← All outputs
Supporting Potential · Systems Scoping (Core) · Working Document

Where AI Can Assist — Documentation and Quality Intelligence

This document maps where AI can add genuine value to the 36 managed documentation controls, what each AI touchpoint does, and what the build looks like. The mapping is honest about sequencing: AI quality is bounded by data quality. Most of these applications require the basic automated workflows (Categories 1–5 in the managed documentation register) to be in place first — because without structured data, AI is reading noise.

The key distinction: Most of the 36 controls in the managed documentation register are automation — if/then rules, scheduled alerts, routing logic. Any well-configured case management system can do this without AI. AI adds value in a different register: detecting patterns humans miss, generating drafts humans review, and making judgements about priority and risk that simple rule engines cannot. This document covers only the AI layer — not the automation baseline.

Build sequence

The three tiers below are roughly sequential. Tier 1 can start now with existing data. Tiers 2 and 3 require the structured documentation workflows to be in place first — because they depend on shift notes and participant records containing meaningful, structured data rather than compliance tick-boxes.

Now — Tier 1
Intelligence on existing documents
Works on plan content (PDFs, Word documents) and incident reports — higher-quality structured inputs that exist today. No dependency on improved shift note quality.
6–12 months — Tier 2
Pattern detection across records
Works on shift note content at scale. Requires structured shift note format (Category 2 adaptation) to be in place — otherwise AI is reading free-text of variable quality.
12–24 months — Tier 3
Embedded generative assistance
AI embedded in the Connect interface, prompting workers in real time and generating drafts. Requires Connect configuration (Categories 1–4 automation) as the foundation and solid data quality from Tier 2 maturity.
The data quality gate

AI reads what it is given. If shift notes are "John had a good shift and enjoyed dinner," AI will identify patterns in good shifts and enjoyed dinners. The Category 2 adaptation in the managed documentation register — restructuring shift note format to include domain-based observation fields — is not just an operational improvement. It is a prerequisite for meaningful AI analysis. A shift note that captures: food texture, volume consumed, behaviour observations, health indicators, follow-up items produces data an AI can reason about. A shift note that captures: "settled and happy" does not.

This means the AI build for Tiers 2 and 3 is gated on the documentation quality improvements that precede it. Getting the structured workflows right first is not a delay to the AI strategy — it is what makes the AI strategy work.

Tier 1 — Intelligence on existing documents

1
Works on plan content and incident reports — can start now

These applications read structured documents that already exist at Achieve: support plans, mealtime management plans, health action plans, behaviour support plans, and incident reports. Input quality is relatively high because these are authored documents, not free-text observation fields. Output is drafts and summaries that a human reviews and approves — AI is not making clinical decisions.

Cat 5 — Staff readiness
Worker participant briefing generation
When a worker is first allocated to a participant, AI reads the active plans (support plan, mealtime management plan, health action plan, BSP, medication protocols) and generates a single worker-facing briefing: who this person is, what matters to them, what you need to know before your first shift, what to do if X. Structured, readable, specific to this person — not a PDF stack.
What it replaces

Currently: worker reads plans themselves (if they read them at all) or gets a verbal handover. AI generates the summary; worker still acknowledges the source plans. Clinician reviews the briefing summary before it is published to confirm accuracy.

What the build looks like

Plans exported from Connect as structured text → AI model with a fixed prompt template (participant briefing format) → output posted back to Connect as a worker-facing document attached to the participant record. Reviewed and approved by SM or clinical nurse before activation. Rebuild triggered whenever a source plan is updated.

Cat 5 — Staff readiness
Knowledge check generation from plan content
Rather than asking a worker to sign "I have read the mealtime management plan," AI generates 4–5 specific questions from the plan content — texture requirements, positioning, volume, what to do if the person coughs or refuses. Worker answers before first solo shift. Demonstrates comprehension, not just attestation.
What it replaces

A signature on a form. Legally equivalent but clinically meaningless. A worker who answers specific questions about this person's plan has demonstrably read it.

What the build looks like

Plan text → AI generates question set → questions delivered in Connect or a linked form → answers recorded against the worker-participant pairing. AI generates questions at plan update, not on a fixed schedule. Wrong answers prompt re-read of the relevant plan section, not an automatic fail.

Cat 3 — Escalation pathways
Incident classification and Commission notification drafting
When a worker submits an incident report in Connect, AI classifies the incident against NDIS Commission reportable incident categories (death, serious injury, abuse or neglect, unauthorised restrictive practice, sexual misconduct, unlawful physical contact). Where the incident is notifiable, AI drafts the Commission notification using the structured fields from the incident report. SM reviews and submits — they do not write from scratch.
What it replaces

Manual classification (error-prone, especially for complex incidents) and blank-page notification drafting (time-consuming, inconsistent quality across writers). Four matters currently open with the Commission — consistent, accurate notifications reduce the risk of Commission escalation.

What the build looks like

Incident report submitted in Connect → AI reads incident description and structured fields → classifies as reportable/non-reportable with confidence level and reasoning → if reportable, generates draft notification in Commission format → SM and OCG review, edit, and submit. AI flags where it is uncertain about classification for human determination.

Cat 2 — Completeness
Intake plan completeness check against Practice Standards
When a new participant's support plan is uploaded or completed in Connect, AI reads the document and checks whether it addresses required domains: communication, health and wellbeing, behaviour support (or documented assessment of need), mealtime management (or documented assessment of need), goals, risk assessment, emergency management. Gaps are flagged before the plan is approved — not discovered at audit.
What it replaces

Manual review by SM or practice leader. AI is faster and applies the same checklist consistently — human reviewer confirms flagged gaps and makes clinical decisions about whether they represent genuine omissions or justified exclusions.

What the build looks like

Support plan document → AI reads against a fixed domain checklist → outputs structured gap report: domain addressed / partially addressed / not addressed / not mentioned. Clinician or SM reviews and approves before plan is marked as active. Checklist is maintained by quality team — AI applies it, humans define it.

Cat 1 — Document currency
Plan currency risk scoring
Beyond "this plan expires in 30 days" — AI reads recent incident data and shift note summaries alongside the plan content and scores which plans are most likely to have drifted from reality, independent of their review date. A plan that was reviewed two months ago for a participant who has had three incidents and a hospital admission since then is higher risk than a plan that hasn't been reviewed but whose participant has been stable for 12 months.
What it replaces

Calendar-only review scheduling. When 15 plans come due in the same month, AI helps the quality team know which three to prioritise — not just which three are oldest.

What the build looks like

Monthly run: AI reads plan review dates + incident count since last review + hospitalisation records + health flag count from shift notes → generates a priority-ranked list for the SM or quality team. Output is a recommendation, not a decision — the human determines whether to bring forward a review. Requires incident data and some shift note data from Connect; does not require full structured shift note reform.

Tier 2 — Pattern detection across participant records

2
Works on shift note content at scale — requires structured note format first

These applications read across dozens or hundreds of shift notes per participant over time, looking for patterns no individual reviewer would catch. They only work if shift notes contain meaningful, structured observations — which requires the Category 2 adaptation in the managed documentation register to be implemented first. Without structured observation fields, AI is reading "good shift" at scale.

Cat 2 — Completeness
Shift note quality analysis
AI reads shift notes not just for existence but for content quality — does this note contain meaningful clinical observation? A note that mentions food texture, volume, and a concern about swallowing passes. A note that says "good shift, John had dinner" fails, even though it is technically filed. Quality score per note, per worker, per site — visible to SM weekly.
What it enables

The current compliance model says "note filed = done." The quality model asks "note filed AND contains what matters = done." SM can see which workers consistently produce thin notes and intervene with coaching rather than waiting for an audit to surface the pattern.

What the build looks like

Shift notes exported from Connect → AI evaluates against a rubric (health observations present, follow-up items documented, plan-specific observations included) → quality score per note → weekly summary report per site for SM. Rubric is maintained by the quality team. AI applies it consistently; quality team reviews outliers.

Cat 2 — Completeness
Health pattern detection — trends humans miss
AI reads across 30–60 shift notes for a participant and flags patterns that don't appear in any single note: "weight described as declining in 5 of the last 12 meal notes," "pain references appear in 4 notes this month across different workers," "food refusal mentioned 6 times in 3 weeks." Each individual note looks fine. The pattern is only visible at scale.
Why this matters

This is the gap behind the health monitoring quality review flag in the exec summary. Four deaths in care, two with unanswered questions about whether lead-up health indicators existed in the weeks before acute presentation. Shift notes may have contained those indicators. AI can read them in a way no individual SM can at scale across 700+ participants.

What the build looks like

Weekly batch: AI reads last 60 days of shift notes per participant → identifies semantic patterns in health, nutrition, mobility, behaviour, and social domains → flags participants with concerning patterns to SM and clinical nurse. Output is a flag with evidence ("here are the 5 notes that show this pattern"), not a diagnosis. Clinical nurse reviews and determines response.

Cat 1 — Document currency
Plan-reality drift detection
AI compares what a participant's active plans say should be happening with what shift notes record actually happening. The plan says pureed food — shift notes describe normal textures. The plan says independent in personal care with verbal prompting — shift notes describe full physical assistance for six months. The plan is technically current; the reality has drifted. AI catches the gap between the document and the practice.
What it replaces

Annual plan review as the only mechanism for catching plan-practice drift. By the time a plan is reviewed, the practice may have been misaligned with the plan for 11 months. AI runs this check continuously and flags interim review when drift is detected.

What the build looks like

Plans from Connect → AI extracts expected practices per plan domain → shift notes from same period → AI compares described practice against plan expectations → drift flags generated for clinician review. False positive rate expected to be meaningful — requires clinical review before any action is taken. The value is in surfacing cases for clinical attention, not in automated plan updates.

Cat 3 — Escalation pathways
Alert prioritisation and triage
When multiple documentation alerts fire simultaneously across sites — three expired plans, two incomplete shift note runs, one overdue incident report — AI ranks them by participant risk rather than by type or date. An expired mealtime management plan for a participant with a known dysphagia risk is more urgent than an expired emergency management plan for a participant with no identified health needs, even if the emergency plan is older.
What it replaces

A flat list of alerts that a SM must sort themselves. In high-volume sites, SMs receiving multiple simultaneous alerts without prioritisation will address the easiest ones first, not the most important ones.

What the build looks like

Alert generated by rules engine → AI reads participant risk profile (health complexity, recent incidents, plan currency history) → assigns priority score to alert → SM receives prioritised alert list with brief rationale ("prioritised because participant has active dysphagia risk and recent incident cluster"). SM can override priority — AI is informing, not deciding.

Cat 4 — Workflow triggers
Hospitalisation review preparation
When a participant hospitalisation is recorded in Connect, AI reads the full participant record and generates a structured summary for the clinical nurse or SM conducting the discharge review: medication history, last clinical plan reviews, recent incident pattern, upcoming plan expiry dates, GP contact history, and any health pattern flags from recent notes. The reviewer arrives prepared rather than pulling this manually.
What it replaces

The clinical nurse or SM reading through 60+ shift notes, three plan documents, and an incident log before the discharge review meeting. This typically doesn't happen — the review proceeds without full context. AI makes full context available in 30 seconds.

What the build looks like

Hospitalisation event recorded in Connect → AI immediately reads participant record (plans, recent notes, incidents, medication, health flags) → generates structured briefing in fixed format → delivered to SM and clinical nurse within minutes. Briefing is a summary tool, not a clinical assessment. The clinical nurse makes the assessment.

Tier 3 — Generative assistance embedded in Connect

3
AI in the workflow — embedded prompts and draft generation at the point of care

These applications embed AI directly into the Connect interface so that workers, SMs, and clinicians are prompted and assisted in real time, not after the fact. This tier requires the full managed documentation foundation (Categories 1–5) to be in place and functioning, and requires solid data quality from Tier 2 maturity. It is the highest-value tier and the furthest away.

Cat 2 — Completeness
Contextual shift note prompting
As a worker completes a shift note in Connect, AI surfaces participant-specific prompts based on the active plans: "This participant has a mealtime management plan — please note food texture, volume, and any swallowing concerns," or "This participant has an active bowel management plan — was a bowel movement recorded today?" Prompts are specific to this person on this shift, not generic reminders. Worker can dismiss, but dismissal is logged.
What it changes

The current shift note prompt is a blank text field, or at best a generic observation structure. This replaces it with a participant-aware guide that pulls from the active plans to tell the worker what to document. It makes the plan visible at the point of care — not just filed somewhere in Connect.

What the build looks like

Connect shift note interface → AI reads active plans for this participant → generates contextual prompt list at note-opening → prompts displayed alongside free-text field → responses structured by domain → AI quality scores the completed note in real time and flags before submission if key domains are missing. Requires Connect front-end integration or a Connect-adjacent interface that feeds back into the record.

Cat 4 — Workflow triggers
Health change detection without explicit flagging
AI monitors shift note language in real time and identifies health concern patterns that a worker has described but not formally flagged: "seemed uncomfortable at dinner," "slower than usual," "didn't finish meals three days this week." Each note, individually, wouldn't trigger an alert. The pattern across four notes does. AI flags to SM without requiring the worker to have explicitly classified it as a health concern.
Why this matters for safety

Workers describe what they observe. They don't always know when a pattern of observations constitutes a clinical concern. AI bridges that gap — reading the observation language and applying clinical pattern recognition that a frontline worker is not expected to have. This is where AI adds something that neither better documentation nor better training can fully replace.

What the build looks like

Continuous note monitoring → AI applies NLP across rolling 30-day note window per participant → pattern detection across health, behaviour, nutrition, mobility domains → threshold flag to SM and clinical nurse with evidence ("3 of the last 5 notes contain language associated with appetite or eating difficulty") → clinical nurse reviews. False positive management is critical: alert rate must be calibrated so SMs respond to alerts rather than ignoring them.

Cat 1 — Document currency
Plan draft generation for clinician review
When a plan review is due, AI drafts an updated plan from: the previous plan version, recent shift notes, incident history, and any new clinical information on file. The clinician reviews and edits a populated draft rather than beginning from a blank template. The draft highlights what has changed since the last version and what AI is uncertain about. Review time reduces significantly; clinician attention goes to what matters, not to reformatting.
What it replaces

Blank-template plan drafting, which requires the clinician to re-synthesise information they partially have from memory and partially need to look up. AI does the synthesis; the clinician does the clinical judgement and participant engagement.

What the build looks like

Plan review task triggered in Connect → AI reads previous plan, last 90 days of shift notes, incident history, and any clinical reports filed → generates draft plan in the required format with inline notes ("no shift note references to swallowing difficulty in the last 90 days — confirm texture recommendation is still appropriate") → clinician reviews, edits, and approves. AI draft is never published without clinician sign-off. Participant review meeting still required.

Cat 3 — Escalation pathways
Escalation routing with context
When a gap triggers an escalation, AI generates the escalation message with relevant context rather than a bare alert. Instead of "Mealtime management plan expired — Action required," the SM receives: "Mealtime management plan for [participant] expired 3 days ago. This participant had a meal-related incident in October 2025. Their plan is authored by [clinician]. Three workers are currently rostered to support them. Priority: high." The person receiving the alert has what they need to act immediately.
What it changes

Alerts without context are often deprioritised because the recipient doesn't know how urgent the situation is. Context-rich alerts are acted on faster. This applies to the same escalation pathways in the managed documentation register — AI adds the context layer to automation-generated alerts.

What the build looks like

Alert triggered by rules engine → AI reads participant record for relevant context → generates contextualised alert message with priority reasoning → delivered to SM via Connect notification or email. Connect front-end displays context summary alongside the action required. SM can see the full participant context from the alert without navigating to the participant record manually.

What Achieve needs before any of this

The Connect API question

All of these applications require programmatic access to data held in Connect — shift notes, incident reports, participant records, plan documents. Whether Connect provides an API or data export mechanism suitable for AI integration is a technical question that needs to be answered before any build is scoped. If Connect does not support API access, the build approach shifts to batch export and processing rather than real-time integration — which affects Tier 3 specifically, and changes the latency on Tier 2 alerts.

Connect went live in June 2026. The implementation team (Karen Moore-Evans) would know whether API access is available or on the roadmap. This is the single most important technical prerequisite to confirm.

Data governance and consent

All of these applications process participant health information, shift note content, and incident data. Before any AI model is applied to this data, Achieve needs to confirm: which AI providers are acceptable under their privacy and data governance policies, whether participant consent (or the existing service agreement) covers AI processing of health-adjacent data, and how audit logs of AI outputs are maintained for regulatory purposes.

The NDIS Commission does not currently have AI-specific guidance for providers, but the NDIS Practice Standards on information management apply. Any AI system processing participant data must meet the same security and privacy requirements as any other system holding that data. Most providers in this space use closed, self-hosted, or privacy-preserving AI deployments rather than sending participant data to general-purpose AI APIs.

Tier 1 applications can be structured to process de-identified or minimally-identified data in many cases (plan documents with identifying fields removed). Tiers 2 and 3 require fuller data access and need the governance framework in place first.

The near-term opportunity that doesn't wait for any of this
Two Tier 1 applications can be piloted now with minimal build cost and no dependency on Connect API access or improved documentation quality: worker participant briefing generation (upload a plan PDF, AI generates the shift briefing, SM reviews and approves) and incident classification and Commission notification drafting (copy incident report into AI tool, AI drafts notification, OCG reviews and submits). Both use existing document inputs and generate human-reviewed outputs. Neither processes raw shift note data. Neither requires Connect integration. They can be run manually as a workflow with AI as a tool used by an administrator, with the output filed back into Connect by hand. Piloting these two applications builds organisational confidence in AI-assisted documentation workflows at low risk — and produces a dataset of quality-reviewed AI outputs that informs the Tier 2 specification.