Purpose
System cards are a key tool for responsible AI development. They provide a clear, accessible overview of an AI system’s purpose, capabilities, and limitations, making it easier for users, developers, and policymakers to understand and evaluate the technology.
Introduction
This document provides an overview of Minikai, including its purpose, key features, limitations, and intended use within disability and aged care settings. It explains how the system works, what it can and cannot do, and the principles guiding its development and use.
Minikai is not an all-encompassing intelligence. Instead, it is a tool designed to help users work more effectively and efficiently. Therefore, an understanding of how to use it is key to its successful implementation.
This document also outlines Minikai’s roadmap for continual improvement, ensuring the assistant evolves alongside sector needs.
What is Minikai?
Minikai is an AI system developed to assist disability and aged care professionals in performing their work more effectively and efficiently.
Each Mini the user selects has access to the data for that person. The Mini can quickly retrieve and analyse a wide range of data on a single person, and the group Mini can answer questions on a group of people. Minis can also help with writing tasks such as care plans, assessments, or writing case daily notes using speech-to-text.
A key capability is that the underlying data used to answer questions will be provided to the user when the Mini responds, enabling the user to validate the response. This is why Minikai is described as a human-in-the-loop generative AI system.
The ability to draw on large volumes of data, stored in a secure environment through conversation with Minis unlocks numerous use-cases that span multiple roles such as frontline workers, allied health professionals, coordinators, managers, compliance teams, and executives looking for high-level insights across their customer base.
How is Minikai different from other AI products?
Data privacy & safety
Data about the person under care is securely stored within the platform, enabling safe and effective use of AI.
Data integration
Minikai integrates with complex organisational data, intelligently navigating it and drawing from it to answer user questions while providing the underlying data in its response.
Custom features
Custom features are developed to support staff in the aged care and disability sectors, such as the ability to ask time-based questions or to query and aggregate data across a set of Minis (for example, everyone supported within a facility).
Team experience and passion
We have a diverse team who have countless hours working with staff in the disability and aged care sectors. Our team members include those who have lived experience with a disability, experience working in these sectors, and family members who receive support in these sectors. Our experience and personal connection enables us to provide a tailor-made product. We are actively building on and improving these features to meet emerging needs.
AI Guiding Principles
Professional ownership
Users maintain control and responsibility for final documentation and decisions. No AI-generated content is submitted without user validation.
Assistive, not autonomous
The AI supports workflows but never replaces human judgment or agency. It assists with decision-making while avoiding expert-level judgments.
Transparent outputs with citation
All outputs are traceable, with clearly linked source data to ensure verifiability. The AI is faithful to the underlying data in the response it provides.
Person-centred by design
The system is built to support care that is respectful, inclusive, and focused on the wellbeing of the person under care.
Continual improvement
The system is continually tested for its applications in the disability and aged care sector. This informs improvement to better meet industry needs.
Intended Use
The Mini provides ready access to information on individuals and can assist with writing tasks. However, it is important to be mindful that Minis can make mistakes and their output should be checked for validity.
Writing
Drafting notes, incident reports, care plans, and assessments.
Assisting with note creation through speech input.
Writing emails to other providers or family.
Writing reports to apply for funding for a person, drawing on customer data from CMS systems to provide as evidence of changing support needs.
Comparing existing data such as incidents or case notes to update documentation such as assessments or care plans to meet the needs of the person.
Retrieval of information
Retrieving and synthesising structured and unstructured data from CMS platforms.
Summarising client plans, assessments, and communication logs.
Analysis
Reviewing documentation for completeness or identifying missing records.
Highlighting inconsistencies, duplication, or expired information in records.
Aggregated analysis answering questions on a large number of people.
Out-of-scope use
Minikai is not intended for:
Clinical diagnosis, treatment planning, or expert-level judgements without human review.
Fully autonomous decision-making.
High-stakes decisions without human verification (e.g., medication management or risk alerts).
Tasks requiring precise mathematical or quantitative calculations (e.g., detailed counts, complex statistics, or financial reconciliations).
Key System Capabilities
Finds the right information: Searches across client records, plans, notes, and assessments by keyword or role, with the ability to filter by time period (e.g., week or month) and record type (e.g., case notes, assessments, charts). (Retrieval engine with smart filtering capabilities).
Summarises for you: Turns long documents or multiple notes into clear summaries. (Chat).
Shows the sources: Every answer includes citations and lets you open the original record to check the detail. (Citation).
Helps you write: Drafts case notes, care plans, reports, and emails with grammar and completeness assistance. (Chat).
Lets you speak instead of type: Converts voice to text for faster and easier note creation. (Speech-to-text).
Translates language: Helps translate notes or communications to and from many languages. (Translation).
Crowd insights across people: Ask questions across a set of people (e.g., all in a facility) to compare, analyse, or surface patterns. (Group chat).
Data Source Integration
Minikai can connect to organisation’s data in two ways: either through a purpose-built integration with the customer’s CMS, or by allowing the customer to securely push their own data into Minikai through our APIs. In both cases, all information remains within Minikai’s secure cloud infrastructure, hosted in Australia, and is never shared externally.
Information Types
Information types Minis have access to is inclusive of but not limited to the list below. Please note that data is often filtered for “active” records. For example, the Mini can retrieve active wound charts but not expired or archived wound charts or personal support plans. Check with your organisation what specific information types the Minis have access to.
Case Notes or Progress Notes.
Assessments (such as Functional Capacity Assessments or Risk Assessments).
Personal Support Plans.
Incidents.
Associated Providers.
Health Records (bowel charts, wound charts, etc).
Contacts (such as GP or family members).
Legal Information (consents and guardianship).
Equipment.
NDIS Goal Achievement Plan.
Residential Information.
Activities Chart.
New Admission Checklist.
Policies and Procedures.
The following table demonstrates the effect of a system change that was made to better identify what type of information is relevant to the user’s question. The percentage of prompts where the correct information type was included in the response increased from 76% to 88%.
|
| Before Uplift | After Uplift |
Information Type | Prompt Tests | Did cited documents have the correct information type? | Did cited documents have the correct information type? |
Alert | 3 | 100% | 100% |
Assessment | 7 | 86% | 86% |
Associated Provider | 3 | 0% | 67% |
Bowel Chart | 4 | 100% | 100% |
Client Contact | 5 | 80% | 80% |
Consent | 4 | 75% | 75% |
Details | 5 | 60% | 80% |
Equipment | 4 | 75% | 100% |
NDIS Goal Achievement Plan | 4 | 75% | 100% |
Note | 4 | 50% | 100% |
Personal Support Plan | 7 | 100% | 86% |
Total | 50 | 76% | 88% |
Testing & Evaluation
Minikai evaluates the performance of the Minis to understand limitations and inform continuous improvement. Minis are tested against the accuracy of their responses, and how useful they are in assisting users. This process includes:
Collecting user feedback: We work with users to resolve issues and help them use the Minis effectively.
Monitoring user engagement: Where users are frequently using or not using the Minis can indicate whether they are assisting the users as intended.
Automated test-prompts: Running regular tests through an evaluation pipeline, where Minis are evaluated on their ability to answer prompts that align to common use-cases.
Analysing real conversations: This includes both manually reading chats, and analysing chats at scale using statistical techniques to understand overall trends.
Structured user-testing: Running structured testing with users in organisations to better understand return on investment (ROI) and Mini behaviour.
Known Limitations and Best Practices
The following table describes limitations of the Minis and best practice in using the Minis. It should be noted that through continually improving the product, these limitations will be reduced over-time and this list will be updated accordingly.
Limitation | Best Practice |
Specific information needed: The Mini may not always interpret broad or vague requests correctly. For example, asking for “plans” may return the wrong type of document, whereas “Therapy NDIS goal plan” is more precise. | Phrase questions as specifically as possible, naming the document type or detail you need. |
Incomplete lists: When large sets of results are requested (e.g., all case notes for a year), the system may only return part of the set, causing items to be missed. | Break large requests into smaller ones (for example, a week at a time rather than over a year) for more complete coverage. |
Record count limits: The Mini can retrieve about 50 records at a time. Requests larger than this may be cut off. | Narrow searches using shorter timeframes or specific record categories. |
Environment differences: Data in different environments (such as a testing sandbox or a live system) may be different. One environment may have less up-to-date records. | Check with your organisation what environment the Minis are connected to. |
Metadata gaps: Certain flags or metadata (e.g., HISS indicators for plan types) are not yet linked to certain records, reducing visibility. | Be aware of gaps and double-check data directly if a flag or tag is missing. |
Expired or inactive records: For most information types, Minis are only given access to records marked as “Active” or “Current”. They may not have access to expired plans and archived wound charts to avoid presenting outdated information. This means they won’t appear even if requested. | Check with your organisation what filters are applied to the data Minis have access to. |
OCR issues: Some documents, particularly scanned PDFs such as legal or consent forms, may be unreadable by the system if OCR (text recognition) fails. | If results seem incomplete, review the original documents to check scan quality. |
Counting and math errors: Large tallying tasks (e.g., “How many bowel movements this month?”) may be over or under counted due to AI limitations with calculations. An example would be returning 37 of 42 bowel charts over a two-week period. | For precise counts, check individual records or export data into structured reporting tools. |
Data source visibility: Minis can only search the data sources that your organisation has connected. If certain records are not integrated (e.g., some plans or assessments), the cannot retrieve them. | Confirm with your organisation which data sources are integrated into Minis, and check directly for data not yet available. |
Human-in-the-loop use: Minis are decision-support tools, not a replacement for professional judgment. Outputs may contain errors, omissions, or outdated content. | Always review AI outputs before submission, verify citations and sources, and report any errors or gaps. |
Search memory limits: Minis remember the answers they previously provided to the user as part of a conversation, but not the full set of records they searched to prepare those responses. If the first response doesn’t include what you need, refine your question to provide more relevant results. | If you don’t get the result you’re after, rephrase or refine your request and ask the Mini to search again. For example, if you asked the Mini to provide behaviour charts over the past six months and no instances of poor behaviour were mentioned, instead ask “provide instances of poor behaviour in behaviour charts over the past six months”. |
Using previous responses: Minis can provide new answers based on their last response, but they don’t automatically re-use that information unless you specifically ask. If you just ask a broad follow-up, the Mini may search data again instead of building on what it already provided. | If you want the Mini to build on a previous response, be explicit. For example, instead of saying “Summarise activities they have been doing”, say “Summarise your last response listing the activities over the past three months.”
|
Roadmap Features
Minikai continues to expand with new features designed to support care and disability teams in practical, person-centred ways. Our roadmap is never static, we continuously adapt and refine priorities based on real user engagement, sector needs, and feedback. This means features evolve with you, ensuring Minis remain relevant, safe, and effective in helping staff deliver better outcomes across aged care, disability support and in-home care.