As companies give AI agents more authority to change purchase orders, move inventory and initiate financial transactions, new challenges are emerging. What should AI be allowed to do on its own, when should a person step in and who is responsible when something goes wrong?
This is particularly important in enterprise resource planning, the essential software that companies rely on for finance, procurement, supply chains, manufacturing, human resources and compliance. ERP systems sit close to the transactions and business rules that keep operations moving, which means an AI mistake can move quickly from a bad recommendation to a real business consequence.
In my recent Forbes article, Why ERP Became The Execution Layer, Not Just The System Of Record, I described how ERP is moving beyond recording transactions and reporting on the business toward helping companies recognize change and act on it. That raises the next question. When software can recommend, route or initiate action, who gives it authority and who remains accountable?
This is fundamentally an enterprise AI governance question. In ERP, I think the real issue comes down to decision rights, meaning the rules that define what a system can decide or do on its own, when a person must intervene and who owns the outcome. If those rules are unclear, AI can expose weak ownership and unclear processes faster and turn them into production risk.
Disclosure. KramerERP provides paid research, advisory and consulting services to technology companies, including ERP vendors listed in this article.
What Happens When AI Can Make Business Decisions
ERP systems can identify supplier risk, inventory shortages, margin pressure, financial close exceptions and workforce issues quickly. But seeing a problem sooner does not mean the organization is ready to respond.
A supplier delay, for example, can affect production, customer commitments, freight cost and working capital. The system may flag the risk quickly, while the response still stalls because procurement, finance, operations and logistics have not agreed on who owns the decision, what thresholds matter or when to escalate.
What I continue to see across enterprises is that technology often moves faster than the operating model around it. More dashboards or agents do not fix a process that falls back to email, meetings and spreadsheets when an exception occurs. The value is not faster awareness alone; it is a clearer path from an event to a decision and then to action.
What AI Should Be Allowed To Do On Its Own
Start by separating what AI can recommend from what it can prepare, route, approve or execute. Those actions carry different levels of risk and should not be governed the same way.
That matters because AI changes the consequences of weak governance. Bad data in a dashboard can mislead a user. Bad data connected to an AI agent can change a purchase order, release inventory, alter a production plan or trigger a financial transaction.
The same applies to unclear processes. If the business has not agreed on how a decision should be made, automation can move that ambiguity directly into action.
For every AI assisted business decision, leaders should define five things. They should know what data and business context the system can trust, who owns the decision, what the AI can do on its own, when an exception requires human review and who is accountable for the result. Those are business design questions first. Technology can enforce the rules, but it cannot invent them.
The ability to stop, correct or reverse an action belongs in the same design. If a team cannot explain why an action occurred or recover from a bad one, the use case is not ready for more autonomy.
Why ERP Still Matters When AI Acts Across Systems
AI driven work rarely stays inside one application. A sales change may start in a customer relationship management system, affect planning, trigger a supply chain response and end as a financial transaction in ERP.
ERP remains essential because it helps ensure transaction accuracy, enforces business rules, manages permissions, maintains financial oversight, promotes compliance and provides audit trails. Even though agents don’t need to work directly within the ERP system, it’s important that they follow established controls.
That creates a practical tradeoff. Agents built into an ERP suite can use the application’s existing business context and permissions. Agents that work across multiple platforms can reach more of the business, but they also create more integration, identity and governance questions.
The goal is not to keep AI inside ERP. It is to keep actions controlled as AI moves across systems. Accountability cannot disappear when work crosses an application boundary.
How ERP Vendors Are Bringing AI Closer To Business Decisions
ERP vendors are taking different paths, but the market is converging on one idea. AI is moving from answering questions to actively participating in the workflows and transactions that run the business.
SAP is bringing Joule agents into business processes. Oracle is extending AI Agent Studio across Fusion Applications. Microsoft is opening Dynamics 365 ERP data and business logic to agents, while Infor is combining AI agents, automation and process intelligence through Velocity Suite.
The same direction is taking shape across the ERP mid-market. IFS is deploying its Digital Workers across industrial workflows. Epicor is expanding Prism with industry focused agents. Acumatica is giving customers and partners tools to build AI-powered workflows, while QAD | Redzone is applying purpose-built agents to manufacturing operations.
The important point is not that ERP vendors have more AI. It is that AI is getting closer to the moment when a recommendation becomes a business action. That changes the risk and the responsibility that comes with it.
For buyers, agent counts, demos and feature lists are becoming less useful measures. I would look harder at what an agent can do, where people remain in control, how exceptions are handled, whether actions can be stopped or reversed and who is accountable for the outcome. That is where intelligent ERP will prove whether it can deliver real business value.
What CIOs Should Define Before AI Can Act
Start with one business decision rather than a broad AI program. Inventory reallocation, supplier exceptions, premium freight, financial close exceptions or compliance checks are useful tests because leaders can identify the event, the people involved and the expected result. Then map the decision from end to end and define what data it needs, who owns the outcome, what the system can do automatically and where a person approves, stops or overrides it.
Increase authority in stages. A system that can recommend does not automatically deserve permission to approve or execute. Test exceptions, conflicting data, policy boundaries and recovery paths before expanding its authority because those hard cases reveal more about readiness than a clean demo.
Measure operational value, not feature adoption. Faster exception resolution, lower premium freight, fewer close delays, better service levels and less manual work are stronger evidence that the approach is working. Prepare the people who remain accountable as well. Employees need to understand what the system can do, when they are expected to intervene and who owns the result. Technology enables transformation. People determine the outcome.
The Real Test Is Knowing When AI Should Stop
ERP is a great example of this governance challenge because it handles transactions related to cash, inventory, suppliers, employees and customers. This extends beyond ERP. AI agents that modify records or start work need a responsible owner to maintain oversight.
The goal is not to give every process more AI. It is to define the rules before more decisions are automated. Those rules should cover what the agent can do, what requires human judgment and how the business stops or corrects an action when something goes wrong.
The test for enterprise AI is not whether an agent can act. It is whether the company knows when it should act, when it should stop and who owns the outcome.
