DCB0129, DCB0160

DTAC evidence provided. Safety case maintained by Toca as manufacturer with a named Clinical Safety Officer.

ISO/IEC 42001

Certified at company level: external assurance across the full AI lifecycle.

Information security

ISO 27001, DSPT and Cyber Essentials maintained; penetration testing evidence available.

Data protection

On-premise processing, no data egress, no retraining on your data. We support your DPIA.

Post-market surveillance

Ongoing monitoring of model performance, with incident reporting aligned to LFPSE and MHRA routes.

Human oversight

Consensus Validation puts a qualified person on every significant decision, as part of the workflow.

Contact us to learn more

FAQs.

Does any data leave our network?

No. Spectral processes documents inside your environment. No identifiable data is sent to an external AI service and no third party sees your records.

Do you train models on our data?

No. Models are fixed, versioned artefacts. Your documents produce output for you, and for nothing else.

Does anything get actioned without a person?

Only where you have set a threshold allowing it, and only where the jury is unanimous. You set and change the thresholds, per function.

Does this replace our EPR, CMS or data platform?

No. Spectral reads the documents you already hold and writes structured output back into the systems you already run.

What do we need in place?

Access to the documents, and somewhere to write the result. APIs and integration engines where they exist; otherwise the application's own screens, under a service account.

How do we buy it?

Through G-Cloud as a standard Call-Off Contract in most public sector cases; directly otherwise. Every deployment starts with a scoped evaluation on your own documents.

Run it against your own documents.

A scoped evaluation on your documents, in your environment, reviewed by your own clinical, legal or finance leads. Every deployment has started there.

Book an evaluationSpectral in the NHSToca Legal