Dominic Jainy stands at the forefront of the next great shift in enterprise technology, bringing years of deep-seated expertise in artificial intelligence and machine learning to the complex world of ERP systems. As organizations move beyond simple automation toward truly autonomous agents, Dominic has become a leading voice on the critical need for governance frameworks that keep pace with innovation. His work focuses on the delicate balance between the efficiency of delegated work and the absolute necessity of human accountability in platforms like Dynamics 365.
In this conversation, we explore the fundamental transformation of ERP security models as they transition from human-centric logins to event-triggered autonomous actions. Dominic breaks down the essential components of a modern governance strategy, including why audit trails must now capture “reasoning paths” rather than just final outcomes. We also discuss the practical implementation of approval thresholds, the danger of behavioral drift in AI agents, and how businesses can retrofit oversight into existing systems without stifling the speed that makes agentic AI so attractive.
Traditional ERP security relies on human logins for accountability, but agentic AI often operates based on events rather than manual triggers. How does this fundamental shift change the way we approach visibility and trust within a corporate system?
The shift from assisted work to delegated work represents a seismic change in the enterprise landscape that many organizations aren’t quite ready for. In the old model, we had the comfort of knowing that a human being made a deliberate choice, but now, an agent inside a platform like Dynamics 365 can research accounts or route service cases with almost no human initiation. This autonomy can feel like a black box if you aren’t careful, as a single event can trigger a sequence of actions that complete several steps before a human even realizes the process has begun. To maintain trust, we have to move away from simple login tracking and toward a model where the system records the specific reasoning and data triggers that led to a decision. Without this granular visibility, organizations risk losing the ability to trace decisions that used to be incredibly straightforward to audit.
Many organizations are eager to see the efficiency gains of AI, leading them to treat governance as a secondary concern. What are the specific risks of deploying an agent first and figuring out the guardrails later?
Treating governance as a compliance checkbox to be revisited after a “go-live” date is a recipe for high-stakes failure. Once an agent has the permission to modify inventory counts or approve purchase orders, you have effectively handed over the keys to your financial integrity. If you try to retrofit oversight after the agent is already in production, you are suddenly faced with a massive backlog of decisions that were never designed to be reviewed or corrected. This creates a dangerous lag where an agent could be making errors in financial records for weeks before anyone notices the pattern. Governance works best when it is woven into the agent’s DNA from day one, defining exactly what requires a human sign-off and what can be handled independently based on the organization’s risk tolerance.
You’ve mentioned that a modern audit trail needs to be more than just a standard system log. What does a “reasoning path” look like in practice, and why is it so vital for accountability?
A standard system log is essentially a digital receipt that tells you what happened, but it tells you absolutely nothing about why it happened. For agentic AI, we need an audit trail that captures the logic used to reach a specific outcome, including the specific inventory data or vendor history that influenced the decision. Imagine an agent autonomously reorders a massive amount of stock; a simple log shows the purchase order, but a reasoning-based trail shows that the agent saw a specific spike in demand data and calculated a new threshold. This depth allows a human reviewer to step in and understand the agent’s “thought process,” making it much easier to spot if the AI is hallucinating or relying on faulty information. It turns a mystery into a manageable business process that can be fine-tuned over time.
There is a common fear that heavy-handed governance will act as an obstruction to the very efficiency that justifies investing in AI. How do you define a “proportional” approach to oversight that keeps the system fast but safe?
The idea that you have to choose between speed and safety is a false choice that often stalls digital transformation. Effective governance is all about proportionality; you don’t need a human to review a low-impact, easily reversible action like triaging a customer service case. However, high-impact actions, such as vendor payments or changes to core financial records, absolutely warrant a checkpoint before they are finalized. We see this working beautifully when organizations set specific dollar thresholds—perhaps an agent can reorder inventory up to $500 automatically, but anything above that routes to a supply chain manager. By tying oversight to the level of risk, you allow the “heavy lifting” to happen at lightning speed while keeping a human in the loop for the decisions that actually move the needle.
One of the more technical challenges you highlight is “drift” in agent behavior. How does this happen in an ERP environment, and what can companies do to catch it before it compounds?
Agent behavior can drift over time as the underlying data in the ERP environment changes or as the agent encounters “edge cases” it wasn’t explicitly trained for. It’s a subtle process where an agent might start making slightly more aggressive inventory predictions or less accurate service routing because the real-world data no longer matches its initial configuration. Continuous monitoring and alerting are the only real defenses here, as they provide ongoing visibility into the agent’s performance long after the initial testing phase is over. If you don’t have these alerts in place, a small error in judgment can cascade across multiple records, creating a mess that requires a massive manual reconciliation project to fix. Catching drift early ensures that the agent remains an asset rather than a liability that slowly erodes your data quality.
For a company that already has agents active in production without a formal governance framework, what is the first practical step they should take to regain control?
The very first thing an organization needs to do is perform a comprehensive inventory of every active agent, identifying exactly what systems they can access and what actions they are authorized to take. It is incredibly common to find that an agent was granted broad system access during its setup phase “just in case,” and those permissions were never tightened back down. Once you have that inventory, you can start retroactively applying role-based permission boundaries so that a customer service agent can’t accidentally touch vendor payment data. From there, you can implement audit logging and establish a regular review cadence, much like the periodic access reviews we already use for human employees in traditional ERP security. It’s about taking those existing best practices and adapting them for a workforce made of software rather than people.
What is your forecast for the future of agentic AI within ERP systems over the next few years?
I predict we are moving toward a “center of excellence” model where centralized governance layers, like those being developed in Copilot Studio, will become the standard for every major enterprise. In the near future, organizations won’t just be managing five or ten agents; they will be managing hundreds of them across finance, HR, and supply chain functions simultaneously. This scale will be impossible to handle without automated governance that can apply data loss prevention policies and monitor for drift across the entire ecosystem from a single control point. We will also see “rollback capability” become a mandatory feature, allowing companies to undo a chain of autonomous actions with a single click, providing the ultimate safety net for delegated work. Ultimately, the companies that succeed will be those that treat data quality, system stability, and governance as a single, unified strategy rather than separate IT projects.