Blog
The Foundations Businesses Need for AI Agents to Use Enterprise Systems
Businesses are exploring how AI agents can move beyond writing content and answering questions to retrieving information from ERP and CRM systems and carrying out business tasks. For example, an agent could check an opportunity in the CRM, compare it with inventory and order information in the ERP, and suggest the next action to a sales representative.
Making this work requires more than choosing an AI model. Businesses need integration that provides access to the right data, permissions that control what agents can do, and operational processes that make their actions traceable. They also need to assess whether using AI for a particular task makes sense in terms of cost and accuracy.
This article explains the foundations needed for AI agents to use enterprise systems, and how to combine AI, data integration, business process outsourcing (BPO), and other approaches according to the needs of each task.
Contents
What Are AI Agents?
AI agents are systems that select information and tools and carry out multiple steps to achieve a given objective. Compared with automation that follows a fixed sequence, agents can choose their next action based on the information they receive.
Business applications could include reviewing an inquiry, searching customer information and support history, and drafting a response. Another example is bringing together opportunity, inventory, and order information to recommend an action.
However, reading information, making recommendations, and updating business systems require different permissions and have different consequences. Businesses need to define the scope of execution for each process.
Decide Whether AI Is Right for the Task
When considering AI, the question is not only “Can AI do this?” Businesses should also ask: What is the most efficient way to complete this task while meeting the required quality standard?
For example, if sales order data arrives in the same format every day and needs to be registered in the ERP, API or ETL integration may be a suitable option. If a task involves interpreting inquiries in different formats, finding relevant information, and drafting a response, AI may be worth evaluating.
For work that requires coordination, an understanding of individual circumstances, or decisions with significant consequences, retaining internal staff or BPO support may be more appropriate. The following table provides a starting point for assessing tasks.
| Task Characteristics | Approaches to Consider | What to Evaluate |
|---|---|---|
| Transferring data with fixed formats and processing rules | API, ETL, or other data integration | Data formats, processing volume, error handling, and operating costs |
| Processing requests or entering records through predefined steps | Standard system features, workflows, or RPA | Coverage of standard features, exception handling, and maintenance effort |
| Interpreting text, organizing information, and drafting recommendations | Generative AI or AI agents | Accuracy with real business data, review effort, and usage-based costs |
| Handling individual circumstances, coordinating stakeholders, and making accountable decisions | Internal staff or BPO combined with supporting tools | Required expertise, service quality, handoffs, and total process costs |
These approaches can work together. AI could classify an inquiry, data integration could retrieve the necessary information, and difficult cases could be passed to an employee or BPO team. Defining which steps use AI, which follow established rules, and which require human judgment helps improve the efficiency of the overall process.
Compare Cost and Accuracy Across the Entire Task
Low AI usage fees do not necessarily mean lower operating costs. If employees spend substantial time checking and correcting the results, the total workload may not decrease. Conversely, a solution with higher implementation costs may deliver value if it consistently reduces manual work across a large volume of tasks.
An evaluation should include the following:
- Total cost: Implementation, system connections, AI usage, maintenance, monitoring, review, and corrections.
- Accuracy: The proportion of tasks that meet the completion criteria, and the consequences of errors.
- Processing capacity: The volume the solution can handle during normal and peak periods, and the time needed to complete tasks.
- Exception handling: The proportion of cases that cannot be completed automatically, and the effort required to hand them over to a person.
- Manageability: The ability to record actions, investigate problems, and accommodate changes to business rules.
For a small-scale trial, use representative business data and exception cases, and compare the proposed approach with the current method under equivalent conditions. The central measure should be the cost and time required to complete one task to the required quality standard.
For customer inquiries, for example, measure more than the time needed to generate a draft. Include checking, editing, sending, and any follow-up work. A fluent AI response and a correctly completed business task are different things and should be evaluated separately.
Five Foundations AI Agents Need to Use Enterprise Systems
Once a business has decided which tasks to assign to AI, it needs to identify the system capabilities that support those tasks.
| Foundation | What to Establish |
|---|---|
| Data management | Define where data resides, what it means, when it was updated, and which system is authoritative. |
| APIs and data integration | Make information retrieval and business operations available through controlled connections. |
| Identity and access management | Verify user and agent permissions and restrict access to what each task requires. |
| Execution records and operations management | Trace results and provide ways to stop, investigate, and recover from abnormal operations. |
| Business rules and approvals | Define when automatic execution is permitted and when human review is required. |
1. Define Data Meaning and Authoritative Sources
Customer codes may differ between the CRM and ERP. Inventory records may have unclear update times. Old and current versions of contracts may exist side by side. In these situations, retrieving information does not necessarily lead to sound decisions.
Start by defining the data needed for the target process and how records correspond across systems. Then identify the authoritative source for each type of information—its system of record.
For example, the CRM might manage opportunity progress, the ERP might hold confirmed orders and billing information, and a document management system might hold approved contracts. The appropriate ownership depends on the company’s processes and system architecture.
Not all data needs to be stored in one place. What matters is retrieving the right information from the appropriate source, with its origin and update time available for verification. Our article on why data integration matters before AI adoption explains this principle in more detail.
2. Prepare the Operations AI Can Use Through APIs and Data Integration
When AI agents use business systems, define the available operations and their inputs and outputs. Examples include checking inventory or creating a draft sales order.
Choose connection methods according to data freshness, processing volume, and other requirements. APIs may support real-time inquiries, while file integration may suit scheduled bulk transfers. Our article on API integration vs. file integration discusses these options and the constraints of existing systems.
Also design validation for required fields and values, controls to prevent duplicate records, and procedures for retrying failed operations. Even when AI proposes the data to register, business constraints should be checked by the integration process or the destination system.
Reusable connections and data transformations can support additional AI use cases as they emerge. Freedom’s S Series provides reusable assets for common SAP integration processes.
3. Control AI Agent Identities and Permissions
An AI agent supporting sales may need to read customer information, but it does not need permission to change a supplier’s bank account. Even read-only applications require controls over access to confidential and personal information.
Define whose request the agent is acting on, which data it can read, and whether it can create, update, or delete records. Grant only the permissions required.
Enforce permissions in connected systems and integration platforms as well as through instructions to the AI. For example, documents or emails may contain malicious instructions. The system should be designed so that content an agent reads cannot, by itself, expand the scope of its permitted actions.
4. Record Actions and Prepare to Respond to Problems
When an AI agent uses multiple systems, a conversation history alone may not provide enough information to identify where a problem occurred.
Link records of the requester, data sources, operations invoked, approvals, and system update results so that the full sequence can be traced. Define what information to log, how long to retain it, and who can access the logs.
Also establish procedures for stopping abnormal operations, notifying the responsible person, and checking or recovering partially completed tasks. As the scope of automatic execution expands, review the monitoring and response arrangements.
5. Define Conditions for Automatic Execution and Human Approval
For each process, define an operating scope such as “read only,” “drafts and recommendations only,” “execute after human approval,” or “execute automatically within defined conditions.”
Checking inventory and confirming a purchase order have different business consequences. Conditions might include transaction value, business partner, or processing volume. Cases outside those conditions—or cases with insufficient information—can be referred to a person.
At the approval stage, reviewers should be able to see the transaction or record involved, the amount, proposed changes, and supporting information. Keep approval tied to the action actually executed, including requiring renewed approval if the proposed action changes.
Start Small and Expand After Evaluating Quality and Cost
Begin by selecting one process and establishing its current processing time, costs, error rates, and rework. Then compare AI, existing automation, BPO, and other options using actual business data.
One approach is to start with information retrieval or recommendations, where the impact is easier to assess. Evaluate whether the approach delivers benefits after review and correction are included. Expand execution gradually once the necessary permissions, approvals, and operating arrangements are in place.
Before rebuilding systems across the enterprise, identify the specific data, connections, and controls missing from the target process. This makes it easier to prioritize investment.
Conclusion: Choose the Right Approach and Build the Foundations to Support It
Using AI agents in business operations requires clear data ownership, APIs and data integration, permissions, execution records, and approval rules.
It also requires comparing AI, data integration, standard system features, BPO, and human work for each task. Success should be measured by whether the overall process becomes less costly or faster while maintaining the required quality.
Deciding where to use AI, which steps to leave to existing systems or people, and how to connect those approaches provides the foundation for making AI a practical part of daily operations.