Differentiating Predictive, Generative, and Agentic AI

Predictive AI analyzes historical and current data to identify patterns and estimate what is likely to happen next.

Some AI systems are designed mainly to classify, detect, recommend, or predict. Generative AI goes a step further by creating new content.

For example, an implementation consultant might ask a generative AI system to:

• Draft a workshop agenda for a finance design session
• Summarize meeting notes and extract action items
• Create sample master data for testing
• Generate a first draft of a business process document
• Rewrite a technical explanation for end users
• Produce SQL, JavaScript, or configuration-related code
• Suggest possible causes of an integration error

The AI is not simply retrieving a single stored answer. It is generating a new response by combining patterns it learned during training with the information provided in the prompt.

Agentic AI refers to AI systems that can work toward a goal by planning and carrying out a sequence of actions, rather than only producing a single prediction or response.

A traditional generative AI interaction might look like this:

Create a test script for invoice processing.

The AI produces the document, and the user decides what to do next.

An agentic AI system may be given a broader objective:

Review failed invoice-processing test cases, identify likely causes, create follow-up tasks, and notify the responsible team members.

To complete that objective, the AI agent might:

1. Retrieve the failed test results.
2. Group similar errors together.
3. Review system logs or documentation.
4. Identify likely root causes.
5. Draft recommended corrective actions.
6. Create tasks in a project-management system.
7. Assign or route those tasks to the appropriate people.
8. Monitor whether the tasks are completed.
9. Escalate unresolved issues.

The key difference is that agentic AI does not only generate content. It can decide which actions are needed, use available tools or systems, evaluate the results, and continue working toward the assigned goal.

An AI agent may use several technologies together, including:

• Large language models
• Machine learning models
• Business rules
• APIs
• ERP transactions
• Search tools
• Workflow systems
• External applications

The agent may also maintain context about what it has already done, which tasks remain incomplete, and whether it needs additional information or human approval.

Traditional automation follows a predefined sequence. For example:

1. Receive an invoice.
2. Validate required fields.
3. Route it for approval.
4. Post it after approval.

The workflow is designed in advance, and the system follows the same defined path each time.

Agentic AI may have more flexibility in determining how to reach an outcome. For example, an agent investigating a blocked invoice might decide to:

• Compare the invoice with the purchase order.
• Review the goods receipt.
• Check supplier master data.
• Search prior support cases.
• Contact the purchasing team for missing information.

The precise path may vary depending on what the agent discovers. However, agentic AI still operates within the permissions, tools, policies, and controls provided to it. It should not be treated as an unrestricted autonomous employee.

In an ERP environment, organizations may require human approval before an agent can:

• Post a financial document
• Change configuration
• Modify master data
• Release a payment
• Send an external communication
• Close an audit or compliance issue

This is often described as keeping a human in the loop.

Imagine a company is implementing a new ERP system and wants to improve invoice processing.

A traditional rule-based solution might include rules such as:

• Block invoices above a certain value
• Reject invoices without a supplier number
• Require approval when the invoice and purchase order differ by more than 5%
• Flag invoices with the same supplier, amount, and date

A machine-learning system could analyze thousands of historical invoices and identify less obvious patterns associated with errors or duplicate payments.

For example, it may determine that a new invoice has an 87% likelihood of being a duplicate based on its supplier, amount, reference number, posting date, and similarity to a previously processed invoice.

An LLM could help the user understand the result by presenting it in plain language: This invoice may be a duplicate because it closely matches an invoice from the same supplier that was posted three days ago.

Generative AI could then create useful content, such as:

• A summary of the issue
• A recommended review checklist
• A message to the accounts payable specialist
• A draft note for the invoice record

An agentic AI system could take the process further. Within its authorized permissions, the agent might:

1. Identify the potentially duplicated invoice.
2. Retrieve the matching invoice and related purchase order.
3. Compare the supplier, amount, reference number, and line items.
4. Check whether either invoice has already been paid.
5. Place the newer invoice on temporary hold.
6. Create a review task for the accounts payable specialist.
7. Draft an explanation of the suspected duplicate.
8. Attach the relevant supporting documents.
9. Notify the responsible user.
10. Monitor the task and release or reject the invoice after receiving an approved decision.

• Machine learning identifies the pattern.
• Predictive AI assigns a probability or risk score.
• An LLM interprets or explains the information in natural language.
• Generative AI creates a summary, recommendation, or communication.
• Agentic AI coordinates and performs the steps required to move the process toward completion.