Two statements about one system can both be true and still look as though they cancel each other out. A large language model samples its output, so the same question put to it on Tuesday and on Wednesday comes back in different words. A control environment is a standing promise that the same question yields the same documented, checkable answer whenever somebody asks it. A board can maintain both beliefs at once, and neither statement gives way. No prompt template, temperature setting or vendor warranty removes the sampling behaviour.
The contradiction resolves at the level of the unit you choose to govern. The token stream is sampling noise and cannot carry a control. Artificial intelligence as a category is too broad to inspect. What can be governed is the deployment envelope behind a specific output: the model and its version identifier, the system instruction in force, the documents the retrieval layer was allowed to read, the person who accepted the risk for that use case, and the record that binds those four elements to a decision. Each section below explains one part of that envelope. No framework requires the model to act in a certain way, because a probabilistic system cannot be required for that.
The register entry that makes a deployment auditable
The envelope becomes governable the moment it is written into a register entry, and the entry fits on one screen. Failures in this area are failures of scope before they are failures of policy. An organisation that governs AI as a category has governed no record an auditor can inspect. An organisation that governs a specific deployment of a specific model for a specific decision has governed something that lasts after a review. Five fields carry the weight, and each field converts a claim into evidence.
deployment: invoice-exception-triage
model: vendor-model-name
model_version: 2026-05-14 # pinned to an explicit version
prompt_template: v7 # under version control, diffs reviewable
retrieval_scope: # document ids at the lowest level
- policy/returns-policy.pdf
- policy/supplier-terms-2026.pdf
authorising_owner: finance-controller # one person with a job title
risk_acceptance: signed 2026-06-02, reviewed quarterly
decision_record: append-only log, retained seven years
With that entry in place, the probabilistic behaviour stops being an unmanageable property. No person must state that the text repeats. The claim is narrower: the conditions were fixed, a specific person accepted them, and both facts can be shown on request. Tredence's guide to LLM governance makes the same point from the technical side, and it is the strongest case in that document, because tracing a decision back to the exact model version behind it turns an audit from guesswork into a five-minute lookup. The same guide places prompt injection, model drift, data exposure and shadow AI spending outside the remit of general access control, and aligns its component list with the NIST AI Risk Management Framework, its Generative AI Profile, and the management-system approach of ISO/IEC 42001.
Five domains in a published framework, and the stall it addresses
A published framework covers the same ground from the strategy end, and it identifies the obstacle organisations meet first. Arizona State University reports that Scott Emett, an associate professor of accountancy, and colleagues Marc Eulerich, Jason Pikoos and David A. Wood built a generative AI governance framework after watching firms stall, that the method went to a peer-reviewed accounting journal, and that the working version was released as a website drawing on more than 1,000 practitioners and academics, among them generative AI specialists, auditors, regulators and C-suite executives. A second version has already been launched.
The stall is not a knowledge gap about the technology. The framework's first instruction is therefore one of order: settle the organisation's risk tolerance, define what the technology is meant to achieve, and then choose which governance domains take priority. Priority is finite, and a framework applied everywhere at maximum depth is a framework no one finishes.
Two further details shape how the work gets budgeted. The team started from accounting on the reasoning that accountants think broadly about process, control, risk and opportunity, which places the framework inside the control function. Emett identifies the failure opposite to reckless adoption as underutilisation: the same report notes that his brother works for a large corporation where the IT department has decided against implementing generative AI at all, on risk grounds. Unauthorised use and blanket refusal are two faces of the same missing decision. The report adds that the models available today would take several years to apply to everything they can already do.
Steps that build the five artefacts
The framework is a production plan in which each domain specifies a thing that must exist. The five artefacts follow, and a deployment needs all five before anybody can describe it as governed.
- Strategic alignment and control environment. The artefact is a written statement of the business objectives the deployment serves and the objectives it leaves alone, with the policy set behind it. The ASU account of the framework records Emett's warning that overuse is the common error, and that putting generative AI everywhere brings problems a narrower deployment avoids.
- Data and compliance management. The artefact is the retrieval scope from the register entry, plus a classification decision on every document set inside it. A folder is the wrong unit of scope, because everything inside a folder that sits in scope is in scope, including the file somebody saved in 2019.
- Operational and technology management. The artefact is the test record. It states what was tested, against which model version, on what date, with what result, and what happens when the test fails. Cybersecurity, monitoring and IT change control all land in this domain.
- Human, ethical and social considerations. The artefact is the workforce decision: who is trained, who is redeployed, who is told what, and on what timetable. The ASU account records Emett's point that technically oriented leaders can under-weight social and reputational risk, and that a deployment which alienates the people expected to use it fails on its own terms.
- Transparency, accountability and continuous improvement. The artefact is the review cycle and the audit trail the register entry feeds. This domain is what keeps the other four from freezing at their launch state.
The framework is guidance, and the published account is explicit that its purpose is to help an organisation work out which risks matter most to it. The organisation makes that judgement, and no external document can make it on the organisation's behalf.
The person responsible for each artefact and the events that trigger escalation
Each artefact above needs a specific owner, and enterprise practice has already settled on four roles to supply them. Tredence's guide sets out which role answers each escalation trigger. An AI governance lead is responsible for policy and convenes the review council, with escalation when model behaviour falls outside the approved risk thresholds. An engineering lead manages the lifecycle, versioning and technical controls, with escalation on performance drift or a benchmark failure. Legal and compliance counsel confirms regulatory alignment and data handling, with escalation on a new regulation, an audit finding or a data incident. A business sponsor sets the limits of the use case and signs the deployment off, with escalation when the use case grows past what was approved.
Four roles, four escalation triggers, and a single accountable person at each step. The point of setting four roles is not to build a committee. It is to place each escalation trigger on a desk where somebody has the authority and the information to act on it. The same guide is explicit about timing: define these roles before deployment, because a role created during a live incident has no time to establish the records it needs. A council that meets quarterly cannot clear a trigger that fires on a Tuesday afternoon. The authorising owner field in the register entry therefore records one person with a job title, and the four roles here are where that person is found.
Where the people who sign the entries learn the job
The authorising owner field expects a job title, and the job title has a qualification route behind it, which is the part most AI governance plans leave unexamined. Governance is a taught and examined profession, and several British master's degrees now place the technical material inside a corporate governance curriculum.
Portsmouth's LLM Corporate Governance and Law is accredited by the Chartered Governance Institute UK and Ireland, and its graduates are eligible for the institute's Grad.CG qualification, with ACG status after relevant experience. Accreditation matters because the graduate answers to a professional body as well as to an employer when a signature is questioned.
Cumbria's LLM Corporate Governance, delivered through Robert Kennedy College, lists a compulsory module in transnational business law and artificial intelligence, which puts the model itself inside the governance syllabus. Bradford's two-year LLM prepares students for the assessments the Solicitors Regulation Authority sets and places them in a law clinic whose work counts towards the two-year Qualifying Work Experience the regulator requires; the same page gives the SQE1 fee as £1,934 and the SQE2 fee as £2,974 from September 2025, charged separately from tuition.
RGU's LLM Law with Corporate Governance lists the roles its graduates move into, among them corporate governance specialist, compliance officer, legal counsel, risk manager, company secretary and regulatory affairs advisor. Those titles map onto the register fields: the risk manager accepts the risk, the company secretary keeps the record, compliance checks the retrieval scope, and legal counsel checks the data classification.
Entry requirements differ enough to matter when a governance function recruits internationally. Portsmouth asks for IELTS 6.5 overall with no component below 6.0; RGU and Bradford both ask for 6.5 overall with no component below 5.5. The threshold is set by the admitting university, and it decides who is eligible to hold the title that signs the entry.
Four events that send an entry back for signature
A signed entry is not a permanent state. Four events reset it, and the register should show which one fired and on what date.
- The pinned model version moves. A vendor retires a snapshot or promotes a new one, and the deployment either migrates with a fresh test record or stops.
- The retrieval scope changes. One document joining the allowed set changes the answers the system can give, so the classification decision runs again on the new document.
- An escalation trigger fires: drift against a benchmark, a threshold breach, a data incident, or a use case that has grown past what the business sponsor approved.
- The rule covering the data in scope changes. The compliance owner checks the model against the new version of the rule and records the date of the check.
Each event writes a dated line into the register, and each one either renews the signature or suspends the deployment until the signature is renewed. Continuous improvement is the domain that keeps this clock running, and the ASU account notes that the research team has already released a second version of its framework, which is the behaviour it asks of the deployments the framework governs.
The alternative to a dated entry is not caution. It is a blanket refusal from the IT function, which the ASU report documents in the case of one large corporation, or a shadow deployment funded from a departmental budget that never reaches the year-end reconciliation, which Tredence's guide treats as a separate category of exposure. Both outcomes leave a board answering questions about a system it has not seen.
Sources
- LLM governance
- LLM Corporate Governance and Law GradCG
- LLM Corporate Governance (Robert Kennedy College)
- International Corporate Law and Governance with Solicitors Practice LLM
- LLM Law with Corporate Governance
Source: Generative AI isn't waiting for business to catch up, Arizona State University (ASU)





