As organisations deploy autonomous AI agents across the enterprise, Peter Manta, AI Strategy and Practice Director, Informatica from Salesforce, says governance must evolve beyond traditional data controls to ensure systems interpret, infer and act on information safely, responsibly and with appropriate human oversight.
Deploying AI agents is a fool’s errand without the right governance policies in place. In an agentic environment, organisations must now think not only about who can see data but how autonomous systems will interpret and act upon it.
Responding to this new reality will not involve tearing up the rulebook and starting again. The foundational work of controlling who can access the data, understanding where it came from and building confidence in how that data will be used remains relevant.
Instead, agentic and traditional control layers will need to work together. Governance must also ensure that, when AI systems autonomously interpret data, draw conclusions and act on them, organisations are protected against unintended or harmful outcomes. To achieve this, decision makers will need to ask questions that existing data policies were never designed to answer.
Agents of change
To understand what needs to change, consider the traditional concerns of data governance, which focused on determining accuracy, deciding who could access data, defining how it was stored and establishing how it should be used within a business setting. That’s still true, but it’s no longer the full picture when agents can make decisions and take action.
Take a lender reviewing mortgage applications. AI is likely to be extremely effective at identifying risk factors buried within hundreds of pages of documentation, highlighting issues that a human reviewer might otherwise miss. But should it make the final lending decision? Few lenders are likely to say yes because the cost of making a bad decision can be so high.
When agents are used to process information and support decision-making, the governance question becomes much deeper, requiring organisations to think carefully about how autonomous systems behave and where human oversight should remain. It might be as simple as deciding that some processes are just too sensitive to be automated. Other cases will require a more nuanced approach.
The agentic data contract
This is where the Agentic Data Contract comes in. It is a framework for governing how agents access, interpret and act on data. Working through it means starting from the business goal, identifying the data needed to serve it and then examining that data against the contract’s dimensions. That process surfaces where the threshold for human involvement should sit. In lower-risk scenarios, agents may operate with minimal oversight. In higher-risk environments, humans should remain responsible for reviewing recommendations and making final decisions.
Let’s say a chatbot has been put to work providing internal support to employees operating in a well-secured corporate environment. In this case, continuous and rigorous human oversight is likely to be unnecessary. If the system is trained exclusively on trusted organisational processes and instructed only to deliver text-based outputs, the boundary does not require heavy policing because the bot is less likely to produce harmful outputs.
Now imagine an agent given responsibility for handling investments. Many people would be concerned to hear that a machine has independently decided where their retirement savings should be invested because of the responsibility inherent in that high-stakes decision. In this case, a human should still make the ultimate decision.
Deciding where the human sits is one lever. But it is not the only one and sometimes it is not even the right one. In a recent experiment, I found that when duplicate data exists, there is no reliable way for an automated system to separate good agent decisions from bad ones. Reviewing every decision by hand defeats the purpose of autonomy. The practical solution is to govern upstream, de-duplicate and prepare the data before an agent ever acts on it.
It’s clear that governance is becoming not only about managing data itself but also about defining where autonomous systems are permitted to act independently and where humans must remain firmly in control.
Unexpected inferences
The human-machine boundary is a relatively visible governance challenge. But a harder-to-see problem lies beneath the surface of any agentic deployment. Autonomous systems can draw inferences from data that are not necessarily incorrect but can still be harmful.
Consider an agent with access to customer interactions, transaction records, support tickets, location data and behavioural information. Each dataset may be harmless in isolation. Yet when combined, they could allow the system to infer sensitive information that was never explicitly provided and that the organisation never intended to expose. If that inference is shared or acted upon, the organisation has a problem on its hands.
Addressing this challenge starts with ensuring the data made available to agents is properly governed from the outset. Rather than assuming problems can be fixed later, organisations need to secure data at the point of origin and maintain strong controls over the information an agent’s ‘brain’ can access at runtime. That means understanding not only what the data contains but what it means in the real world and how it relates to other organisational information.
Context is also critical. Data only becomes meaningful when interpreted within the correct situational frame so that systems understand not just what information says but what it means and how it relates to the wider business environment.
Without that foundation, AI systems can make technically logical but fundamentally mistaken decisions. Governance must provide this context before data is ever made available to an agent, ensuring it is classified, understood and handled responsibly.
Creating guardrails for autonomy
Governance, therefore, needs to be treated as a strategic lifecycle discipline rather than a one-and-done exercise. The decisions made during foundational model training can continue shaping an agent’s behaviour long after it enters production. To ensure safety throughout an agent’s operational life, models must operate under robust data governance policies from the moment they begin training while agents must access and act on enterprise data under appropriate governance controls during deployment.
Comparisons to training a human are not necessarily relevant here. Once an agentic system has learned something from the data it works with, that learning becomes extremely difficult to reverse. This is not a matter of opening the hood, tweaking some code and removing the knowledge. The nuance a system absorbs from runtime data cannot be cleanly extracted, and undoing it often means discarding everything else the system has learned and falling back to the underlying foundation model.
That is a costly and imperfect remedy and it is precisely why the data a system can access and learn from must be governed before it ever acts. The organisations that succeed with agentic AI will be those that governed it long before it made its first decision.
