Early Support Before Liability: Track the Process, Not the Person

Early support should move without waiting for liability, but its data must not become a shadow claim assessment. A two-lane AI board can track voluntary support and human action while keeping evidence, consent, privacy and every liability judgement under separate human control.

Support can start before liability is decided. Scoring cannot. The safe role for AI is to show whether a human offered voluntary support, whether the employee chose to participate and what authorised action is due next, working only on de-identified records.

Do not ask the model whether a psychological injury looks serious, work related, credible or likely to be accepted. Do not let participation, refusal, withdrawal or contact speed become a proxy for any of those things. That turns an early-support program into an undisclosed liability screen.

Build two lanes instead. One holds the minimum process data needed to deliver voluntary support. The other indexes evidence for the human claims process. A controlled bridge records any authorised transfer between them, while an anti-inference dictionary states what every process status does not mean.

What does support before liability actually mean?

Comcare's current early-intervention guidance says early intervention may be used for injuries and illnesses that are work related or non work related, can happen whether or not an employee has lodged a workers compensation claim, and should be voluntary for an employee. It also says early intervention does not affect an employee's ability to make a claim. Those are the foundations of the support lane.

The better-practice early-intervention resource adds that a participant may withdraw from the program at any time and that a program should not be structured in a way that hinders or prevents a worker's right to lodge a claim at any time. Its model includes informed consent at initial contact, clear entry and exit criteria, a worker's ability to nominate a preferred treating health practitioner, and a process for transition to a claim for workers who require further treatment, care or support.

That does not mean every action before liability is an informal voluntary service. Comcare's rehabilitation information for employers distinguishes an organisation's early-intervention or rehabilitation policy from statutory rehabilitation. It says rehabilitation should start as soon as possible and regardless of the employee's intention to lodge a claim under the SRC Act, but if a rehabilitation assessment is arranged under the SRC Act there are requirements the rehabilitation authority must comply with before arranging the assessment and making the determination.

The current in-force Safety, Rehabilitation and Compensation Act 1988 is Compilation No. 82, with a compilation date of 1 July 2026. Section 14 is the general compensation-liability provision. Sections 36 and 37 separately concern rehabilitation assessments and programs. If a human rehabilitation authority invokes those provisions, send the matter to the controlled statutory rehabilitation workflow, where the section 36 assessment has its own AI boundaries. Do not label that determination process as voluntary early support or ask AI to make it.

Funding also needs a role check. Comcare's claims information for employers describes pre-liability treatment reimbursement and rehabilitation-cost arrangements inside a section expressly limited to Australian Government agencies and statutory authorities. A self-insured licensee should not copy those mechanics as its own entitlement or reimbursement rule. Its authorised people must apply the executed licence, SRC Act, internal funding policy and applicable employment arrangements. Comcare also notes that for a self-insured licensee the contracted claims manager or reviewer named in the licence does not have authority to make or reconsider any rehabilitation determination, and that those powers can only be exercised by the chief executive officer or managing director, or an officer or employee to whom they have delegated the power.

The liability lane remains separate. Comcare says the claims manager gathers and reviews information from the employee, their employer and treating medical practitioners to determine liability, and reviews the information provided, including medical evidence, before determining liability for the claim. An offer of support does not prove section 14 liability. Accepting support is not an admission, and declining it is not contrary evidence.

This prompt creates the board without assessing the person. A human support owner and a human claims professional must verify every entry against the source systems before operational use.

Create a two-lane process board for de-identified matter [MATTER_ID].

Use these inputs separately:
SUPPORT PROCESS RECORDS: [SUPPORT_RECORD_IDS]
CLAIM EVIDENCE INDEX: [CLAIM_RECORD_IDS]
SYSTEM DATE: [SYSTEM_DATE]

Lane A, VOLUNTARY SUPPORT PROCESS:
- date the organisation was notified
- approved program and route
- human contact owner
- offer date and channel
- employee choice recorded as NOT YET RECORDED, ACCEPTED, DECLINED or WITHDRAWN
- consent scope, date and source record
- support activation status
- next human action and internal due date

Lane B, LIABILITY EVIDENCE INDEX:
- claim lodged date, if supplied
- source document ID and received date
- evidence status only: REQUESTED, RECEIVED, SOURCE NOT VERIFIED or HUMAN REVIEWED
- human claims owner and next action

Create an empty BRIDGE LOG. Do not join the lanes, summarise clinical content, infer causation, assess credibility, score injury severity, predict claim outcome, triage the employee or treat support participation as evidence. Mark unsupported fields UNKNOWN. Human source review is required.

How should the two-lane board work?

Lane A should be deliberately boring. It is a support tracker, not the recover-at-work and suitable-duties record, which sits downstream and under separate control. It needs the support offer, human owner, contact channel, employee choice, consent scope, activation status and next action. It does not need diagnosis, symptom narrative, manager impressions, EAP content, practitioner notes or a model-generated risk level.

For psychological matters, ban sentiment analysis, emotion detection and behavioural scoring. Do not mine email, calls, leave patterns, productivity measures or meeting transcripts to find people who might need help. The board should begin with an approved event, such as employee self-referral, a human-reviewed notification or a claim event, then track the organisation's response.

Lane B is not a second determination file. It is an index showing whether identified evidence has been requested, received, source-verified and assigned to a human. The full evidence stays in the controlled claim system. AI does not weigh it, apply section 14, analyse the section 5A reasonable administrative action exclusion or draft an outcome.

The anti-inference dictionary is the control most teams miss:

Measure the organisation through process clocks: time to first human contact, time from consent to activation, overdue actions and unresolved bridge requests. Comcare's better-practice resource recommends collecting program data and using it for prevention and monitoring, but it also says organisations must recognise their employees' right to privacy and ensure that any collection, use or disclosure of their personal information is authorised under and compliant with the relevant privacy legislation. Outcome evaluation belongs in a separately governed, appropriately aggregated view, not a claimant score on the operational board.

What information is allowed to cross the bridge?

Nothing crosses automatically. A human bridge decision records the source, exact field, receiving lane, purpose, authority or consent relied upon, approver and date. The receiving user opens the original record rather than trusting an AI paraphrase.

Comcare's claims information for employers says employers hold personal and medical information not related to the compensable injury, that this should be held separately to the compensation file, and that employers need to be aware of obligations under the Privacy Act 1988 and manage this information in line with the purpose in which it was obtained. The two-lane design makes that purpose boundary visible rather than relying on users to remember it.

Where Australian Privacy Principle 3 applies, the OAIC's current APP 3 guidance says collection must be reasonably necessary for the entity's functions or activities, with a directly-related alternative for agencies. Sensitive information generally requires consent unless an exception in APP 3.4 applies. The OAIC describes valid consent as adequately informed, voluntary, current and specific, with capacity to understand and communicate consent, and says collection must use lawful and fair means. It also says that where an entity creates personal information with reference to, or generated, inferred or observed from, other information it holds, this is a collection of personal information and APP 3 obligations apply, giving artificial intelligence, automated decision making and data analytics as examples.

Participation consent and information authority are not the same field. Record what the employee agreed to participate in, what information may be shared, with whom, for what purpose and for how long. A human privacy practitioner must determine which privacy rules, exceptions and other obligations apply to the particular organisation and record.

This prompt tests the board for hidden profiling and uncontrolled transfers. A human claims lead, support owner and privacy reviewer must investigate every flag and approve any correction.

Red-team the de-identified two-lane board for [MATTER_ID].

Inputs:
[LANE_A_SUPPORT_ROWS]
[LANE_B_EVIDENCE_INDEX]
[BRIDGE_LOG]
[ANTI_INFERENCE_DICTIONARY]
[APPROVED_ROLE_MATRIX]

Flag any:
- claimant, injury, severity, credibility, causation, recovery or outcome score
- inference from ACCEPTED, DECLINED, WITHDRAWN or NO CONTACT COMPLETED
- clinical narrative or manager opinion stored in the support lane
- support-program data used to fill a liability-evidence gap
- automatic join, copy or summary across lanes
- transfer without exact source, purpose, approver, date and recorded authority
- AI-generated employee contact or substantive claim action
- internal service date presented as a statutory deadline

Return PROCESS CONTROL FLAGS only. Do not decide whether collection is lawful, whether consent is valid, what evidence proves or whether liability is accepted. Humans decide and document every response.

Fictional worked example

This scenario is fictional and de-identified. [EMPLOYEE_NAME] notifies [EMPLOYER_NAME] of psychological symptoms and later lodges claim [CLAIM_NUMBER]. The support lane records that [SUPPORT_OWNER_ROLE] offered a human contact through [CONTACT_CHANNEL], the employee accepted [SUPPORT_OPTION], and consent source [CONSENT_RECORD_ID] permits specified scheduling information to go to [APPROVED_RECIPIENT] until [CONSENT_END_DATE].

The support lane does not store the employee's account of workplace events, a diagnosis or the provider's clinical notes. The liability lane records that statement [STATEMENT_ID] and medical document [MEDICAL_RECORD_ID] were received and await human source review. It does not import the support choice.

A manager asks whether quick acceptance of psychological support suggests the condition is genuine and work related. The board refuses the inference and records no bridge event. A human claims decision-maker assesses the claim evidence under the SRC Act. The support owner continues the agreed process without receiving the liability view.

Do this Monday

1. Draw the boundary. Name the voluntary support system, claim system and separate statutory rehabilitation workflow, including each human owner.
2. Minimise Lane A. Keep only offer, contact, choice, consent scope, activation and next-action fields. Remove diagnosis and free-text manager observations.
3. Install the anti-inference dictionary. Make every support status display its prohibited meanings beside the field.
4. Create a human bridge log. Require source, field, purpose, authority, recipient, approver and date before information moves.
5. Ban person scoring. Test for sentiment, severity, credibility, causation, recovery and likely-outcome fields or proxies.
6. Use fictional failures. Test refusal, withdrawal, no contact, a late claim and a manager request to share clinical details before using de-identified operational records.

Bottom line

Early support should be fast, voluntary and human. Liability assessment should be evidence-based, authorised and separate. AI can track whether your process moved and expose where your organisation has not acted. It must never convert support behaviour into a view about the person or the claim.

This article is general information and education only. It is not legal advice, and it is not advice about any individual claim. Decisions under the Safety, Rehabilitation and Compensation Act 1988 are made by human decision-makers on the individual merits of each claim, and claimants have reconsideration and review rights in respect of determinations. Seek advice specific to your scheme and circumstances.

References

1. Federal Register of Legislation, Safety, Rehabilitation and Compensation Act 1988, Compilation No. 82, compilation date 1 July 2026. https://www.legislation.gov.au/C2004A03668/latest/text
2. Comcare, Intervene early. https://www.comcare.gov.au/safe-healthy-work/healthy-workplace/intervene-early
3. Comcare, Better practice: early intervention programs. https://www.comcare.gov.au/about/forms-pubs/docs/pubs/research/better-practice-early-intervention-programs.pdf
4. Comcare, Claims information for employers. https://www.comcare.gov.au/claims/employer-information/claims-information-for-employers
5. Comcare, Rehabilitation information for employers. https://www.comcare.gov.au/claims/employer-information/rehabilitation-information
6. Office of the Australian Information Commissioner, Chapter 3: APP 3 Collection of solicited personal information. https://www.oaic.gov.au/privacy/australian-privacy-principles/australian-privacy-principles-guidelines/chapter-3-app-3-collection-of-solicited-personal-information