Enterprise AI has given data governance the sponsor it never had

The funding problem, in its historical form

For twenty years chief data officers have sought funding from executives who experienced governance as overhead. Support arrived reliably only during a regulatory or governance crisis, and receded once the crisis closed. This is the pattern behind Gartner's forecast that 80 per cent of data and analytics governance initiatives will fail by 2027, and behind the decay curve set out in part one: a programme funded by a crisis lasts about as long as the crisis does.

The mandate described in part one changes the obligation without changing that dynamic. A board obliged to evidence its critical operations will fund enough governance to survive the next supervisory review, which is a floor rather than a capability. Something else is required to fund the work continuously, and for the first time it exists.

The direction of dependence has reversed

Agentic systems cannot operate reliably on data that has not been classified, catalogued, quality-assured and access-controlled. An ungoverned estate produces output that is confident, unattributable and unauditable, which is precisely the combination an operational risk function cannot accept. AI programmes, meanwhile, attract funding at a scale never available to governance programmes, because executives treat them as central to strategy rather than as compliance overhead.

A chief data officer therefore no longer competes with the AI programme for budget. The executives sponsoring that programme cannot deliver it without the governance work, which makes them the first well-funded sponsors in two decades whose own objectives fail when governance fails.

The clearest evidence is an acquisition, not a product feature

Most vendors have automated parts of the governance task. Microsoft Purview scans for sensitive data across its cloud estate, Databricks generates documentation and lineage natively in Unity Catalog, and Collibra tracks models as governed assets alongside data. These capabilities reduce effort. None of them alters who pays.

The more significant development was a transaction. Salesforce signed a definitive agreement to acquire Informatica on 27 May 2025 for approximately US$8 billion, and completed the acquisition on 18 November 2025, taking on Informatica's metadata management, cataloguing, lineage and governance capability (CNBC 2025; Salesforce 2025). Marc Benioff put the reasoning in terms a data practitioner will recognise: “You have to get your data right to get your AI right. Data and context is the true fuel of Agentforce, and without clean, connected, trusted data there is no intelligence – only hallucination” (Salesforce 2025).

Our reading is that Salesforce priced governance as a precondition of its product rather than as an accessory to it, and paid accordingly. That inference is ours, drawn from the transaction and the company's public positioning, not a statement of Salesforce's strategy.

Gartner's cancellation forecast points the same way

Gartner forecasts that over 40 per cent of agentic AI projects will be cancelled by the end of 2027, attributing the cancellations to escalating costs, unclear business value and inadequate risk controls (Gartner 2025). The forecast rests on a January 2025 poll of 3,412 webinar attendees, so it measures practitioner sentiment rather than audited project outcomes, and should be read as an indicator rather than a measurement.

Read carefully, it still supports the argument. Inadequate risk controls is a governance failure under another name, and unclear business value is frequently what an organisation discovers when a model trained or grounded on an unreliable estate produces answers nobody will underwrite. An AI programme that treats data readiness as a caveat rather than a funded deliverable is describing its own cancellation in advance.

The same programmes create new objects to govern

Retrieval-augmented generation grounds a model's answers in an organisation's own approved policy and process documents rather than in generic training data, which is why it dominates enterprise deployments and why it improves the relevance of answers considerably. Its reliability is bounded by the corpus beneath it. A model retrieving from stale or conflicting policy documents returns authoritative-sounding wrong answers, at scale and on demand.

A confidently wrong answer about policy is a control failure, not a defect in user experience.

The model itself also becomes an asset requiring governance. Someone must version it and its prompts, evaluate answer quality as the corpus changes, control who may query what, and evidence all of it to an auditor. Under the mandate described in part one, that evidence is not an internal nicety. Where the model supports a critical operation, it forms part of the account an entity gives APRA of how that operation runs.

What to do now

Two actions convert the funding shift into governance capability, and a third stops the new assets from becoming an unmanaged liability.

The move for chief data officers is not to buy AI for governance, but to require that AI programmes pay for governance.

For two decades governance needed substantial funding and rarely received it. Every funded AI programme now depends on work the governance backlog already describes, which gives a chief data officer an argument that does not rely on a regulator or an incident. Organisations that use AI merely to run the old governance model more cheaply will join Gartner's 80 per cent, having automated a process nobody was using.