DefinitionOperating model · operating layer · Operating SystemVersion 1.025 October 2026
AI operating model: who decides, funds and runs AI
An AI operating model is the organizational design that sets who decides, funds, builds, approves and runs AI in a company, with which roles, decision rights and cadence; the AI operating layer is the software and operating discipline that enforces those decisions on every request, and Hikari Blue’s Operating System is the method that installs both. The definition is Hikari Blue’s: no neutral reference text defines the term, so this entry names the frameworks it draws on and the test that shows a model is enforced.
On this page
An operating model is written once. An operating layer applies it to every request.
In this page, "we" means Hikari Blue, which engineers and operates the AI operating layer. The definition of the AI operating model is ours: no neutral reference text defines the term. The roles, decision rights and cadence named here are those of our Enterprise AI Governance Framework v1.0, published on 25 September 2026.
Buyers often merge 3 objects under one word: the organization that decides about AI, the software that enforces those decisions, and the method that installs both. This entry keeps them apart.
Operating model, operating layer, Operating System.
The operating model decides, the operating layer enforces, the Operating System installs. Each has its own artefact and its own holder.
| Term | What it is | What it produces | Who holds it |
|---|---|---|---|
| AI operating model | The organizational design that sets who decides, funds, builds, approves and runs AI, with which roles, decision rights and cadence. | Roles, a table of decision rights and a cadence of reviews. | The board, which sets the AI risk appetite, and one accountable executive. |
| AI operating layer | The software and operating discipline between AI models and a company’s systems. It routes each request to the right model, applies policy, records every decision and keeps a way to stop. | For each action: a policy decision, a record and a stop path. | The operator: during an engagement it may include us; after transfer it is the client team. |
| Hikari Blue’s Operating System | The method that installs both, in 6 phases: Diagnostic, Architecture, Build, Governance, Adoption, Run. | AI in production that your team can own, defend and stop, by a set end date. | A named partner, who signs the engagement. |
The definition of the operating layer in the second row is the one served on our definition page, word for word. The 6 phases are those of our Operating System page.
Roles, decision rights and cadence: 3 decisions of design.
Our governance framework answers each one with named objects a third party can check.
- 7 roles
Roles
Board, accountable executive, AI risk owner, system owner, human verifier, operator and model provider: roles, not job titles. One person may hold several in a small deployment, except that a verifier never approves an output of a workflow they own.
- 10 decisions
Decision rights
Each decision has one accountable role, in the RACI convention: from the AI risk appetite, set by the board, to the kill switch, which any named operator or verifier may trigger. 2 accountable names on one decision means none.
- 5 rhythms
Cadence
Per action, daily, weekly, quarterly and annual. The fastest rhythm needs no meeting: on each action, the layer evaluates the policy, enforces the corridor, checks the scope and writes the record.
A decision right that no system enforces is a policy memo.
An operating model written in a memo says who may approve a workflow, switch a model or stop an agent. It does not prevent anyone else from doing it. Enforcement starts where the layer evaluates the rule on each request, denies what the decision rights forbid and records who acted.
3 decisions of our framework show the split: the model names who answers, the layer makes the answer true at runtime.
| Decision | Operating model: who is accountable | Operating layer: what enforces it |
|---|---|---|
| Set or change a policy or residency corridor | The AI risk owner; the operator executes. | The policy version in force is evaluated on each request and written into the record of each action. |
| Select or switch a model provider | The system owner; the operator executes. | The switch is a routing change, recorded in the trail with its author and timestamp. |
| Trigger the kill switch | Any named operator or verifier. | The stop path revokes the agent’s session and tool permissions, and the trigger, its author and the halt time are recorded. |
An enforcement point that sits only at the API boundary leaves agents and their tools outside it: AI operating layer vs AI gateway
Ask which rule blocked the last refused action.
Pick the last action your AI systems refused, or should have refused. Ask 3 questions: which rule blocked it, in which version; which role owns that rule; where the refusal is recorded.
Pass
The answer comes from the record in minutes: the policy identifier and version, the accountable role named in the decision rights, and a timestamped entry that nobody assembled by hand.
Fail
Nobody can name the rule, the rule exists only in a document, or the refusal left no record. The operating model then exists on paper only.
The person who answers for one AI use, by name: Named owner
3 references frame the operating model, each from one side.
Each text below frames one part of an operating model: lines of accountability, leadership duties, governance as a function. The correspondence with our definition is ours. Mapping prepared by Hikari Blue; the IIA, ISO and NIST have not reviewed it.
- IIA · Three Lines Model
Lines of accountability
The Institute of Internal Auditors published its Statement of Position on the Three Lines Model on 8 July 2026, replacing the position paper of 2020. Management, the first line, owns and manages risks; second-line roles provide expertise, support, monitoring and challenge; internal audit, the third line, gives independent assurance and advice.
In our reading, our framework places the system owner in the first line and the AI risk owner in the second; internal audit tests the controls and never operates them.
- ISO/IEC 42001:2023 · clause 5
Leadership duties
The international standard for AI management systems, published on 18 December 2023, sets the duties of top management in its clause 5, Leadership.
In our reading, the roles and decision rights of an operating model are where an organization shows how it meets that clause.
- NIST AI 100-1 · GOVERN
Governance as a function
The NIST AI Risk Management Framework, published in January 2023, describes GOVERN as a cross-cutting function that is infused throughout AI risk management. Its category GOVERN 2 asks for accountability structures, so that the appropriate teams and individuals are empowered, responsible and trained.
In our reading, GOVERN describes what the operating model must achieve; the operating layer produces the evidence that it does.
Our governance framework v1.0 cites the 2020 version of the Three Lines Model. The 2026 statement keeps the 3 lines and their roles as summarized above.
Each of the 72 NIST subcategories, assigned to one of our 15 controls: NIST AI RMF implementation, control by control
The questions asked about the AI operating model.
What is an AI operating model?
The organizational design that sets who decides, funds, builds, approves and runs AI in a company, with which roles, decision rights and cadence. The definition is Hikari Blue’s; no neutral reference text defines the term.
What is the difference between an AI operating model and an AI operating layer?
The operating model sets who decides and runs AI. The operating layer enforces those decisions on each request: it evaluates the policy, writes the record and keeps the stop path. A model without a layer stays on paper; a layer without a model enforces nobody’s decisions.
Is Hikari Blue’s Operating System an operating model?
No. It is our method, in 6 phases: Diagnostic, Architecture, Build, Governance, Adoption, Run. It installs the operating model and the operating layer, and each engagement has a set end date, when your team takes over the system.
Who owns the AI operating model?
In our governance framework, the board sets the AI risk appetite and approves the use-case classes the enterprise accepts, and one accountable executive answers for AI in production. Each of the 10 decisions of the framework has one accountable role.
How do you test that an AI operating model is enforced?
Ask which rule blocked the last refused action. If the record names the policy and its version, the accountable role and a timestamp, without manual assembly, the model is enforced. If nobody can answer, it exists on paper only.
Reference documents and primary sources.
Every role, date and quotation on this page comes from the sources below, read on 25 October 2026.
Reference documents
- Enterprise AI Governance Framework v1.0 Roles, decision rights, 15 controls, evidence and cadence. CC BY 4.0.
- Operating System The 6 phases of the method, from Diagnostic to Run.
- What is an AI operating layer? The definition this entry builds on, and the 4 tests for buyers.
- AI strategy Where AI changes the P&L, workflow by workflow, before the operating model assigns it.
- Glossary of the AI operating layer The index of entries, with their status and citation.
Primary sources, consulted on 25 October 2026
- IIA · Statement of Position, Three Lines Model: Assurance and Advice in Support of Effective Governance (opens in a new tab) Published on 8 July 2026. Replaces the position paper of 2020. Executive summary.
- IIA · The IIA’s Three Lines Model: An update of the Three Lines of Defense (opens in a new tab) Position paper of September 2020, marked as superseded on the IIA site.
- ISO/IEC 42001:2023 · Information technology, Artificial intelligence, Management system (opens in a new tab) Edition 1.0, published on 18 December 2023. Clause 5, Leadership: title only, the text is sold by ISO and IEC.
- NIST AI 100-1 · Artificial Intelligence Risk Management Framework (AI RMF 1.0) (opens in a new tab) January 2023. Section 5.1, the GOVERN function, and Table 1, category GOVERN 2.
How to cite this document.
This URL is permanent and always serves the current version. Each change of substance increments the version and is listed in the changelog below, so a citation stays traceable to the text it quoted.
Preferred citation
Hikari Blue, "AI operating model: who decides, funds and runs AI", Glossary of the AI operating layer, v1.0, 25 October 2026, https://hikariblue.com/glossary/ai-operating-model
Section anchors are stable: #status, #three-terms, #decides, #enforcement, #test, #frameworks, #faq, #references, #cite.
Changelog
v1.0 · 25 October 2026. First public version.
License. Published under CC BY 4.0: free to cite, quote and reuse with attribution to Hikari Blue and a link to this page. The license covers Hikari Blue’s text, tables and diagrams on this page; it does not cover third-party data, quotations or legal texts, which keep their own terms, nor the Hikari Blue name and logo, trademarks of UNITED4 LLC.
Name the rule that blocked your last refused action.
30 minutes with a named partner. We take one AI workflow you run and check its operating model against what the system enforces: who holds each decision right, which rule the system applies, and what the record shows. You leave with the gaps, whether or not we work together.
- One workflow, its decision rights mapped
- What the system enforces and what is only written
- Who should own each gap