A superior technology such as AI cannot wash away the sins of poor implementation which inevitably results in poor data quality and as source data to AI. Will this poor quality of data not escalate problems with the embedding of AI in an ERP software? Analytics, the darling of consultants for over a decade, could well become an unmitigated disaster. It is firmly established that poor data quality alone remains one of the principal causes of failures in the efficient use of AI systems. Interestingly, the database as the OS, an idea articulated by Professor Michael Stonebraker, could be the opening to a new way of looking.
There have been sharp reactions in India and elsewhere to the US Federal Government order denying or limiting access to foreign companies to Anthropic’s two most advanced AI models – Fable 5 and Mythos 5, citing nation security concerns but without mentioning any specific threat. According to a news story in The Guardian, “It is Anthropic’s understanding that the government believes there is a method of bypassing, or “jailbreaking”, a safeguard that would prevent Fable 5 from being used in identifying software vulnerabilities. Anthropic has disagreed with the government’s view but has no choice but to abide by it.
Perhaps this is a blessing in disguise. Consider it as a recess time to clean up.
Poor quality data, not just out-dated infrastructure
Superior technology may exacerbate weaknesses and inadequacies in a system on which it is deployed. Ever since ‘Analytics’ became the dominant linguistic currency, enterprises, governments and other institutions have been talking incessantly about data, letting flourish an entire business around data; data cleaning became a business by itself. And ‘data-driven decisions’ became the new corporate and government slogan, with phrases like ‘evidence-based medicine’ being bandied about.
It is hardly surprising therefore to read that ‘AI data quality’ is the key to AI success, a message articulated by the tech media and tech companies, especially those with a substantial stake in the business of data. Data quality encompasses both obvious features such as accuracy, completeness, reliable with a goodness of fit and technical characteristics such as “as representativeness, bias, label accuracy and irrelevant variations (noise)—which can affect model behavior”, as pointed out by IBM, which has a huge business centered on data. “The IBM Institute for Business Value’s 2025 CEO Study found that only 16% of AI initiatives have successfully scaled across the enterprise, while MIT’s NANDA study3 reports that up to 95% of generative AI pilots fail to progress beyond experimentation”.
The clamour for replacing ‘out-dated infrastructure’ might well be justified though it does seem like an invitation for fresh investment, given the current valuations. However, that is in the future, even assuming everything goes well. The problem is here and now: poor data quality.
Surprisingly, for all the song and dance about ‘AI data quality’, there is no attempt to define what is the ‘source’ of this ‘AI data quality’, a highly dubious term. There is a more serious and fundamental problem than just data silos, the oft-mentioned term, which goes to the root of the problem: ERP implementation failures. It is ironic that while there is a separate and independent discussion on failures in ERP implementation, there has been no attempt to highlight it in the discussions on ‘AI-data quality’.
Continuing implementation failures
History teaches us that progress is a non-linear process, not an arrow which moves only in one direction. It is five decades since an operating system became an independently saleable product and more than three decades since the dawn and adoption of ERP, accompanied by a proliferation of discrete software packages addressing specific requirements, leading to the needs for system integration. For all the progress made by software systems, basic functions such as implementation continue to be problem areas.
A recent report (May 22, 2026) finds that more than 70% of ERP implementation initiatives fail to meet their initial objectives and almost 25% of them completely end up as failures. ERP is a foundational system, whatever the nature of an organisation, capturing the bulk of transactions. As a corollary, failure in implementation will have repercussions across the organisation deeply vitiating the work flow, with poor data quality an inevitable consequence leading to poor management decisions.
A LinkedIn post by Steve Novak in January 2026 says: “This research documents over $1 billion in combined financial impact across identified cases, including project costs, lost revenue, settlements, and remediation expenses. The pattern is remarkably consistent: compressed timelines, inadequate testing, poor data migration, and insufficient change management create predictable disasters regardless of whether organizations choose SAP, Oracle, Workday, or other platforms”. The post refers to law suits instituted by clients against the firm implementing the software, highlighting that “Zimmer Biomet’s SAP S/4HANA implementation stands as the most extensively documented ERP failure in the medical device industry, culminating in a $172 million lawsuit filed against Deloitte in September 2024”.
As the study (referred to above) explains, “One of the primary reasons behind this trend is the increasing complexity of ERP implementations. Modern ERP systems are no longer standalone platforms. They are deeply interconnected ecosystems that involve cloud environments, third-party integrations, regulatory compliance requirements, and real-time data processing”.
In a September 2025 study titled “Top 10 ERP Failures of the Last Three Decades (yep, they’re still relevant)”, Panorama Consulting sums up thus: “While the case studies below span three decades, their strategic missteps are just as relevant in today’s ERP landscape, especially as organizations integrate cloud architectures and AI in ERP”. (https://www.panorama-consulting.com/top-10-erp-failures/)
This raises an obvious question about the prevailing skill levels in the IT and tech industry.
Summing up
However bitter a pill this might be to swallow, it is not easy to move from implementing packaged software and filling in gaps through a Gap Analysis to dealing with agentic AI models. It is an entirely a different world. The ‘AI industry’ is attempting to build a huge edifice driven by the desire to cash in on valuations, without considering foundational dimensions. A sophisticated AI infrastructure cannot magically overcome the obstacles thrown up by continuing failures in implementation of software systems like ERP.
If data the is fulcrum, perhaps it is time to think of the database as the operating system, an idea propounded by Professor Michael Stonebraker.
