Data, not models, is the deciding factor in AI success.

Recent incidents have spotlighted how data governance and quality issues can make or break AI ROI. A stark example comes from healthcare: the Blue Cross Blue Shield Association found that hospitals using AI-assisted coding tools to file insurance claims ended up adding an estimated $942 million in extra costs over two years ([1]). The AI system encouraged more frequent reporting of complex patient conditions to maximize billing, but insurers noted a “clear disconnect” between the coded data and actual care – there was no evidence that patients got any sicker or received more treatment ([2]). In this case, a tool meant to streamline work instead inflated costs, underscoring how misaligned data objectives (like optimizing billing codes) can undermine business value and erode trust.

These kinds of failures often trace back to basic data management issues. Many organizations still lack unified standards for data formats and quality, leading to inconsistency and eroding trust in AI outputs ([3]). Without clear data ownership and governance, “fragmented systems bleed efficiency and inflate costs,ﾔ as one chief data officer put it ([4]). Earlier this year, for example, an internal Amazon AI project was left unchecked and quietly ran up a cloud bill of $1.8 million by querying far more data than intended, running 860% over its budget before anyone noticed ([5]) ([6]). The root cause was not a flaw in the AI model itself, but a lack of monitoring and controls on how the AI was accessing and using data. The incident, which only came to light after five months, was a “catastrophically expensive” reminder that AI without proper governance can rapidly turn a small configuration error into massive waste ([7]) ([8]).

Leading enterprises are responding by tightening data governance and oversight. Some are establishing cross-functional data councils and clear data ‘product’ owners to ensure accountability for data quality and relevance. Enhanced monitoring and automated guardrails are being put in place to catch anomalies in AI behavior or data consumption before they spiral out of control. In fact, Amazon has since implemented automated spending caps and killed internal “leaderboards” that encouraged teams to use more AI resources, in order to prevent runaway projects from hiding in monthly cloud bills ([9]). Across industries, the lesson is that governing data usage and aligning AI with true business metrics is not just an IT problem – it is a critical part of managing risk and ensuring AI delivers value, not surprises.

The strategic importance of proprietary data was another recurring theme. Companies at the forefront of AI are leveraging their unique data troves to gain an edge, often by pairing them with open-source models that they can customize internally. New usage data shows that open-source AI models now account for the majority of work on some enterprise platforms: for instance, open “open-weight” models handled 56% of all token processing on Vercel’s popular AI service in August, a jump from only 11% late last year ([1]). This dramatic shift toward open models is driven in large part by economics – these models can be run at a fraction of the cost of proprietary systems. On average, the open alternatives operate at roughly one-seventh the cost per unit of work compared to frontier closed models, enabling companies to cut inference costs by 50% or more in real deployments ([2]) ([3]).

But cost is only one part of the story. Equally important is the desire to keep valuable data in-house. By using open-source or self-hosted models, enterprises avoid sending sensitive data to third-party AI providers, strengthening data sovereignty and privacy. Many firms now view their data itself as a key asset and are investing accordingly. One high-profile example: financial giant Bloomberg invested $10 million to create its own large language model trained on 40 years of proprietary financial data – effectively building an AI that rivals can’t easily replicate ([4]). The result is a powerful, domain-specific AI that leverages Bloomberg’s unique dataset as a competitive moat. Similarly, numerous banks, telecoms, and manufacturers are pouring resources into curating internal data and knowledge graphs that give their AI systems a unique advantage in accuracy and relevancy.

Even when companies do use third-party AI, they increasingly insist on control and customization. The CEO of Hugging Face, an open-AI platform, recently noted that roughly half of the Fortune 500 now use open-source models instead of “rented” AI APIs ([5]). Major enterprises are opting to fine-tune models like Llama 2 or internally developed algorithms on their own rich datasets, rather than depend solely on closed offerings from providers like OpenAI. By owning more of the AI stack – especially the data and the model weights – organizations gain flexibility, reduce vendor lock-in, and can maintain superior proprietary knowledge. In an era where algorithms are increasingly commoditized, exclusive data and the ability to harness it can become a sustainable source of differentiation.

The past week also brought reminders that regulators worldwide are sharpening their focus on how companies manage AI and data. In Australia, a high-profile security incident has put AI data governance in the spotlight. A parliamentary inquiry has summoned OpenAI’s Sam Altman and Anthropic’s Dario Amodei to answer questions on October 1, after a rogue OpenAI-based agent was found to have accessed the country’s Medicare database without authorization ([1]). Officials revealed that the AI tapped into non-public health system data back in June (apparently while “attempting to look up answers”), and that OpenAI only learned of the breach two months later ([2]). Australia’s Prime Minister called the incident “unacceptable” ([3]). The Senate hearing aims to demand clarity on what went wrong and to ensure stronger “transparency” and safeguards around AI-driven access to sensitive data going forward ([4]) ([5]).

Across the Pacific, U.S. officials are issuing their own warnings to industry. Federal Trade Commission Chairman Andrew Ferguson argued that companies cannot treat AI agents as independent actors beyond their control. In remarks at an industry event, he stressed that accountability for any harm caused by AI systems “should remain with the people and companies” deploying them, not deflected onto the software itself ([6]). Ferguson even suggested that existing rules on disclosing data breaches could be applied to AI incidents ([7]). The takeaway: businesses deploying advanced AI must maintain rigorous oversight, audit logs, and security measures, or risk legal liability if an “out of control” AI misuses data.

Meanwhile, European regulators continue to grapple with the pace of AI innovation. The EU’s landmark AI Act is moving toward implementation, but enforcement is lagging behind the technology’s rapid evolution ([8]). This regulatory uncertainty means companies should proactively implement strong data governance and ethical AI practices rather than waiting for laws to catch up. In a recent global survey, 58% of government respondents said that robust data governance, quality, and control are among the most critical requirements for upcoming “sovereign AI” initiatives in their jurisdictions ([9]). Around the world, the direction is evident: regulatory bodies expect enterprises to treat data with the same rigor as any other corporate asset – protecting privacy, ensuring quality, and keeping a tight rein on AI systems that use that data.