Skills SA runs four agents at VET compliance audits

Digital Attitudes, 2 July 2026

The Thursday read is one South Australian government agency running the annual VET compliance audit on a supervisor and three specialist agents built on Databricks, one Teach First survey that puts the AI strategy gap in England at 98 per cent of the school sector against 60 per cent weekly teacher use, one German randomised study whose year nine cohort rarely asked the AI tutor to check the arithmetic and finished the term behind where they started, one Anthropic and Microsoft product announcement that lands Claude Opus 4.8 and Claude Haiku 4.5 inside the Azure tenant Australian universities are already running, and one OpenAI product policy hire off the LEGO Group child rights desk.

Skills SA runs the annual VET compliance audit on a supervisor plus three agent Databricks stack

Ry Crozier at iTnews reported on Thursday 2 July that Skills SA, the South Australian government regulator that oversees the state's vocational education and training providers, has moved the Organisational Self Assessment process onto a multi agent AI workflow built on Databricks, with a supervisor agent orchestrating an operational agent, an analysis agent and a decision support agent against the six student support standards each provider is measured on. The regulator receives hundreds of submissions annually, each a 40 to 50 question self assessment form carrying structured data and heavy free text, and Skills SA AI, Data and Integration Lead Jarrad Taylor presented the design at the Databricks Data and AI Summit in San Francisco earlier this month. Taylor named the problem: "OSA is a compliance process used to assess training providers against a defined set of standards each year. We receive hundreds of these submissions, which are quite large and contain both structured and unstructured information and large volumes of free text within the responses." He named the failure mode the workflow targets: "the process is highly manual, requires significant assessor effort, and more importantly it relies on people reviewing large volumes of information and applying standards consistently across every submission." Taylor named the three layer benefit: "for our assessors, the goal is to reduce the time spent on routine activities and allow them to focus on the submissions that genuinely required further human judgment, investigation, and/or follow-up. For our operational staff, it provides some real-time visibility into workflow statuses and finding bottlenecks and trends across the assessment process. And for Skills SA, it helps apply standards more consistently while maintaining a complete audit trail of how recommendations and decisions were reached." Development ran for around 18 months and Skills SA now treats the pattern as reusable: "while we've used OSA as the example, we see it as a reusable pattern for operational AI in regulated environments. Anywhere you have a high volume of information, clearly defined standards, human decision-makers, and a requirement for governance and auditability, you'll often encounter the same challenges." iTnews on Skills SA treating the multi agent workflow as a reusable pattern, 2 July.

Why this matters: For the South Australian Department for Education and the Skills SA sister agencies inside the same portfolio reading the OSA rebuild against the standards based compliance work each of them carries into every school registration cycle, the operational read is that the state's own AI, Data and Integration team has just published a supervisor and three agent architecture that treats a large annual free text audit as a reusable pattern rather than a one off automation project, and the question the next portfolio briefing answers is whether the AISA independent schools registration audit, the SACE Board assessment moderation workflow and the annual government school compliance return get the same supervisor plus specialist agents treatment inside the current forward estimates or whether each program still writes its own pilot from scratch. For NSW Education Standards Authority, the Victorian Registration and Qualifications Authority, the Queensland College of Teachers registration desk and the ACARA data collection team reading the Skills SA pattern against the same standards based paperwork they process, the procurement read is that South Australia has just put a Databricks agentic reference architecture on the Australian public sector shortlist, and the question the next AI governance paper answers is whether the standards based audit workflow the interstate regulator inherits from twenty years of legacy tooling gets an agent architecture inside the 2027 budget cycle or whether the next parliamentary question about assessor backlog carries the same manual queue answer it did in 2024. For the ASX listed education software vendors selling compliance, moderation and assessment products into the Australian regulator market (Janison, TechnologyOne, ReadyTech, Litmos), the read is that a state education portfolio has now published a working supervisor plus specialist agents design against a real free text corpus, and the vendor whose 2027 product roadmap does not carry a governed multi agent story against the regulator standards workflow is the vendor whose next Skills SA, TAFE Queensland or NESA renewal paper reads thin against the incumbent architecture.

Teach First survey finds two per cent of English schools have a formed AI strategy against sixty per cent weekly teacher use

Emma Thompson at EdTech Innovation Hub carried the Teach First survey on Wednesday 1 July, which drew responses from nearly 200 secondary school leaders (approximately six per cent of the English secondary sector) supplemented by 30 semi structured interviews and six specialist interviews, and found the gap between AI in classrooms and AI in policy: 60 per cent of school leaders report using AI at least weekly (16 per cent daily and 44 per cent weekly), 14 per cent report no AI use, 12 per cent have an AI policy in place, 20 per cent provide any AI focused staff training and only two per cent describe their school as having a fully formed AI strategy. The barriers respondents named were limited staff confidence or skills (63 per cent), data privacy concerns (51 per cent) and limited understanding of educational AI applications (41 per cent). 79 per cent agreed AI could improve teaching and learning outcomes. London leaders reported daily AI use at 29 per cent against a 12 per cent national daily figure. Teach First CEO James Toop named the risk on the day the survey landed: "variable access to technology and training could deepen existing inequalities" and "teachers and school leaders are being left to navigate these changes without the guidance or support they need." EdTech Innovation Hub on the Teach First study finding AI use outpacing strategy in English schools, 1 July.

Why this matters: For state Department of Education chief digital officers and the AISNSW, ISV, AISWA and Catholic system AI leads reading the English 60 versus two split against the Australian Framework for Generative AI in Schools the education ministers endorsed last June and each system has been implementing across the last twelve months of term work, the operational read is that the neighbouring anglophone sector has just published the same gap Australian principals told the Grattan Institute survey they were living with a year earlier, and the question the next term 3 planning paper answers is whether the school AI policy at the front of the staff handbook is the fully formed strategy the two per cent line describes or the informal policy the twelve per cent line describes. For state education ministers and the federal Department of Education reading the daily AI use figure against the national framework rollout, the read is that a two year old policy framework does not by itself close the strategy gap when weekly teacher use is running at 60 per cent, and the question the next Education Ministers' Meeting answers is whether the framework picks up an implementation guarantee (an explicit strategy publication requirement, mandatory professional learning, and an audit trail schools can point to) before the 2027 sector self report reads the same 60 versus two split back into a Senate estimates hearing. For the Australian ed-tech vendors selling professional learning, policy templates and AI training into the K to 12 sector (Grok Academy, EdMuse, Impact Coaching, the AISA and AISNSW commercial arms), the read is that the survey has just named the demand signal (63 per cent of leaders citing staff skills as the barrier), and the vendor whose 2026 product roadmap does not have a term 3 principal ready implementation toolkit against the Australian framework is the vendor whose next renewal paper carries the wrong reference customer.

Tübingen randomised study finds year nine students rarely asked the AI tutor to check their work

Emma Thompson at EdTech Innovation Hub carried a University of Tübingen research paper on Wednesday 1 July from Rania Abdelghani, Peter Kaiser and Kou Murayama that ran 98 year nine students (aged 14 to 15) across three German schools in Baden-Württemberg through a mathematics AI tutor and examined the 1,616 chat exchanges the cohort produced. Before the sessions started, 82.9 per cent of the students selected step by step examples as a learning goal and 69.7 per cent said they wanted the AI to check their understanding. In practice, only 16.3 per cent of the requests they actually sent involved verification, 72.9 per cent of student messages were direct requests for help rather than monitoring or evaluation, and time spent monitoring their own understanding or evaluating the AI's responses averaged 5.7 per cent and 3.4 per cent respectively. Average pre test performance sat at 67.5 per cent, average post test performance sat at 56.9 per cent, and the decline was statistically significant. Abdelghani named the finding: "prompt quality alone is not enough to understand students' ability to use AI in advancing their learning." EdTech Innovation Hub on the Tübingen study of teenagers rarely checking AI answers, 1 July.

Why this matters: For year 8 and 9 mathematics faculty leads across the Victorian, NSW, Queensland, SA and WA government secondary systems who have been running AI enabled tutoring pilots against Mathletics, MathSpace, Education Perfect, Khanmigo and the in house teacher built tools sitting inside the Google Workspace for Education tenant, the operational read is that the closest randomised study to the cohort each of those pilots serves has just measured a 10.6 point drop between pre test and post test performance under an AI tutor and named the mechanism as the gap between what the students said they wanted and what they actually asked for. The question the next department AI in mathematics review answers is whether the AI tutor deployment inside the 2027 secondary mathematics program specifies a verification prompt scaffold the student must complete before the model answers the arithmetic question, or whether the current design still lets a year 9 cohort choose "just tell me" 72.9 per cent of the time. For the Victorian and NSW education departments' AI in classrooms guidance authors reading the study against the term 3 secondary teacher briefing, the read is that a peer reviewed randomised design has now put a number on the concern the classroom teacher union has been naming for two years, and the question the department AI in schools policy answers is whether the verification scaffold, the mandatory pre and post assessment loop, and the teacher inspection of the student prompt log gets called out as a term 3 expectation before the 2027 NAPLAN cycle records the same effect at the population level. For the ed-tech vendors selling AI tutor products into Australian secondary mathematics (MathSpace, Education Perfect, Cluey Learning, Matific, Mathletics, Stile) the read is that the peer reviewed evidence base has now caught up with the marketing claims, and the product whose classroom rollout does not carry a verification scaffold and a monitored prompt log is the product whose next department procurement paper reads with the Tübingen result on the same slide.

Anthropic ships Claude Opus 4.8 and Haiku 4.5 into Microsoft Foundry on the Azure tenant Australian universities already run

Emma Thompson at EdTech Innovation Hub reported on Wednesday 1 July that Anthropic has moved Claude Opus 4.8 and Claude Haiku 4.5 into general availability through Microsoft Foundry, with both models available on Azure from Monday 29 June, both supporting the Messages API, and both launching on Foundry with prompt caching and extended thinking. Anthropic and Microsoft named the enterprise controls the joint deployment carries: Microsoft identity integration, Azure networking controls, consolidated billing through Azure accounts, a US specific data zone as the first data residency option, access controls managed through existing Azure administrative tooling, and compatibility with Microsoft Enterprise Agreements for spending commitments. Two deployment routes exist inside the same tenant: Azure hosted with full Azure integration, and Anthropic hosted with the broader API surface Foundry has not yet lit up. Anthropic has not published a full pricing schedule or a complete supported processing locations list. EdTech Innovation Hub on Anthropic bringing Claude to Microsoft Foundry with Azure hosting and governance controls, 1 July.

Why this matters: For the AARNet, University of Melbourne, Monash, UNSW, USyd, ANU, UQ and Curtin CIO offices reading the Foundry announcement against the Azure tenant each institution has already stood the corporate identity, Microsoft 365 for Education, Azure networking and enterprise agreement rails on, the operational read is that the second frontier lab has now put a governed general availability path into the same tenant OpenAI has occupied since ChatGPT Edu landed on the Australian VC's desk, and the question the next enterprise architecture paper answers is whether the university's academic and research AI stack picks up Claude Opus 4.8 for high stakes marking, research assistance and student assessment work under the same identity, network and audit tooling the finance, HR and student records systems already run behind, or whether the sovereign data residency question (no Australian data zone announced) still keeps the deployment inside the Anthropic hosted API. For the state Departments of Education CIO offices reading the Foundry release against the enterprise Microsoft 365 for Education tenant each government school system runs, the read is that the same governance surface the AI enabled Copilot in Classroom pilot depends on now covers a second vendor's models, and the question the next enterprise AI procurement paper answers is whether the state's classroom AI strategy stays single vendor or opens a two vendor architecture before the 2027 curriculum cycle. For the Australian ed-tech vendors building AI features on top of the university and school Microsoft tenant (Stile, Atomi, Education Perfect, the local AI marking startups), the read is that the API surface each of those vendors integrates against has just doubled its inventory and the vendor whose product architecture picks up the multi model option inside the customer tenant reads better on the next enterprise renewal slide than the vendor still locked to a single API.

OpenAI hires John Buckley off the LEGO Group child rights desk to head child safety and youth wellbeing policy

Emma Thompson at EdTech Innovation Hub reported on Tuesday 30 June that OpenAI has hired John Buckley to lead child safety inside the Product Policy team and to support the company's youth and wellbeing initiatives. Buckley arrives with fifteen years of child protection and online safety experience across the LEGO Group (Director and Head of Child Rights and Safety), Google (child safety policy lead), YouTube (Director of child safety, YouTube Kids and livestream policy enforcement across Europe, the Middle East and Africa), Meta (child sexual abuse prevention and policy enforcement at Facebook and Instagram), Ireland's SpunOut.ie youth service, the Irish Society for the Prevention of Cruelty to Children, Childline, and a ministerial advisory committee to the Irish Government on youth work. His remit inside OpenAI covers designing AI products with child safety as a first order requirement, responding when harm occurs, and building preventative programming into the product surface. OpenAI's existing child safety architecture includes the Child Safety Blueprint, the Under 18 Principles inside the Model Spec, parental controls, and age appropriate protections on linked teen accounts. Buckley on the day of the announcement: "one of the most important places to work right now for children and vulnerable people to work is AI." EdTech Innovation Hub on OpenAI appointing John Buckley to a child safety and youth wellbeing role, 30 June.

Why this matters: For the state Department of Education AI governance leads and the AISNSW, ISV, AISWA and Catholic legal teams sitting behind the ChatGPT Edu tenant contracts and the ChatGPT under 18 usage the Australian K to 12 sector has been onboarding across the last twelve months, the operational read is that the OpenAI product policy stack that governs the linked teen account, the parental control surface and the Under 18 Principles inside the Model Spec now has a fifteen year child rights operator running it, and the question the next term 3 vendor conversation answers is whether the school system's own age assurance, parental notification and account provisioning workflow can lean on a named senior counterpart at the vendor when the next child safety incident hits the eSafety Commissioner's escalation queue. For the eSafety Commissioner's office and the OAIC reading the appointment against the Children's Online Privacy Code that becomes law on 10 December and the amended online safety legislation the government pushed through last week (the AUD $99 million penalty ceiling, the compel powers), the read is that OpenAI has just put a regulatory counterpart in place who can answer the Australian regulator's compliance questions in the same operating vocabulary, and the question the next joint eSafety and OAIC letter to the vendor asks answers itself. For the Australian ed-tech vendors building on the ChatGPT API stack (the local AI marking startups, the AI tutor products, the school administrative AI copilots), the read is that OpenAI has just signalled the child safety layer as a product first order concern rather than a policy afterthought, and the vendor whose product policy answer to the next school system procurement question does not carry the same architecture as the OpenAI parent stack reads with the wrong references on the first page.

Yesterday's brief covered the federal eSafety amendment and the ASD Essential Eight retirement; today's Skills SA and OpenAI hire stories sit against the same regulatory arc. More at digitalattitudes.com.au.