5 Questions Business Leaders Should Ask Before Adding AI to Their ERP

Artificial intelligence is becoming part of modern enterprise resource planning (ERP) systems. From summarizing financial information to assisting with invoice processing, forecasting, and business reporting, AI can help employees work with business data in new ways.
However, adding AI to an ERP system is not simply a software upgrade. ERP platforms manage important business processes, including finance, procurement, inventory, sales, and supply chain operations. An AI feature that produces inaccurate information or operates with inappropriate permissions can create problems beyond the technology department.
Microsoft's official documentation for Dynamics 365 and Power Platform Copilot, for example, explains how AI features use organizational data, follow existing access permissions, and operate within defined security and privacy controls. Microsoft also cautions that AI-generated responses are not always factual. These considerations make business readiness as important as the technology itself.
Before investing in AI for ERP, business leaders should ask five questions.

1. Which Business Problem Are We Trying to Solve?** The first question should be about the business outcome, not the AI feature. Organizations may be interested in AI because it is receiving attention across the industry. But a technology investment needs a clear operational purpose. Leaders should identify a specific problem that currently consumes time, creates errors, delays decisions, or limits visibility. For example, a finance team might spend considerable time reviewing supplier invoices, matching purchase orders, and investigating discrepancies. A procurement team might struggle to identify unusual spending patterns across departments. Inventory managers may need better ways to investigate recurring stock shortages or excess inventory. These situations may offer opportunities for AI assistance, but each requires a different approach. **_A useful starting point is to document the existing process:**

• How is the task performed today?
• How much staff time does it require?
• Where do errors, delays, or repeated manual checks occur?
• Which decisions require professional judgment?
• What measurable improvement would justify the investment?

Practical recommendation: Establish a baseline before introducing AI. Depending on the use case, this could include invoice processing time, exception rates, reporting turnaround, or the time required to investigate inventory discrepancies.
The proposed AI solution can then be evaluated against the existing process rather than against a general promise of greater efficiency.
A clearly defined problem also makes it easier to reject features that look impressive but do not address a meaningful business need.

2. Is Our ERP Data Reliable Enough for AI?
AI-assisted ERP processes depend on the quality, relevance, and accessibility of the information they use. If supplier records are duplicated, product descriptions are inconsistent, or historical transactions contain errors, an AI feature may produce misleading summaries or recommendations.
Consider an Oracle Cloud ERP consulting firm whose ERP system contains multiple records for the same supplier. An AI-assisted spending analysis might present an incomplete picture if purchases are split across those records. Similarly, outdated product lead times could undermine an inventory recommendation, even if the recommendation is presented clearly.
The issue is not limited to incorrect data. Information can also be incomplete, outdated, inconsistently formatted, or unavailable to the users and systems that need it.
Before implementation, businesses should examine:
Master data: Are customer, supplier, product, and account records accurate and consistent?
Transaction history: Are historical records complete enough for the intended use?
Data ownership: Who is responsible for correcting errors and maintaining records?
System integration: Can the AI feature access the relevant ERP information reliably?
Access permissions: Can users access only the information they are authorized to see?
Not every AI use case requires the same level of historical data. A tool that drafts a description from approved product information has different requirements from a forecasting model that relies on historical demand and supply patterns.
The National Institute of Standards and Technology (NIST), through its AI Risk Management Framework, identifies validity, reliability, security, transparency, and accountability among the characteristics organizations should consider when managing AI risks.
Practical recommendation: Conduct a data-readiness assessment before selecting a use case. Resolve material data-quality issues first, and document any limitations that could affect the results.

3. How Will We Protect Business Data and Control Access?
ERP systems often contain commercially sensitive information, including payroll records, customer details, supplier contracts, financial results, pricing, and purchasing activity. Connecting AI to this information requires a clear understanding of how data is accessed, processed, stored, and protected.
Leaders should not assume that every AI product handles organizational information in the same way. Security and privacy arrangements vary by product, feature, configuration, deployment model, and contractual terms.
Microsoft's official Copilot documentation for Dynamics 365 and Power Platform describes how these features use existing permissions to limit access to organizational data. It also explains the provider's stated data-handling and security practices. Organizations should verify the applicable documentation and configuration for the specific feature they intend to use rather than generalizing those statements to all AI tools.

Before approving an implementation, ask:

_- What business information will the AI feature access?

• Does it inherit existing ERP roles and permissions?
• Where is the information processed and stored?
• Is data shared with external services or third-party extensions?
• What are the applicable retention, deletion, and data-use terms?
• How are user activity, access, and potential security incidents monitored?
• Do the arrangements meet relevant contractual, regulatory, and internal requirements?_ These questions are especially important when AI connects to several systems or uses third-party integrations. For example, a purchasing assistant may need access to supplier prices and order histories. It may not need access to employee payroll information. Limiting access to the information required for a specific task reduces unnecessary exposure. **Practical recommendation: **Involve IT security, data protection, legal, and ERP administrators in the evaluation. Test permissions using representative user roles, and verify how the feature behaves when a user attempts to access information outside their authorization. Security should be reviewed before deployment and revisited when integrations, permissions, or AI capabilities change.

4. Where Should Human Review Remain Mandatory?
AI can help summarize information, identify patterns, draft explanations, and recommend possible next steps. However, a confident-looking response is not proof that the underlying information or conclusion is correct.
Microsoft explicitly notes in its Dynamics 365 and Power Platform Copilot documentation that AI responses are not always factual. NIST also identifies risks associated with people placing excessive trust in generative AI outputs.
The appropriate level of human oversight depends on the consequences of an error.
For instance, AI might prepare a draft explanation of a monthly expense variance. A finance professional can verify the figures and approve the final report. By contrast, automatically releasing payments, changing supplier bank details, approving significant purchases, or posting material financial adjustments may require stricter controls and authorized human approval.
A sensible implementation distinguishes between assistance and authority.
Businesses should determine:
_

• Which outputs can be generated automatically?
• Which outputs require review before being used?
• Which transactions need approval from an authorized employee?
• How will employees verify AI-generated figures and recommendations?
• What happens when the system produces an uncertain, incomplete, or conflicting result?
• Can users correct errors and escalate unusual cases?

For repeatable, low-risk tasks, a sample-based review may be appropriate if it is supported by testing and internal controls. For high-impact decisions, individual approval or additional verification may be necessary.
The exact controls should reflect the organization's risk tolerance, applicable obligations, and the capabilities of the selected ERP feature.
Practical recommendation: Define approval thresholds, exception-handling procedures, audit trails, and escalation paths before moving from a pilot to production. Review the controls whenever the scope of automation expands.
The goal is not to eliminate human involvement. It is to direct human attention toward decisions where judgment, accountability, and verification matter most.

5. How Will We Measure Success and Manage the Long-Term Cost?
An AI project should be evaluated as a business investment, not just as a technology deployment.
The total cost may extend beyond the initial license or subscription. Depending on the solution, it can include configuration, integration, data preparation, testing, employee training, security reviews, ongoing monitoring, and support. Additional costs may arise when usage grows or when business processes need to be redesigned.
Before proceeding, leaders should agree on how the project will be evaluated.

Potential measures include:

• Average time required to process an invoice.
• Percentage of transactions requiring manual intervention.
• Time taken to prepare and validate management reports.
• Frequency of incorrect recommendations or classifications.
• Number of exceptions escalated to employees.
• Adoption rates among the intended users.
• Total operating cost compared with the existing process. These are suggested evaluation measures, not guaranteed outcomes. The appropriate metrics depend on the use case, and improvements should be demonstrated using actual business data. For example, if an AI-assisted invoice workflow is intended to reduce manual review, the pilot should measure both processing time and the accuracy of the resulting records. A faster workflow is not successful if it increases posting errors or creates additional work for the finance team. Leaders should also consider what happens after launch. AI features and their underlying platforms may change, business data may evolve, and user behavior may differ from pilot conditions. Monitoring is therefore an ongoing responsibility rather than a one-time implementation task. The NIST AI Risk Management Framework provides a useful structure for this work through four functions: Govern, Map, Measure, and Manage. These functions support the identification of responsibilities, assessment of risks, evaluation of system performance, and continued management of issues throughout the AI lifecycle. ERP implementation and consulting teams, including SoftArt Solutions, can be relevant to this evaluation when businesses need to assess their existing processes, integration requirements, and implementation priorities before introducing AI capabilities. Practical recommendation: Begin with a limited pilot, document the baseline, set acceptance criteria, and agree on a review period. Expand only when the results, controls, user feedback, and ongoing costs justify the next step.

A Practical Way to Begin
Business leaders do not need to introduce AI across the entire ERP environment at once. A focused approach allows the organization to test assumptions and identify operational issues before expanding the scope.
**A practical sequence is:

• Select one use case**:Choose a clearly defined business problem with an identifiable owner.

• Assess data and permissions: Confirm that the required information is reliable and appropriately protected.

• Define controls: Specify the tasks AI can support, the decisions employees must approve, and the procedures for handling exceptions.

• Run a controlled pilot: Test the feature against realistic scenarios, including unusual cases and potential errors.
  Evaluate the evidence: Compare results with the baseline, document limitations, and calculate the full cost of operating the solution.

• Review before scaling: Expand only when performance, security, accountability, and business value meet the agreed requirements.

This approach also helps organizations distinguish between capabilities already included in their ERP subscription and those that require additional licensing, configuration, or third-party services. Product documentation and commercial terms should be checked for the specific deployment.
Conclusion
Adding AI to an ERP system can create opportunities to improve reporting, support repetitive tasks, and help employees interpret business information. But the value depends on more than the availability of a new feature.
Business leaders should begin with a clearly defined problem, examine data quality, verify security and access controls, establish appropriate human oversight, and measure results against agreed business objectives.
The strongest starting point is not the most ambitious AI project. It is a manageable use case with reliable data, clear accountability, measurable success criteria, and a plan for ongoing review.
By answering these five questions before implementation, organizations can make more informed decisions about where AI belongs in their ERP environment—and where human judgment should remain central.