For the past few years, the discussion has centred on generative AI, consumer use of AI, and model sophistication. The industry also focused on building smarter models, stronger reasoning, and better performance. As enterprises have started experimenting with these models, they face a stark reality: data quality, data governance, process optimisation, process orchestration, and a single source of truth for data and processes remain the core challenge. In short, AI is mainly a data problem.
Organisations grow both organically and inorganically, and as they grow, their processes evolve, compliance becomes more complex, and data becomes fragmented. At times, technology and adoption lag behind the pace of the organisation’s growth. In other words, the problem is more about governance and data than intelligence.
Suresh Anantpurkar, Founder and CEO, Manch Technologies
Importance of data
It is important to understand the key roles of data and governance in the success of AI deployment. Organisations may have structured, unstructured, and multimodal data. A lot of intelligence may be hidden in the multimodal and unstructured data. If the organisation wants to find the original payment terms, contract terms, and commercials, and how they have evolved over time, it may need to scan a large volume of documents, extract the relevant information, build a verification process, and then pass it to subsequent systems for further action.
Often, due to inorganic growth, customer and supplier information may be scattered across multiple systems. Even before embarking on any AI journey, it is imperative to consolidate and clean the data to create a single source of truth, often called the Golden record. An AI model run on fragmented and incomplete data will yield suboptimal or, at times, disastrous outcomes.
Metadata management, data lineage, and data provenance are critical for successful AI governance. Well-defined metadata enables integration, automation, and business logic. It also helps define the context and standards for the organisation to use the data consistently across business functions and systems.
Data lineage and data provenance play a critical role in the origin, changes, data flow, and data transformation throughout the data life cycle. Both are critical, as lineage helps debug data issues, and provenance helps ensure compliance. The agentic AI framework has evolved rapidly; now agents can consume data, perform tasks, invoke APIs, interact with the external world, and make decisions, all with little human involvement. Intelligence is moving into the operational fabric of the organisation very quickly. These workflows don’t necessarily need a frontier model; the focus is more on trusted outcomes than speed and intelligence. With models emerging so quickly, it’s fair to say that intelligence is abundant, but trust has yet to be established.
Why governance matters
Inefficient governance of agents can result in major loss or embarrassment to the organisation. Imagine an agent extracting outdated MSME certificate information and making a decision on vendor payment. Though every step was executed properly, every workflow was correct, and every communication was carried out as required, the business outcome was flawed. This is not an AI failure; it is a data and governance failure.
Governance, or agent governance as it is now called, depends on two critical aspects: data governance and control governance. Data governance means the agent operates on trusted, accurate, complete, verified, and contextual information, and control governance means the agent has the right permissions, the behaviour can be observed and explained, it is fully auditable, and it complies with privacy, security, and accessibility parameters set by the organisation.
The organisations that create value may not necessarily use the most advanced or powerful models. They will operate autonomous systems with confidence, control, and accountability, delivering well-defined outcomes.
The agentic AI framework shifts the task from human to machine, but it is important to keep humans in the loop, build guardrails, and provide tools and methods that help humans understand why a particular decision is made. Without appropriate tools and methods and due to the black-box decision-making approach, the agentic AI framework will face headwinds in scaling.
Use of AI requires a systematic governance model to scale. Guidelines such as transparency, accountability, and fairness are mandated by regulators and governments; their implementation requires strategic intent and an adaptive governance model as AI systems continue to scale and new use cases emerge.
Risk stages
- Development Risk – During the development phase, use skewed or incomplete training data sets, incomplete validations and rule identifications, and identification of the role of human, i.e., appropriate human-in-the-loop implementation.
- Deployment Risk – Deployment risk primarily arises from changes in data over time and from the inability to detect and alert appropriately. The human-in-the-loop plays a critical role during deployment because AI output needs to be explainable to stakeholders or end users, and without an appropriate reasoning or evaluation (eval) model, end users’ trust may erode.
Key pillars of governance
AI actions must be aware of the identity (i.e., user identity) and the classification of the data, keeping sensitivity, need-to-know, and privacy norms in mind – i.e., who can use it, how it can be accessed and used, for what purpose, and for how long.
Audit log and data lineage – AI output, inference, evaluation, and input must be available for audit checks; data should be traceable, and all changes must be properly tracked. Existing governance and regulatory requirements should be enforced, logged, and, if needed, presented for verification and audit checks.
Summary
In summary, the agentic framework in a B2B setup largely depends on data quality, data lineage, data provenance, the unbiasedness of the model that uses this data, the unbiasedness of the training data, and the organisation’s governance and regulatory setup. The article has tried to articulate these points in the above sections.
About The Author
