AI capability selection principles

Before selecting a solution, understand the features available in Pega Infinity™ 26. Each feature addresses a specific category of business problems. Selecting the appropriate option keeps the solution governable, explainable, and cost-effective without introducing AI where it is not required.

Pega Infinity 26 provides the following AI and automation options:

• Deterministic workflow and decisioning: Use Case processing, automation, decision tables, business rules, and direct integrations for requirements with defined behavior.
• Predictive and adaptive decisioning: Use Predictive Models to estimate outcomes, Adaptive Models to learn from captured responses, and decision strategies to combine models, rules, and data.
• Generative and knowledge features: Use bounded generative operations for defined outputs. Use knowledge features to provide answers grounded in governed sources.
• Agentic orchestration: Use Agents when an application must select from multiple approved tools or actions and adapt to intermediate results.
• Specialized AI: Evaluate document-processing, voice, and process-analysis features when the input or outcome requires these functions.
• Interoperability: Use Model Context Protocol (MCP) for Agent-to-tool access. Evaluate Agent2Agent (A2A) communication when Agents must communicate across a supported boundary.

These options address different business requirements and are not interchangeable. For example:

• A request for answers to policy questions does not necessarily require an Agent.
• A workflow that performs one summarization operation does not require agentic orchestration.
• A requirement based on stable business rules does not require generative AI.

Treat a request to "add AI" as an architecture decision. Select the approach that provides the required outcome with the lowest appropriate level of complexity.

The role of a Lead System Architect (LSA) includes determining whether an AI feature is appropriate, what controls it requires, and how the application responds when the feature fails.

Moving from content generation to autonomous action increases enterprise risk. Each AI feature introduces inference expense, governance overhead, and consequences if the output is incorrect. Use a structured framework to determine which feature to use, when to use it, and who owns the resulting decisions.

Can deterministic processing produce the required outcome?

If stable rules and validated data consistently produce the required outcome, use deterministic workflow, decisioning, or integration.

Evaluate AI when the requirement includes one or more of the following conditions:

• Interpreting or generating natural language
• Extracting meaning from variable-format documents
• Predicting an outcome from data
• Selecting dynamically from multiple approved actions
• Adapting a response or action to intermediate results

Deterministic processing is not error-free. Its behavior still depends on correct requirements, rules, data, configuration, and integrations.

What outcome does the user need?

The user might need information, a result, a prediction, or an action. Match the required output to the appropriate feature to avoid introducing a mechanism that is more complex than the requirement.

Does the requirement involve a bounded operation or a multistep objective?

Determine whether the requirement can be fulfilled through a single, well-defined operation or requires coordinated execution of multiple actions and decisions.

Bounded operation: Use a Pega GenAI Connect Rule when the requirement consists of one clearly defined operation with known inputs and expected outputs. Examples include extracting information, summarizing content, drafting a response, classifying an item, or recommending a next step from predefined criteria.

Multistep objective: Evaluate an Agent when the requirement includes multiple coordinated actions, context-based selection of tools or workflows, adaptation to intermediate results, or exception handling. Examples include processing complex documents, resolving customer inquiries across multiple systems, and coordinating verification, routing, and approval activities.

Are the external tool and its inputs known in advance?

Determine whether the required tool invocation can be defined during design time or must be selected dynamically at runtime. This distinction helps you choose between a predictable workflow-driven approach and an Agent-driven approach that uses reasoning to select the appropriate tool and inputs.

Known tool and defined inputs: When the external tool and its required inputs and outputs are known, invoke the tool directly from the workflow by using an appropriate MCP integration or workflow connector. A predefined invocation provides a predictable processing path without requiring Agent reasoning. This approach reduces complexity, cost, and latency.

Dynamic tool selection: When the appropriate tool or its inputs depend on runtime context, use an Agent with access to approved MCP tools. The Agent interprets the request and selects a tool based on the available context.

Does the requirement cross a system or Agent boundary?

Determine whether the solution requires interaction across application, system, or Agent boundaries. The communication pattern helps identify whether the architecture needs secure access to tools and data, task delegation between Agents, or a combination of both.

External system requiring access to Pega tools or data: Evaluate MCP when an external AI system requires governed access to Pega tools or data. Expose only approved functions as tools, and apply authentication, authorization, and auditing controls.

Agent delegating work to another Agent: Evaluate A2A communication when one Agent must delegate a task to another specialized Agent. A2A communication supports information exchange, task delegation, and result delivery across supported boundaries.

Key distinction: MCP gives an Agent access to tools and contextual data. A2A communication supports task delegation and collaboration between independent Agents. A solution can use both protocols when the architecture requires both interaction patterns.

Is the engagement reactive or proactive?

Determine who initiates the interaction. The engagement pattern influences the level of autonomy, business controls, and governance required in the solution, and helps identify the appropriate Pega capabilities to use.

Reactive engagement: Use a reactive pattern when a person initiates an inquiry, request, or action. Depending on the requirement, the application can use an Agent, Knowledge Agent, Pega GenAI Connect, or deterministic workflow to provide information, perform a bounded operation, or route work for human action.

Proactive engagement: Use a proactive pattern when the application initiates contact to request information or prompt an action. Apply business controls, human oversight, transparency, and auditability because the application initiates the interaction.

Does the requirement involve a specialized input format?

The type of input the application must process can influence the capabilities required in the solution. Evaluate whether the requirement involves unstructured documents, voice interactions, or structured data to determine the most appropriate processing approach.

Unstructured documents or images: Evaluate document-processing features when the application must analyze PDFs, scanned forms, images, or other variable-format documents. Document processing can extract and validate data, populate Case fields, classify documents, and pass results to a workflow.

Examples include invoice extraction, claim-form processing, contract analysis, and document-based questions.

Voice or speech input: Evaluate voice features when the requirement involves live conversations, hands-free interaction, or real-time transcription.

Structured data: Use standard processing when the input comes from defined sources, such as application forms, APIs, databases, or Data Pages. Focus the architecture on workflow, decision logic, integration, and the appropriate use of Agents or Pega GenAI Connect

The framework requires periodic reassessment as features, models, business requirements, risks, and organizational experience change.

Include the following controls:

• Version tracking: Record the feature, model, prompt, and configuration used in production.
• Performance monitoring: Track quality, cost, latency, exceptions, and business outcomes.
• Controlled testing: Compare alternative features or configurations when appropriate.
• Feedback and drift analysis: Use corrections, overrides, failures, and user feedback to identify degradation.
• Periodic reassessment: Reapply the framework when requirements, risks, or available features change.
• Retirement criteria: Define when to update, replace, restrict, or retire the selected feature.

The following figure summarizes a simplified sequence for selecting an implementation approach:

Design AI governance into the solution from the beginning. Do not add governance only after deployment.

Keep each AI-generated recommendation, decision, and action transparent, accountable, and traceable. These controls help the organization understand what occurred, apply human oversight, investigate failures, and implement its compliance strategy.

Record the following information for each AI-driven outcome:

• Input transparency: The data provided to the AI system.
• Retrieved context: The context or knowledge sources used to produce the outcome.
• Output visibility: The content, recommendation, or decision that the AI system produced.
• Action recording: The action taken in response to the output.
• Decision ownership: The person or role that approved, rejected, or overrode the decision.
• Exception handling: The fallback or exception path used.
• Audit trail: The records required for investigation, traceability, and continuous improvement.

Check your knowledge with the following interaction: