The conversation around AI readiness is changing as organisations move beyond experimentation towards using AI in live environments. While preparing enterprise data for AI remains fundamental, it only represents one side of the equation. Organisations also need to consider whether the AI itself is ready to operate within the controlled and accountable environment that enterprise use demands. Jaco van Niekerk, CTO of Intenda and the driving force behind the Fraxses platform, has been closely involved in addressing the challenges at the intersection of enterprise data and AI. Here, he shares his thoughts on what enterprise readiness for AI means, and what organisations should have in place before introducing AI into production environments.
Generative AI ranks as one of the most transformative technological innovations we’ve ever seen, but the characteristics that make it powerful also introduce uncertainty. While LLMs have been in existence for several years, their widespread use only took off with the release of ChatGPT in late 2022. Anyone who has engaged extensively with LLMs will have experienced how they can produce answers that seem entirely plausible but are factually incorrect, or respond differently to the same or similar inputs. For everyday users this unpredictability is manageable, but within enterprise systems where AI-generated outputs inform operational, financial or strategic decisions, the stakes are higher and the potential consequences more significant.
This unreliability becomes particularly important for organisations as they move beyond experimentation and begin integrating AI into production environments and business processes. A result generated during a controlled test is one thing. Allowing AI-generated outputs to influence live enterprise operations is another. Of course it would be nice if trust in enterprise AI could be achieved through better prompts, but it isn’t as simple as that, nor will it be anytime soon. At this point in the evolution of enterprise AI, trust needs to be established through the architecture and processes surrounding the AI.
An enterprise-ready approach to AI starts by recognising that an AI model should be treated as a component within a controlled system, rather than as an authority in its own right. Boundaries should be established around what the model can access, what it’s permitted to generate and what happens to its outputs before they are used.
The underlying enterprise data remains critical because metadata, schemas, relationships and business definitions provide the context within which AI operates. This information can be used to ground the AI in what is actually known about an organisation’s data, rather than allowing the model to infer structures or relationships independently. This reduces the space in which the model is able to make unsupported assumptions. Control also needs to extend beyond the information supplied to the AI. Validation, defined workflows and human oversight can ensure that generated outputs are checked before they progress to the next stage of a process or enter a production environment. In simple terms, AI can make suggestions, but humans and governed systems should make the decisions.
The advance of generative AI has led to much discussion about increasingly autonomous systems. Rapid as the evolution has been, we haven’t reached the singularity yet, and in the enterprise context, removing people from the equation is neither necessary nor desirable. A human-in-the-loop approach combines the speed and generative capability of AI with human oversight.
This approach doesn’t mean that every AI-generated action becomes a manual process. Approval gates can be introduced to processes at the key points where decisions have meaningful consequences. AI can perform time-consuming tasks such as analysing structures, generating documentation or proposing outputs, with humans remaining responsible for reviewing and approving critical decisions.
This creates a useful division of labour. AI handles the grunt work, maximising speed, efficiency and scale, while people provide judgement, context and accountability. For organisations operating in regulated industries or sectors where data accuracy and traceability are vital, this oversight also contributes to a clear audit trail. Organisations can understand what was generated, what information it was based on, what validation occurred and who approved the outcome.
Within a governed AI architecture, outputs can be validated against information that is known to be correct before they are accepted. In a data environment for example, an AI-generated SQL statement can be checked against the actual schema, relationships and data types discovered within the underlying source. If an output doesn’t align with what is known about the data, it can be rejected rather than passed downstream.
This represents an important shift in how organisations approach AI reliability: rather than assuming the AI model itself to be infallible, the surrounding system is designed around the possibility that it may be wrong. The objective is therefore not to change the probabilistic nature of generative AI, but to control how probabilistic AI interacts with systems that demand dependable outcomes.
Control is also closely linked to another key enterprise requirement: understanding how AI-generated outcomes were reached.
A governed approach creates traceability between the source data, the AI-generated output, the validation applied to it and the final approved result. Organisations have a documented path showing how an output progressed through the system – a far safer approach than accepting a black-box response with no explanation of how the AI arrived at the result.
This has implications beyond technical accuracy. Explainability and auditability are increasingly important considerations for governance, risk management and regulatory compliance. People are far more likely to trust AI-supported processes when they understand where information came from and know that checks have taken place.
Enterprise AI governance therefore isn’t only about restricting AI, but about giving organisations the confidence to incorporate it into their operations.
A controlled approach to enterprise AI starts with a governed data foundation that connects and contextualises information across enterprise sources. Data Intelligence capabilities establish metadata, relationships and business meaning, providing AI models with the context they need to work effectively with enterprise data.
Within this environment, AI can be treated as a non-deterministic component operating within a governed system. The process begins with deterministic discovery of the underlying data and metadata. AI can then assist with activities such as generating entity relationship diagrams, documentation and data objects – but it operates only within the structures that have actually been discovered.
Human approval gates and automated validation are introduced throughout the workflow. For example, generated SQL can be validated against known schemas, relationships and data types before it is approved and materialised. Crucially, AI does not write directly to production.
The result is an approach in which the creativity and speed of generative AI can be harnessed without requiring organisations to trust its outputs by default.
Up to this point, the conversation around enterprise AI readiness has concentrated primarily on data. While accessible, governed and meaningful data remains fundamental, the definition of AI readiness needs to expand as organisations progress towards production use. Not only does the data need to be ready for AI – the AI itself needs to be ready for the enterprise.
This means grounding AI in reliable data, establishing boundaries around what it is allowed to do, validating what it produces and retaining appropriate human oversight over consequential decisions.
Enterprises don’t need to change generative AI’s probabilistic nature in order to benefit from it. They simply need systems capable of managing this fundamental characteristic. By combining trusted enterprise data with governance, validation and human oversight, organisations can move beyond simply experimenting with AI and begin using it with the control, traceability and confidence required for enterprise environments.
Thank you for contacting us.
We will be in touch shortly.
