Technical Note
Mapping the Five Cs of AI Governance onto the financing and institutional themes of digital public infrastructure safeguards.
Digital public infrastructure safeguards fail at a predictable point. Not at the drafting stage, where principles are articulated with care, and not at the political stage, where they are endorsed. They fail at the budget hearing, where a finance ministry official asks what a safeguard costs, what happens if it is not funded, and how anyone would know whether it worked. Frameworks that cannot answer those three questions are cut, and they are cut early, because a line item with no defined failure mode is indistinguishable from an aspiration.
This is not a financing problem in the first instance. It is a risk and controls problem. Enterprise risk management has spent three decades building the machinery that converts a stated principle into a funded, owned and testable control, largely in service of regulated banks, insurers and infrastructure operators. That machinery has not been systematically applied to digital public infrastructure, and its absence explains a great deal about why safeguards remain chronically underfunded relative to the platforms they are meant to constrain.
A safeguard becomes fundable when three conditions hold simultaneously. Each is ordinary practice in enterprise risk management and each is routinely missing from DPI safeguards documentation.
The difference between a safeguards framework a ministry admires and one an auditor can test is whether every recommended safeguard carries four attributes: an owner, a control type, a tolerance threshold, and an evidence requirement.
The Five Cs of AI Governance were developed as a jurisdiction-neutral core, on the principle that regulatory regimes change faster than institutions can rebuild their governance architecture. Individual regimes — the EU AI Act, Korea's AI Basic Act, India's IT Rules, sector guidance from central banks — layer onto the core as packs rather than being run as separate programmes. The same property makes the framework portable across DPI deployments in jurisdictions at very different stages of regulatory maturity.
Every system named, owned, risk-tiered and purpose-bound. An inventory is the precondition for everything else; a safeguards policy that cannot enumerate the systems it governs is a document, not a control.
Safeguards that bind before deployment rather than after: accuracy and bias testing against defined thresholds, human oversight at specified decision points, documented limits on permissible use.
Continuous, demonstrable alignment to legal and standards obligations, held in a live register rather than reconstructed annually for an audit.
Evidence that external parties can rely on: disclosure to affected persons, provenance and labelling of machine-generated content, explanations proportionate to the decision's consequence, and honest statements of system limitation.
Lifecycle safety: drift monitoring, incident response, vendor safeguards, and planned decommissioning. For a horizontal layer serving population-scale services, continuity is not a technology concern but a public service availability concern.
The 2026–27 DPI safeguards agenda is organised around financing and institutional questions. Each maps onto a specific piece of risk machinery.
| Safeguards theme | What the risk discipline supplies |
|---|---|
| Cost items and trade-offs | Continuity of a horizontal layer is a business impact analysis question under ISO 22301: which public services stop, for how long, and what recovery capacity must therefore be funded. A recovery time objective stated against a named service is a costable number in a way that "resilience" is not. |
| Institutional coordination | A matrix of institutional responsibility is a Three Lines of Defence problem. First-line delivery ministries own the risk; a second-line safeguards function sets standards, tolerance thresholds and escalation; third-line independent assurance reports outside the delivery chain. Where the third line reports to the first, assurance is structurally compromised regardless of the competence of those performing it. |
| Public–private governance | Vendor lock-in, data sovereignty and exit capability are contractual risk positions. They can be registered, tiered, assigned owners, and carried into standard procurement clauses with defined triggers — rather than being negotiated afresh, and inconsistently, on each contract. |
| Monitoring, evaluation and learning | Standardised indicators require an obligations register mapping each safeguard duty to an owner, a control and an evidence artefact. This is what allows MEL findings to enter annual public financial management review cycles as budget evidence, rather than terminating in a report. |
Of the components above, the obligations register does the most work and is the most often absent. It is a simple structure: every legal, regulatory and contractual duty the deployment is subject to, expressed as a discrete obligation, each mapped to a responsible owner, the control that discharges it, and the artefact that evidences the control ran.
Its value is that it converts between two languages that otherwise do not meet. On one side, a safeguards principle — meaningful human oversight, say. On the other, a budget line — two full-time reviewers, a case management system, a quarterly sampling audit, a defined escalation path. The register is the translation layer, and without it the conversation between a safeguards community and a finance ministry remains one in which each side is convinced the other is being unserious.
It also survives personnel change, which matters more in public administration than in most enterprise settings. A register with named owners and dated evidence retains institutional memory across transfers and electoral cycles in a way that tacit knowledge held by a committed official does not.
Risk management supplies structure, not legitimacy. It can tell a government that an exclusion risk is scored high, owned by a named agency, and treated by three typed controls. It cannot tell that government what level of exclusion is acceptable. That is a political and rights question, settled through participatory processes and constitutional constraint, and it must be settled before a tolerance threshold can be set.
The discipline is therefore downstream of the rights conversation, not a substitute for it. Its contribution is to ensure that once a society has decided what it will not tolerate, the machinery exists to fund the safeguard, assign it, test it, and prove it ran.