Kitbag Consulting
Contact
← Thought leadership

Operating Logic as Compliance: DOLIUM and EN 18286

11 November 2026 · Perspectives · By Patrick Robertson

Operating Logic as Compliance: DOLIUM and EN 18286

Society is increasingly recognising the risks that artificial intelligence carries. As AI moves from experimentation into the decisions that shape people's work, finances and access to services, public trust has fallen and concern has risen, a shift the University of Melbourne and KPMG have tracked since their 2022 study, which predates the release of tools such as ChatGPT. That recognition is now visible in the institutions that act on society's behalf. During 2026, three of them, with no shared mandate, reached the same judgement that AI which cannot be shown to be governed carries a risk they are no longer prepared to absorb. They arrived at that judgement from different directions, one commercial, one legislative and one supervisory, and the alignment between them marks the point at which AI governance moves from aspiration to a condition of operating.

The commercial signal comes from insurers. Since January 2026, more than 60 US property and casualty insurance groups have filed to adopt standard exclusions that remove generative AI losses from general liability cover, and by April state regulators had approved more than 80 per cent of carrier filings to exclude AI-related damages. Insurers are withdrawing from AI whose construction, validation and controls they cannot see, while the specialist carriers that continue to write AI cover do so on the strength of documented governance evidence.

The legislative signal comes from Europe. On 14 October 2026 the European Commission published the reference of EN 18286 in the Official Journal, making this quality management system standard for providers of high-risk AI the first harmonised standard under the EU AI Act. Providers that build their quality management system on it receive a presumption of conformity with the quality management system requirements set out in Article 17 of the Act. The standard defines quality as compliance with the AI Act and with the law protecting fundamental rights, health and safety, and it requires that compliance to be organised, documented and evidenced across the lifecycle of each AI system.

The supervisory signal comes from Australia. APRA's letter of 30 April 2026 to all regulated entities found that AI governance frameworks are not operational in practice, that accountability across the AI lifecycle is unclear and that post-deployment monitoring is weak. It placed boards on notice that APRA will take stronger supervisory action, including enforcement, where AI risk is not adequately identified, managed and controlled. The Government has since gone further, announcing on 10 November 2026 that it will legislate mandatory reporting of AI incidents and establish a dedicated AI regulator.

Each of these institutions is responding to the same recognition of risk, and each turns on the same distinction between governance that exists on paper and governance that operates in the work. Insurers will not price what they cannot see, the EU now requires a quality management system that produces evidence, and APRA has found frameworks that are documented but not operational. This article examines that growing recognition of risk, sets out what EN 18286 requires, and explains why DOLIUM, as a System of Work that holds an organisation's operating logic and generates evidence through execution, aligns naturally with the framework the standard establishes.

A wider pattern

The three institutions above are part of a wider response. Legislators, procurement authorities and the public itself are expressing the same recognition of risk through different instruments, and the convergence of those instruments points to a structural shift in how AI will be permitted to operate.

The public's recognition of the risk behind these responses is well documented. A global study by the University of Melbourne and KPMG found that most people want AI regulated and doubt that current laws are adequate to keep its use safe. Australians are among the least trusting of AI of any country surveyed, and most say they would be more willing to trust AI systems where assurances such as adherence to international standards and responsible governance practices are in place.

The insurance retreat extends beyond standard exclusions. W.R. Berkley, for example, has proposed excluding any actual or alleged use of AI across its directors and officers, errors and omissions, and fiduciary liability lines. The retreat is driven by an evidence problem, as insurers generally cannot see how a policyholder built, validated or constrained the AI systems it depends on, and a risk that cannot be seen cannot be priced. Where affirmative AI cover is available, it is written on the strength of that evidence. Munich Re requires AI solutions to pass its technical due diligence before it will cover them, and Armilla integrates third-party assurance and ISO/IEC 42001 evidence directly into its underwriting. The market is withdrawing from ungoverned AI while continuing to insure AI whose governance can be demonstrated. CSIS has gone so far as to describe insurance as the most consequential de facto regulator of AI deployment in the United States, operating through thousands of individual underwriting decisions in the absence of federal rulemaking.

APRA's letter followed targeted engagement with industry in late 2025 and was issued to banks, insurers and superannuation trustees alike. APRA now expects boards to understand AI well enough to set strategic direction and provide effective challenge. The finding that governance frameworks are not operational in practice describes the gap between documented process and executed work that a structured quality management system is designed to close. ASIC reinforced the message on 8 May 2026 in an open letter calling on all financial services licensees to strengthen cyber resilience as frontier AI intensifies the risk environment. The letter cited the Federal Court's judgment in ASIC v FIIG Securities, which held that cyber risk management must be demonstrably effective and supported by clear governance.

The Commonwealth is now moving from supervision to legislation. In July 2026 the Prime Minister announced that the Australian Standards for AI would be legislated and established an Office of AI within the Department of the Prime Minister and Cabinet, and National Cabinet agreed the framework on 26 August 2026. The case for those standards sharpened after an OpenAI agent accessed a Services Australia Medicare statistics portal without authorisation on 18 June 2026, an incident the government was not told of until 10 September. On 10 November 2026 the Prime Minister announced that the government had accepted all recommendations of the resulting review, including mandatory reporting by AI companies of misaligned behaviour and cyber incidents, independent evaluation of frontier AI models and a new AI regulator to oversee security controls across the development and deployment of AI systems. OpenAI and Anthropic had both told a parliamentary committee that they would support mandatory incident reporting, which indicates that the developers recognise the same risk the public has identified. The government intends to introduce the legislation by the end of 2026 and pass it in early 2027, with mandatory incident reporting at its centre. A reporting obligation presumes that an organisation can see what its AI did, when and under whose authority, which is the evidence a governed quality management system is designed to produce.

Governments are codifying the same demand in statute and in procurement. South Korea's AI Basic Act entered into force on 22 January 2026. Its definition of high-impact AI closely mirrors the EU's high-risk category, and it requires risk identification, assessment and mitigation throughout the model lifecycle. The United Kingdom has chosen to build a market, and its Trusted Third-Party AI Assurance Roadmap of September 2025 projects an assurance sector that could reach £18.8 billion by 2035. In the United States, which has no federal AI statute, the requirement has arrived through purchasing power. A new General Services Administration contract clause took effect on 19 October 2026 and imposes binding data safeguarding obligations on large language model systems that process government data under federal contracts.

Public opinion, insurance underwriting, prudential supervision, legislation and procurement are separate instruments operating in separate jurisdictions. They converge on a single requirement, that an organisation deploying AI must be able to show how that AI is governed, by whom, under what controls and with what evidence. EN 18286 is the most detailed expression of that requirement to date, as it specifies the management system that produces the evidence.

The standard asks for operating logic

Article 17 of the EU AI Act requires every provider of a high-risk AI system to put a quality management system in place, and its central obligation is that the system be documented in a systematic and orderly manner as written policies, procedures and instructions. That obligation describes operating logic, which is the defined, governed and auditable account of who decides, on what evidence and under what controls. Holding operating logic is DOLIUM's core function. DOLIUM is a sovereign, IRAP-assessed and DTA-listed System of Work that sits above an organisation's Systems of Record and codifies how its work is defined, executed and governed. Most organisations will approach EN 18286 by writing a manual and preparing for assessment against it. An organisation using DOLIUM will express the same policies, procedures and instructions as the executable structure through which its work runs, so the evidence of conformance will accumulate as each task is performed.

Where the standard and the System of Work align

Article 17 sets out thirteen aspects that a provider's quality management system must cover at a minimum, and the most fundamental of them describe the way an organisation governs itself. A provider must hold a strategy for regulatory compliance that includes its procedures for conformity assessment and for managing modifications to the AI system. It must manage its communication with national authorities, notified bodies, other operators and customers. It must keep records of all relevant documentation and information, and it must maintain an accountability framework that sets out the responsibilities of management and staff. Each of these obligations is a statement of operating logic, and each will be held in DOLIUM as part of the structure through which the work runs.

In practice, the compliance strategy will sit in DOLIUM as governed operating logic linked to the obligations it serves, and every proposed modification to the AI system will pass through a defined change process with impact assessment and recorded approval. Communication with regulators and other parties will follow defined pathways assigned to accountable roles, with each exchange recorded against the obligation that triggered it. Role, responsibility and authority are first-order objects in DOLIUM, so each task will carry its owner, its approver and its link to the objective it supports, and the record of what was done, by whom, under what authority and against what standard will be generated as a by-product of execution.

EN 18286 reinforces these obligations through its leadership and support requirements, which call for clearly defined roles, responsibilities and authorities, structured communication processes and the visible involvement of top management. Kitbag's position that assurance should be built into the work maps directly onto these requirements. A quality management system that exists only as documentation invites the decoupling between written process and actual practice that the research on ISO 9000 certification has recorded over three decades, whereas a quality management system that operates as the structure of the work leaves no gap in which that decoupling can occur.

DOLIUM ensuring governance

Much of the remainder of Article 17 concerns the lifecycle of the AI system itself, including the techniques used to design the system and to control and verify that design, the techniques used in its development, quality control and quality assurance, the examination, test and validation procedures applied before, during and after development, the risk management system, post-market monitoring and the reporting of serious incidents. These activities rely on specialist methods, and EN 18286 leaves providers free to adopt the methods that suit their systems. What the standard requires is that those methods operate within a governed process, and DOLIUM will provide that process. For each activity, DOLIUM will define the sequence of work, the outcomes that must be achieved from each activity, the authorities who sign off and the evidence each step must produce, while the engineering content of each activity will follow the methods and standards the provider adopts.

Risk management illustrates the arrangement. EN 18286 requires a risk management system to operate across the lifecycle of the AI system, while the risk process itself will be the subject of a separate harmonised standard on AI risk management, prEN 18228, which remains in development. DOLIUM will host whichever method the provider adopts and will ensure that it runs on schedule, with escalation paths and an auditable trail. Serious incident reporting shows why that discipline matters. The AI Act requires providers to report serious incidents to national authorities within fixed deadlines, and a procedure that is triggered automatically from a recorded incident will meet them more reliably than one that depends on an individual recalling the obligation under pressure.

The standard also asks providers to manage their resources, including measures to secure supply, and here DOLIUM's architecture carries particular weight. DOLIUM is sovereign to the enterprise and runs on a containerised architecture using open-weight models, so a provider's quality management system will remain operable when a foreign vendor changes its licensing terms, withdraws a service or becomes subject to sanctions. Security of supply then extends to the system that governs compliance as well as to the AI system being governed.

Where DOLIUM governs and others perform

Two further aspects of Article 17 sit closest to the engineering of the AI system, covering the technical specifications and standards the system must meet and the management of the data used to build it, from acquisition and labelling through to storage, filtering and retention. DOLIUM will govern these activities by defining the procedures, owners and controls that apply to them and by recording that each step occurred, while the pipelines and engineering tools that perform the work will remain in the provider's data and development platforms. Stating that boundary plainly matters, as EN 18286 rewards a quality management system whose claims match its practice.

Relevance beyond Europe

The AI Act applies to providers outside the EU that place AI systems on the EU market, and the high-risk obligations will apply from 2 December 2027 for the systems listed in Annex III of the Act, which covers uses such as biometrics, employment, education and critical infrastructure. For Australian providers in that position, EN 18286 sets out the recognised route to a presumption of conformity. For Australian agencies outside its scope, the standard offers a benchmark for what a structured and evidenced quality system for AI should contain, at a time when government is beginning to deploy AI agents in operational volume. The governing question for those agents is the one quality management has always asked of human work, namely what was done, under whose authority and to what standard. EN 18286 now gives that question a regulatory form, and DOLIUM is built to answer it.

Conclusion

The insurers' retreat from AI liability and APRA's finding that governance frameworks are not operational in practice rest on the same proposition that governance which cannot be seen in the work cannot be relied upon. An organisation whose AI operates within a governed System of Work will generate the evidence that underwriters, supervisors and regulators now require, as a by-product of doing the work.

Sources

More thought leadership

View all insights →

Simplify complexity. Let’s start a conversation.

Get in touch →