What is the clearest sign that an organisation’s data estate is not ready for enterprise AI? 

The point where most businesses realise they aren’t ready for enterprise AI is when pilot projects stall with little to no measurable impact. Most of the time, AI isn’t falling short because of the tools, it’s because of the data, which becomes clearer when AI starts amplifying existing problems rather than fixing them, or demonstrating a valuable outcome. 

Weak governance and a lack of clear ownership also present themselves early on in AI experimentation cycles, with staff using LLMs in ways that leak sensitive data, usually without the knowledge of leadership, and when there’s no clear answer to who controls the data and who has access to it. 

Without the right data foundations and frameworks, the challenges already facing the business will only be magnified when AI is involved.  

Where do issues around inconsistent, poorly governed, or inaccessible data usually become visible first? 

For AI agents, issues arise when they need data access in order to complete a task, showing up when people and systems are over-permissioned and agents act autonomously beyond their instruction, and when they are under-permissioned or simply can’t access what they need due to fragmentation, poor data quality, or incompatibility. 

For generative AI and AI automation, poor data is immediately obvious when there’s poor accuracy and quality of AI outputs. While prompting and commands may form part of the issue, more often it’s a case of poor underlying data that’s restricting the AI system from providing high-quality results. 

In both cases, it relates to a direct outcome that stops the person or agent from completing a task. 

How can leaders tell whether an AI project is being held back by weak data foundations rather than the AI technology itself? 

One question is whether the same AI project would work with a different AI system, because if the same issues would arise across AI systems, then it’s clearly a data issue. That can span inconsistent and poor-quality outputs, agents requiring too much data access, or lots of manual inputs in order to function effectively.  

Evaluating data foundations and the overall digital health of the business is important before moving on to AI rollouts, otherwise aspects such as governance and unified data become bolt-ons and cause their own problems. A total of 97% of businesses with good digital health see ROI from AI data insights, according to Zoho’s Digital Health Study 2026, compared with just 45% of businesses with poor digital health. AI raises the ceiling of digitally adept businesses and lowers the floor of those who aren’t. 

What should CIOs, CDOs, and transformation leaders assess before approving the next phase of AI investment? 

Start with a clearly defined pilot programme with clear, measurable goals such as data insights, chatbots, and fraud detection. In areas such as these, it’s obvious when the AI is having a valuable impact and when it isn’t. Many businesses fall into the trap of trying to automate too much too quickly and lose grasp of the outcomes that showcase whether the AI system was successful or not. 

All of that should be underpinned by governance structures that define data access and audit trails that showcase how an AI system is interacting with the business functions. That way, it’s easier for CIOs, CDOs and transformation leaders to understand whether data foundations are strong, where AI systems are going wrong, and have confidence and trust in the outcomes, which becomes crucial for customers and end-users down the line. 

Getting AI right isn’t about picking the best tool. It’s about the groundwork underneath it. From Sachin’s insights, here’s what that actually looks like. 

• Audit your data before you blame your AI. 
  If a pilot stalls, the tech usually isn’t the problem. Poor data is. Check the foundations first. 
   

• Assign clear ownership. 
  Someone needs to own who controls the data and who can access it. Without that, you get staff leaking sensitive information into LLMs and nobody even knowing it’s happening. 
   

• Get agent permissions right. 
  Too much access and agents act beyond their brief. Too little and they can’t do the job at all. Both fail for the same reason: nobody set the boundaries properly. 
   

• Run the swap test. 
  Struggling with an AI project? Ask if a different AI system would hit the same wall. If it would, it’s not the tech. It’s your data. 
   

• Start small, measure everything. 
  Pick one pilot with a clear goal, fraud detection, chatbots, whatever fits. Then build governance and audit trails around it before you scale. Businesses with strong digital health see AI ROI more than double those without it. That gap isn’t about better AI. It’s about better groundwork. 

AI doesn’t fix a broken data estate. It just makes the cracks louder.