Hybrid AI Workflows: Choosing the Right Pattern for AI and Workflows - Workflow Automation CoE

[Lisa Holenstein]

Workflow Automation CoE > Hybrid AI Workflows

Goal 

In this Workflow Automation Center of Excellence article, you will learn how to choose the right automation pattern when combining AI capabilities with deterministic workflows. You will understand when to build AI-first workflows that leverage flows as tools versus when to build deterministic workflows that call AI Agents for specific steps. 

Technology 

Workflow Studio is our central, low-code workflow automation builder that consolidates flow, subflow, playbook, decision table, and action design into a single environment. 

The following product features enable hybrid workflow automation: 

Feature	Description
Execute Skill action	Flow action that invokes a published skill from Now Assist Skill Kit
Use an AI Agent action	Flow action that invokes a published AI agent from AI Agent Studio
AI Agent Studio	Design environment for creating and managing AI agents
Now Assist Skill Kit	Design environment for creating custom GenAI skills

Understanding the Two Patterns 

ServiceNow now offers two approaches for combining AI with workflow automation. Understanding when to use each pattern is critical for successful implementation. 

Pattern 1: AI-First Workflows (Agent as Orchestrator) 

In this pattern, an AI Agent serves as the orchestrator, reasoning through requests and calling flows and subflows as tools to accomplish tasks. 

How it works: The AI Agent receives a request, reasons about which tools (including flows) to use, creates a plan, and executes by calling those tools. After each tool invocation, the agent can modify its plan and repeat until the desired outcome is reached. 

Characteristics: 

Aspect	Description
Orchestration	AI Agent decides what to do and when
Flexibility	High; agent adapts based on context and results
Human involvement	Agent decides if and when to prompt a human
Predictability	Lower; outcomes may vary based on reasoning

Pattern 2: Deterministic Workflows with AI Steps (Workflow as Orchestrator) 

In this pattern, the deterministic workflow remains the orchestrator, with AI Agents or Skills called at specific steps where adaptive intelligence is needed. 

How it works: The workflow defines the sequence of steps, conditions, and human interactions. At designated points, the flow calls an AI Agent or Skill for tasks requiring reasoning and/or non-deterministic behavior, then continues with deterministic logic. 

Characteristics: 

Aspect	Description
Orchestration	Workflow controls sequence and decisions
Flexibility	Targeted; AI flexibility only where specified
Human involvement	Workflow explicitly defines where humans participate
Predictability	High; workflow structure ensures consistent paths

Decision Framework: When to Use Each Pattern 

Use the following questions to determine the right pattern for your use case. 

Choose AI-First Workflows (Agent as Orchestrator) When: 

1. The task is undefined or highly variable 
   • Example: "Automate a timesheet based on a week's activities" 
   • Example: "Determine and execute follow-ups from a business dinner" 
   • The agent needs to reason about what steps are even required depending on the situation 
2. Context determines the path 
   • The "right" approach cannot be predetermined 
   • Each execution may require different tools or sequences 
   • Discovery of information changes what actions are needed 
3. Conversational interaction is primary 
   • Users interact through natural language 
   • The agent needs to ask clarifying questions 
   • Response generation is the main output 
4. Speed of implementation over governance 
   • Rapid prototyping or experimentation 
   • Internal productivity tools with lower risk 
   • Controlled user populations who can validate outputs 

Choose Deterministic Workflows with AI Steps When: 

1. Rules and compliance drive the process 
   • Regulatory requirements mandate specific steps 
   • Approvals must follow defined hierarchies 
   • Audit trails must be explainable and reproducible 
   • Example: Provisioning a laptop to a new employee 
   • Example: Processing an expense report approval 
2. The process is well-defined with targeted AI needs 
   • 80% of steps follow known business rules 
   • Specific steps benefit from AI (summarization, classification, drafting) 
   • AI enhances the process; it does not define it 
3. Existing flows should be enhanced, not replaced 
   • Teams have invested in proven workflow automation 
   • Rebuilding as agentic workflows would duplicate effort 
   • AI capabilities should augment current investments 
4. End-to-end visibility is critical 
   • Teams need to understand the complete automation 
   • Troubleshooting requires clear execution paths 
   • Separate agents "listening" for events creates complexity 
5. Human touchpoints are fixed by policy 
   • The workflow must define exactly when humans participate 
   • Certain decisions cannot be delegated to AI reasoning 
   • The organization requires "human in the loop" at specific stages 

Real-World Scenario: Financial Services Operations 

This example illustrates why many enterprise use cases benefit from deterministic workflows with AI steps. 

The Challenge 

A financial services team has a flow that processes ACH disputes. They want to add AI capabilities to accelerate resolution. They have two options: 

Option A: AI-First (Agent as Orchestrator) Build an AI Agent that handles ACH disputes end-to-end, using flows as tools. 

Option B: Deterministic with AI (Workflow as Orchestrator) Enhance the existing flow to call an AI Agent at specific steps. 

Why Option B Was Chosen 

The team chose to embed the AI Agent within their existing flow for these reasons: 

Factor	AI-First Approach	Deterministic with AI
Compliance	Agent decides dispute handling sequence	Flow enforces required regulatory steps
Visibility	Agent and flow are separate entities	Complete end-to-end process visible in Workflow Studio
Maintenance	Two separate automations to maintain	Single flow with AI step to maintain
Predictability	Agent may handle similar disputes differently	Consistent handling with AI-enhanced intelligence
Existing Investment	Would require rebuilding current automation	Enhances proven flow with new capability

The Implementation 

With the Use an AI Agent action in Workflow Studio, the team: 

1. Kept their existing dispute intake and validation steps (deterministic) 
2. Added an AI Agent step where reasoning about dispute resolution is needed 
3. Continued with deterministic steps for approvals and notifications 
4. Maintained full audit trail with clear AI contribution visibility 

Hybrid Patterns in Practice 

Many implementations combine both patterns based on the specific requirements of different process phases. 

Example: Investigative Case Management 

In investigative case management, different phases have different requirements: 

Phase	Recommended Pattern	Rationale
Case Intake	Deterministic with AI	Validation rules are fixed; AI helps with document extraction and case narrative drafting
Lead Triage	Deterministic with AI	Escalation criteria are policy-driven; AI assists with entity matching and data enrichment
Active Investigation	Hybrid consideration	Dynamic nature may benefit from AI-first for certain investigative tasks
Reporting	Deterministic with AI	Report structure is defined; AI generates content within the template
Post-Investigation	Deterministic	Archival, retention, and legal integration follow strict rules

Key Insight: It Is Not Either/Or 

The most effective implementations often use deterministic workflows as the primary orchestrator with AI agents called for specific capabilities: 

• Summarization: AI Agent summarizes case notes, emails, or documents 
• Classification: AI Agent determines category, priority, or routing 
• Content generation: AI Agent drafts responses, reports, or recommendations 
• Reasoning: AI Agent analyzes data to suggest next steps (human reviews) 

Implementation Guidance 

Starting with Deterministic Workflows and Adding AI 

If you have existing flows and want to add AI capabilities: 

1. Identify steps where AI adds value 
   • Which steps require interpretation, summarization, or reasoning? 
   • Where do fulfillers currently spend time on manual cognitive tasks? 
2. Choose the right AI capability 
   • For targeted tasks (summarization, classification): Use a Skill 
   • For complex reasoning with multiple tool calls: Use an AI Agent 
3. Configure the AI step appropriately 
   • Define expected outputs if subsequent steps need the data 
   • Set execution mode (autonomous vs. supervised) based on risk 
   • Specify support user for supervised agents 
4. Maintain governance 
   • AI operates within the workflow's governance structure 
   • Audit trails capture both deterministic steps and AI contributions 
   • Maintain human touchpoints at policy-defined locations 

Transitioning from AI-First to Hybrid 

If you have AI agents that would benefit from more structure: 

1. Identify predictable sequences 
   • Which steps does the agent perform consistently and in the same order? 
   • Where are rules being "re-derived" that could be enforced? 
2. Extract deterministic logic 
   • Build flows for the predictable portions 
   • Keep the agent for reasoning-intensive portions 
3. Use the agent as a step, not the orchestrator 
   • Workflow controls the sequence 
   • Agent provides intelligence at specific points 

Governance and Trust Considerations 

When deciding between patterns, consider the governance implications: 

Auditability 

Pattern	Audit Trail
AI-First	Agent conversation logs capture reasoning, but path varies
Deterministic with AI	Workflow execution history shows exact steps; AI contributions are logged within the workflow context

Explainability 

For regulated environments, consider who needs to understand decisions: 

• Auditors and regulators: May require step-by-step explainability that deterministic workflows provide 
• Technical teams: Can interpret agent reasoning logs 
• Business users: Benefit from predictable workflow structures 

The Trust Threshold 

As AI capabilities mature, the threshold for when to use AI-first patterns may shift. Currently, most enterprise use cases benefit from deterministic orchestration with targeted AI enhancement. Over time, as AI explainability and reliability improve, more processes may be candidates for AI-first approaches. 

Conclusion 

The introduction of hybrid workflow automation creates new possibilities for combining the best of deterministic and AI-driven approaches. The key takeaways are: 

1. Choose based on requirements, not technology preference: Governance, compliance, and predictability needs should drive pattern selection 
2. Deterministic workflows provide the "guardrails": AI operates inside the workflow, ensuring governance and auditability 
3. Existing investments can be enhanced: You do not need to rebuild proven flows; you can add AI capabilities where they provide value 

Resources 

Workflow Automation Center of Excellence 

Getting Started with Playbooks 
Jumpstart your Workflow Automation: Playbooks (Impact Offering) 
Platform Fundamentals Academy: Use Workflows as Tools for AI Agents 

Identifying the right use case(s) for Agentic workflows 

Design considerations - traditional vs agentic workflows 

Training 

Workflow Automation Training on ServiceNow University 
Now Assist AI Agents Deep Dive on ServiceNow University 
AI Builder Fundamentals on ServiceNow University 

Center of Excellence Navigation

1. Workflow Automation - Center of Excellence
2. Workflow Automation Migration Considerations
3. What's new for Workflow Automation?
4. Resource Hubs
   1. Flows
   2. Decision Tables
   3. Playbooks
5. FAQ
6. Checklist
7. Training
8. Workflow Academy