Most organisations deploying AI have tight controls around their finances and operations. Very few apply the same rigour to the AI systems now making decisions inside those same processes.
The eight standard internal control objectives, Completeness, Accuracy, Authorisation, Validity, Existence, Handling Errors, Segregation of Duties, and Presentation and Disclosure, were designed precisely to catch the kinds of failures that probability-based systems will inevitably produce. Applied to AI, they become a practical governance architecture rather than a theoretical framework.
Here is how each one translates.
1. Completeness: Account for what the model doesn’t know it doesn’t know
A complete AI system is one where the training data represents the full problem space, including rare but critical edge cases that only appear once in ten thousand transactions, and disproportionately when things go wrong.
- AI Application: Audit data pipelines for systematic omissions. Gaps in representation don’t announce themselves. They appear as unexplained performance degradation in specific cohorts or conditions.
- Control Metric: Every decision or transaction processed by the AI is logged and auditable. If you cannot reconstruct why the model produced a specific output, you do not have a complete system. Build automated testing of edge cases during development, with independent adversarial testing as a separate gate.
2. Accuracy: Precision starts with the data, not the model
Model performance is only as reliable as the data it consumed. Sophisticated architecture cannot compensate for poor-quality inputs. In generative AI, accuracy failures manifest as confident, fluent misinformation.
- AI Application: Implement data validation at ingestion, not retrospectively. Establish monitoring for model drift in predictive systems and hallucination rates in generative ones. A model that was accurate six months ago may not be accurate today.
- Control Metric: Regular benchmarking against ground-truth datasets with defined error tolerances. When the error rate exceeds the business threshold, that is a control failure, not a technical inconvenience.
3. Authorisation: Define what the algorithm is permitted to do
Authorisation in AI has two distinct dimensions: who can modify the system, and what the system is permitted to do autonomously. Most organisations manage the first. Fewer define the second clearly enough to enforce it.
- AI Application: Restrict access to model weights, training datasets, and deployment pipelines through Role-Based Access Control (a system where permissions are assigned by job function, not individual). Define explicit “human-in-the-loop” thresholds. These are the decisions that require human approval before execution, regardless of the model’s confidence score.
- Control Metric: Digital signatures for model versioning and deployment. Every change to a production model should be attributable to an authorised individual and logged with a timestamp.
4. Validity: Is the model solving the problem you think it’s solving?
A valid AI system produces outputs that align with the business intent behind it. This is harder to confirm than it sounds. Models are adept at finding patterns that predict the training outcome without understanding the underlying mechanism. This is called spurious correlation, where the model is technically right for entirely the wrong reasons.
- AI Application: Conduct regular “model validation” audits where independent experts review the model’s architecture and logic. The question is not just “does it produce correct outputs” but “is it using the right reasoning to get there”.
- Control Metric: Documented validation reviews at defined intervals, and whenever the business context the model operates in changes materially.
5. Existence: Know what AI is running in your organisation
Traditional existence controls verify that the assets on your balance sheet are real. The AI equivalent is maintaining a comprehensive, accurate inventory of every model operating within the organisation.
The risk here is what practitioners call “Shadow AI“, which refers to unauthorised or forgotten models running without oversight. These emerge when business units deploy AI tools outside of formal approval processes, or when models built for one purpose continue running after that purpose has changed or ended.
- AI Application: Maintain a central AI registry that records every model in production: its purpose, owner, training data lineage, version history, and approval status.
- Control Metric: Automatic reconciliation between the registry and the models running in production. Discrepancies are governance failures.
6. Handling Errors: Build for failure, not just performance
AI systems are probabilistic. They will make mistakes. The control question is not how to prevent all errors, because that is impossible, but how the system behaves when it errs.
The organisations that manage AI risk well design their failure modes before deployment, not after an incident has forced the question.
- AI Application: Develop incident response plans specific to AI failures. This includes automated fallback mechanisms, such as routing decisions to manual review when model confidence drops below a defined threshold, and clear escalation paths when automated systems produce anomalous outputs.
- Control Metric: All AI errors and near-misses are logged, reviewed, and fed back into retraining or process improvement. One failure is information. The same failure twice is a system problem.
7. Segregation of Duties: No single point of unchecked authority
Segregation of duties prevents fraud and error by ensuring no single person controls every stage of a consequential process. In AI, this principle applies to both people and the AI systems themselves.
The teams that develop models should not be the same teams that validate them. The people managing training data should not control deployment decisions. An AI agent that can initiate a financial transaction must not also be the agent that approves it. Approval authority requires independence, whether that approver is human or a separate, independently governed system.
- AI Application: Formalise the separation between development, testing, and deployment roles. Apply the same separation to AI agents operating in financial or high-stakes workflows.
- Control Metric: Mandatory independent review before any model moves from sandbox to production. No exceptions for speed or urgency. Independent audit of workflows to confirm separation is maintained in practice, not just on paper.
8. Presentation and Disclosure: Transparency is a control, not a courtesy
How AI-driven decisions are communicated to internal stakeholders, customers, and regulators is itself a governance requirement. Explainability is what allows oversight to function.
- AI Application: Implement Explainable AI practices so that any significant decision can be examined, challenged, and understood by a non-technical reviewer. Disclose clearly when users are interacting with an AI system rather than a person. Produce “Model Cards”, the standardised documentation that records the model’s intended use, known limitations, and potential biases.
- Control Metric: Adherence to applicable transparency standards, including the EU AI Act for organisations operating in or selling into European markets. Disclosure practices should be auditable, not aspirational.
Summary: Internal Control Objectives Applied to AI
| Objective | Traditional Definition | AI Governance Implementation |
|---|---|---|
| Completeness | All transactions recorded. | Data representation and decision logging. |
| Accuracy | Correct amounts and details. | Model precision and hallucination prevention. |
| Authorisation | Validated by management. | Access controls and human-in-the-loop thresholds. |
| Validity | Real, legitimate transactions. | Alignment with business intent and logic. |
| Existence | Assets actually exist. | AI inventory and Shadow AI prevention. |
| Handling Errors | Corrections are made. | Incident response and fallback mechanisms. |
| Segregation of Duties | Divided responsibilities. | Separation of development, testing, and audit roles. |
| Presentation | Clear and fair reporting. | Explainability and regulatory disclosure. |
The frameworks for managing AI risk already exist. They were built for exactly this kind of problem: systems that process large volumes of decisions at speed, where errors compound before they are noticed, and where accountability needs to be documented rather than assumed.
The organisations that govern AI well are not the ones building the most sophisticated models. They are the ones applying the most systematic controls.