Casefile
Reads the file
Document understanding
Medical, legal, engineering, financial. Scanned, photographed, handwritten. Tuned to the Australian document set rather than a translated American one.
217 document typestarget v1
Claims intelligence for Australian insurers
intactly reads the entire claim file, checks what the documents say against each other, and tracks every deadline the claim is under: Code, statutory and scheme. Findings arrive with the source pages attached. In this market, an answer you cannot evidence is worth less than no answer at all.
Output from the obligation engine and the integrity examiner. Every row cites the instrument behind it. Every finding carries its source pages.
The problem
They are judged on delay, on denial, and on whether they can explain a decision eighteen months later to someone who was not there. That is where the complaints land, and it is where the regulator is currently looking.
Platform
Four of these match what the global platforms already do well. Two of them do not exist anywhere else, for the simple reason that no other regulator asks for them.
Medical, legal, engineering, financial. Scanned, photographed, handwritten. Tuned to the Australian document set rather than a translated American one.
Reconciles what each document asserts and surfaces the disagreements: dates, capacity, quantum, coverage, duplication, document provenance.
Ask the file anything. Every sentence of every answer carries a document, a page and a quote. Without an anchor the system will not answer.
ABN and ASIC registers, property and title records, Bureau of Meteorology event data, address resolution, ICA catastrophe declarations.
Every timeframe the claim is under, evaluated continuously: the Code, RG 271 dispute resolution, and the statutory clocks of eight jurisdictions.
The record your board, ASIC and AFCA all ask for. Produces the APP 1 automated decision disclosure and the CPS 230 service provider pack without a drafting exercise.
Approach
Speed is the easy part. The hard part is being able to show, two years on, exactly what the system saw and why anyone acted on it.
Most claims systems only see what somebody typed into a form. The problem is almost always in what nobody typed: the third page of a specialist report, a handwritten note in the margin of a certificate of capacity, a revised scope buried in a builder's email attachment.
Casefile takes the file as it actually exists. Faxes, photographs of paperwork, mixed-orientation PDFs, long correspondence threads. It turns all of it into facts that can be queried and traced back to where they came from.
It knows what a claim is. It does not know that a certificate of capacity is a NSW workers compensation instrument with a statutory review cycle, that a cash settlement carries Code obligations a repair does not, or that the insurer on a CTP file may be discharging duties owed to a regulator rather than to a policyholder.
That difference is the difference between pulling a value out of a document and understanding what the document does.
When a claims officer cannot tell why the system flagged a file, one of two things happens. They trust it when they should not, or they stop reading it altogether. Both end up in the same determination.
Every intactly finding names its factors in plain English, cites the pages it came from, and states the innocent explanations next to the adverse one. A finding that only argues one side is not evidence. It is advocacy, and it will be treated as such by anyone reviewing it.
APRA wrote to regulated entities in April 2026 naming third party AI risk, models buried in vendor platforms with upstream dependencies nobody can account for, as the largest gap it found. A claims system that cannot explain its own supply chain has become a finding waiting to be written up.
From 10 December 2026 your privacy policy has to describe what decisions are made by automated processes, what personal information they use, and broadly how they work. Ledger keeps that description current as the system changes, rather than leaving legal to reconstruct it once a year.
Comparison
The offshore claims platforms are good software. They were written for the FCA, the NAIC and the Canadian codes. What they lack here is not capability.
| Capability | intactly.ai | Offshore claims AI | General purpose LLM |
|---|---|---|---|
| Document understanding at claim file scaleThe core engine | Yes, Australian corpus | Yes, US and Canadian corpus | Partial, no layout model |
| Cross document integrity checkingWhere sources disagree | Yes, anchored findings | Yes, scored signals | No, no persistent file state |
| Code of Practice clocksAcknowledge, decide, expert reports, four months | Native | Not modelled | No |
| State scheme statutory clocksSIRA, icare, TAC, WorkSafe, WorkCover, RTWSA | 8 jurisdictions | Not modelled | No |
| RG 271 dispute resolution30 calendar days, then AFCA | Native | Not modelled | No |
| Vulnerability and hardship handlingFamily violence, hardship, mental health | Native, walled off from integrity | No local equivalent | No |
| APP 1 automated decision disclosureRequired from 10 December 2026 | Generated | Your problem | Your problem |
| Data residencyClaim content stays in Australia | ap-southeast-2 | Varies, often US inference | Varies by provider |
| CPS 230 service provider packTolerances, dependencies, exit plan | Pre-built | On request | Not applicable |
| Declines, reductions, recoveriesThe adverse calls | People only | Automatable | Unconstrained |
A finding, in full
This is the real output shape. Note what is here that a scoring product does not produce: both source pages, the named factors, and the innocent explanations.
Certificate DOC-0114 restricts lifting above 5 kg from 14 April to 12 May 2026. Rehabilitation report DOC-0139, dated 29 April 2026, records lifting estimated at 12–15 kg, inside that period.
Lines of business
Australia does not have one claims regulator. It has a national Code, a corporate regulator, a prudential regulator, an ombudsman, and a separate personal injury scheme in every state.
Value
Every input is yours to set and every line of output states where it comes from. No industry averages, no borrowed percentages. Move the sliders. The assumptions are all on screen.
Commitments
These sit in the data model and the test suite, not in a policy document. A promise that survives a roadmap argument is the only kind worth writing into a contract.
Declines, reductions and recovery actions are recommendations only. There is no setting that changes this, which is usually the first question a board asks.
The system cannot state a fact about a file without a document, page and quote attached. Extractions without an anchor are discarded rather than downgraded.
Family violence, hardship and mental health signals are walled off from every integrity, liability and recovery path. Crossing that wall raises an error, not a warning.
Claim content stays in Australian regions, including when a model is being called. No cross border inference, and no exceptions for capacity.
Written into the contract rather than a best effort undertaking. Your file is not our training data.
Priced on claims under management. Never on savings share, recoveries or declines. A supplier that earns more when claims are denied cannot be defended at AFCA.
Talk to us
Six weeks on one portfolio, measured against a matched control on four numbers agreed before we start: days to decision, projected breaches per thousand claims, rework rate at internal dispute resolution, and AFCA referral rate. We publish the method with the result.
Prefer email? hello@intactly.ai · Open the live demo