🔒 Permissions: This article applies to everyone who uses Minikai, in any role. No special permissions are needed. Learn more about roles →
When you ask a Mini about someone you support, the response informs real decisions about a real person's care. That is a responsibility we take seriously, and it shapes how Minikai is built, tested, and improved. This article explains what we consider a quality answer, how we measure it, how changes are tested before they reach you, and the part you play in using Minikai safely.
Every response is grounded in the person's records
A Mini does not answer from general knowledge or the internet. It searches the records on that person's timeline and builds its response from what it finds there. Those records may have been synced from your care management or client management system, imported from documents or spreadsheets, or created directly in Minikai. The record types vary from organisation to organisation, for example progress notes, care plans, assessments, incident reports, and charts.
Wherever a record came from, it is treated the same way: indexed on the person's timeline and searchable by their Mini. Every response includes numbered citations, so you can open the exact records the information came from and check it for yourself. For more on how this retrieval and grounding process works, see How Minikai's AI Works.
What makes a quality answer
We hold every response to the same standard, defined by four things:
Grounded. Everything in the response is supported by the records on the person's timeline, not general knowledge or assumption. We call this groundedness, and it is the foundation of everything else.
Cited. Every claim can be traced back to the record it came from, so you never have to take the response on faith.
Precise. The response answers the question that was asked, without adding detail the records do not support or padding the answer with things you did not ask about.
Honest about gaps. If the records do not contain the answer, a quality response says so. "There is nothing on the timeline about swallowing difficulties" is a better answer than a confident guess.
How we measure quality
These are not qualities we assume, they are things we test. We maintain an evaluation suite built from realistic care-sector scenarios: questions a support worker, nurse, or coordinator would actually ask, run against records like the ones organisations hold, each with a known correct answer drawn from those records.
Responses are scored against that expected answer, including whether the response is grounded in the right records, whether it is precise, and whether it acknowledges gaps rather than filling them. Because the same scenarios are run again and again over time, we can see whether quality is holding, improving, or slipping, rather than relying on spot checks or anecdote.
Evaluation before every change
Changes to the AI, including upgrades to the underlying model, are run against the evaluation suite described above before they are enabled. A change only goes ahead if response quality is maintained or improved against those benchmarks.
We maintain a change log of the AI models and behaviour in use, and we announce meaningful changes through our release notes.
Impact assessment and safety review
Minikai operates a certified AI management system under ISO/IEC 42001, the international standard for the responsible management of AI, alongside ISO/IEC 27001 for information security. In practice, this means every material change to how the AI behaves is also reviewed for its effect on the people whose information Minikai holds, separately from whether it improves response quality. See How Minikai keeps your data safe and compliant for our full certifications and compliance position.
Impact assessment. New or materially changed AI behaviour goes through an AI impact assessment before rollout, weighing privacy, fairness, and the risk of over-reliance on AI output.
Honest about what's still maturing. Structured bias and fairness testing is an area we are actively building out, and it is not yet as mature as our quality evaluation suite.
What Minikai is for, and what it is not
Minikai is an administrative and information management tool. It helps you find information, prepare documents, and capture records, and it is not a medical device, therapeutic good, or clinical decision support system. It must not be used for diagnosis, screening, clinical advice, treatment recommendations, or medication decisions.
Clinical decisions, care planning, and treatment decisions always sit with qualified people, not with the AI. A Mini can gather the history and draft the paperwork, and a person decides what happens next. This principle, human in the loop, runs through the whole product and is reflected in our Terms & Conditions.
The part you play
Quality is a shared responsibility, and a few habits make a real difference:
Review before you rely on it. Read anything a Mini produces before you submit it, share it, or act on it, the same way you would review a colleague's draft. See Human in the Loop for more on what this means in practice.
Check the citations. If a detail matters, open the cited record and confirm it says what the response says it does.
Ask precise questions. "What changes have there been in Maria's mobility since her last physio review?" will get a more grounded response than "tell me about Maria".
Flag anything that looks wrong. If a response misstates a record or misses something important, tell us through the Help Desk. Reports like this feed directly into our quality process.
⚠️ Always review before acting: A Mini's output is a draft or a summary, never a decision. Before you submit a record, send an update, or act on a response, review it and apply your own professional judgement. You remain responsible for what you do with the information.
ℹ️ Good to know: Your organisation's data is stored and processed in its selected data region and is never used to train AI models. For how information is protected, see AI and Privacy and How Minikai keeps your data safe and compliant. Our certifications, sub-processors, and security documentation are published on the Trust Centre, and the Minikai System Card describes the system in more technical depth.
Need help?
If you have questions about AI quality or safety, or you want to raise something a Mini got wrong, message us through the help icon in the bottom-right corner of Minikai. We're here to help.