Secure Innovation
Book a demo
Menu

Secure Innovation Governed AI for regulated work

Your AI pilot didn't fail. It stopped at the risk review.

Radmis is built to get it through. AI runs as a permissioned user, on your premises or your cloud. Every action has an owner, an approval and a trail your auditor verifies without us.

In two weeks your risk owner has a go, fix or stop answer, with the evidence behind it. Defined outcomes at a fixed price. Stop after any stage.

Built for organisations including NSW Health · NSW Department of Education · NSW Telco Authority · Optus

What the risk review asks

Five questions decide whether your AI goes live.

Most pilots stall because nobody can answer them with a record. The platform keeps the record for each one.

  1. 01Who owns it?
  2. 02What may it do?
  3. 03What did it do?
  4. 04Who checked?
  5. 05How do we stop it?

01 Who owns it, and for what?

An owner on record, before it runs.

The AI Register holds one entry per AI system: its owner, purpose, risk band and review date. A check finds any AI co-worker or workflow that is not registered, registers it and raises a ticket.

02 What is it allowed to do?

AI is a user. Not a superuser.

An AI co-worker is an account with an access level, on the same access tree as your people. It reaches your data only through two controlled entrances, and only through calls you have registered. A call above its level does not exist for it.

03 What did it do?

A trail your auditor verifies without us.

Every change to audited data is recorded by the database itself in a chain of SHA-256 hashes, so a change made outside the screens is recorded too. Each AI turn is kept with its request, its response, the tools it called and its sources.

Your trail on this page · held in your browser, sent nowhere

    04 Who checked?

    The work waits until a person decides.

    A workflow stops at its approval step until a person decides, and the decision is recorded with their name. A workflow that mails what an AI wrote cannot be saved unless an approval stands on every path.

    A co-worker's work is held for approval

    05 How do we stop it?

    One switch stops the AI.

    The AI Kill Switch stops the gates, the models and the classifier from one screen. Each flip is recorded and mailed.

    This page's switch · not yet tested

    Why it passes review

    Most of what you need is automation. Not an agent.

    The less of the work a model does on its own, the less there is to defend at the risk review.

    Most of the work

    If it repeats, automate it.

    An LLM builds the automation quickly. Once it is built, it runs without one.

    The edge cases

    If it needs judgement, reason.

    Keep the LLM for the complex cases your current systems cannot cover.

    Always

    Guardrails, so it is compliant.

    Access levels, human approval and an audit trail on both. The compliance report comes out of the same record.

    The demo, step by step

    From an email to a working register. Four minutes.

    1. A compliance head emails Eva, an AI co-worker, from a train.
    2. Eva asks two questions.
    3. Eva writes three short text files. Not code.
    4. Three working, access-controlled screens.

    We run this build live when we meet.

    The Radmis AI Platform home screen: a menu of tickets, dashboard, co-workers and admin, with cards for ticketing, the site model, users and access, and the audit log.
    The platform's home screen
    How one request travels
    1. YouDescribe it in an email
    2. AI co-workerAsks, then writes a text file
    3. GateChecks the access level
    4. Sketch filePlain text. No code runs
    5. PlatformBuilds the table and screens
    6. RecordAudit trail and reports

    Outside the platformThe only way inInside, with your data

    See the whole file Eva wrote for the form
    @sketch 1
    @record
    @table gifts
    @width 900
    
    The gift:
    'Received by' fullname*: text | 'Email' email: text | 'Department' department [select: finance, operations, sales, board]: code
    'From' giver*: text | 'Date received' received* [date]: date | 'Value' amount*: money
    'What it was' description* [textarea]: long
    
    The decision:
    'Manager' manager: text | 'Decision' decision [select: pending, kept, returned, donated]: code
    'Why' reason [textarea]: long

    Nobody asked for what came with it

    • Access and logging

      From the first request.

    • Privacy

      Personal data found and inventoried.

    • Retention

      Old rows removed on schedule.

    • Audit

      Every change in the hash chain.

    AI stays outside. Your data stays in. See the security design

    What you can do with it

    Ten jobs it takes off the risk register.

    1. 01Get stalled AI pilots past the risk committee
    2. 02Replace spreadsheets with governed registers
    3. 03Use frontier AI on sensitive data
    4. 04AI-drafted letters with human sign-off
    5. 05Personal data requests on a 30-day clock
    6. 06APRA CPS 230 risk and service providers
    7. 07Audit and regulator evidence on demand
    8. 08AI incidents, complaints and appeals
    9. 09A policy and knowledge assistant
    10. 10Inbound mail triage on your own servers

    Track record

    Twenty-six years of systems that had to survive audit.

    Every capability in the platform was built for a real system first, in health, government, telecommunications and payments.

    • 26 years

      of research and development behind the platform.

    • 100+

      deployments of data management systems and services.

    Telecommunications · infrastructure audit

    3,000+ sites. Three auditors instead of eight.

    Problem
    No system for the job, and too many vendors whose products had never been integrated successfully.
    What we built
    An interim audit platform with secure mobile tools and automated compliance reporting.
    Result
    The interim system did the work. The incumbent vendor then used it as the working model and shortened its own implementation.
    As reported by the Network Infrastructure Program Lead, telco operator

    Government · Department of Education

    Weeks of reconciliation, now running in the background.

    Problem
    A limitation in the legacy ERP system, and several disparate systems that did not talk to each other.
    What we built
    A workaround that connects those systems using AI, automating policy-driven incentives and payroll adjustments for thousands of employees.
    Result
    Reconciliation that took weeks now runs in the background, with fewer errors.
    As reported by the Director, Workforce Policy Division
    • 200+ hours of manual data entry saved in radiology
    • 53 million payment transactions processed
    • 50 government health systems built to pass audit

    Read the case studies

    Technology partners

    • AWS Partner
    • Microsoft Azure
    • Akamai

    How we engage

    Five stages. Stop after any of them.

    We agree what done means, what it costs and what evidence comes with it before work starts. Then we carry the delivery risk.

    1. 01Two weeks

      Unblock

      A controls document your reviewers can sign, and a go, fix or stop answer.

    2. 02Three to four weeks

      Define

      The outcome, acceptance criteria, access boundaries and a fixed price.

    3. 03Eight to sixteen weeks

      Build

      Fixed price. Weekly evidence, not status reports.

    4. 04At go-live

      Accept

      Your risk owner runs five checks. Failures are fixed at our cost.

    5. 05Optional

      Operate

      Monitoring, a monthly evidence pack and kill-switch drills.

    Coverage

    Mapped, not certified.

    Thirteen regulations and guidelines, point by point, against the code on 28 October 2026. We show what is not met yet, because your reviewers will ask.

    MetIn partNot yet

    • All thirteen191/365
    See each of the thirteen, with its count
    • APRA CPS 23020/35
    • APRA CPS 23411/14
    • Privacy Act 198817/22
    • OAIC APP guidelines11/12
    • DTA AI policy v2.17/19
    • DTA impact assessment tool20/42
    • DTA assessment guidance6/18
    • NSW AI Assessment Framework7/21
    • NSW AI Operational Policy v1.119/41
    • NSW Circular DCS-2026-022/5
    • NIST AI RMF 1.026/52
    • EU AI Act12/42
    • GDPR33/42

    Ask for the full map, point by point

    Product roadmap

    What is still open.

    No dates are promised on this page. Ask us where each one stands.

    • One stop per AI system
    • Ask before a write
    • Explanation on request
    • Co-worker checks in Platform Health
    • An AI mark on an account
    • A six-month retention floor
    • Other regulators' fields on the AI Register

    Not ready to talk yet?

    Fifty questions your AI program will be asked. How many can you answer with evidence?

    Tick the ones you cannot answer yet and see which team owns each one: legal, privacy, compliance, security, risk or commercial.

    Five to ask any AI vendor. Including us.

    Can our auditor verify the audit log without your help?

    Yes. The trail is a hash chain written by the database itself. A change made outside the screens is recorded too.

    If an agent is tricked by a prompt, what can it write, delete or read?

    Only what its access level opens, through calls you have registered.

    When AI builds a screen, is any code generated?

    No. The AI describes and the platform builds.

    Where does personal data go before it reaches the model?

    To a local model first. Before an outside model, through the sanitiser where it is set.

    How long before the first governed use case is running?

    About an hour to install. A first register can be on screen the same day.

    For your risk committee

    For government · PDF · 8 pages

    Governed AI for government: from policy to practice.

    For accountable officials. Monitoring AI policy, and a secure, compliant gifts register as the worked example.

    Download

    Product paper · PDF · 8 pages

    Radmis AI Platform.

    The five questions, the worked example, the security design and coverage regulation by regulation.

    Download

    Technical paper · PDF · 8 pages

    What is in the platform, and how far it goes.

    For architects, reviewers and auditors. The architecture, the full feature list, the ISO mapping and coverage with every count.

    Download

    Next step

    Bring us the pilot that stopped at the risk review.

    Tell us what is stuck. In two weeks you have a controls document your reviewers can sign, and a go, fix or stop answer. Then you decide whether to go further.

    1. 1You tell us what is stuck, in a few lines.
    2. 2We run the four-minute build live and take your five vendor questions.
    3. 3The two-week Unblock starts. One of the three answers is always stop.