Use deterministic workflow automation when the steps and decisions can be expressed reliably in rules. Use an AI agent when the task requires interpretation, tool selection or adaptation across variable inputs—and only when that flexibility creates measurable value greater than its additional risk and cost.
Direct answer
Start with the least complex architecture that can meet the outcome:
- ordinary software or rules;
- a workflow containing one or more model-assisted steps;
- a single agent with bounded tools; and
- multiple agents only when separate roles, context or permissions materially improve the result.
OpenAI distinguishes agents from simple chatbots and classifiers by whether a model manages workflow execution and can select tools to complete a task. Its implementation guide recommends starting with clear instructions, tools and guardrails. See A practical guide to building AI agents.
Decision matrix
| Factor | Deterministic workflow | Single AI agent | Multi-agent system |
|---|---|---|---|
| Input variation | Low to moderate and classifiable | High language or context variation | High variation across distinct specialist domains |
| Decision path | Known in advance | Chosen within defined bounds | Delegated or routed among bounded roles |
| Correctness | Rule and test based | Statistical plus rule and outcome evaluation | Adds handoff and coordination evaluation |
| Permissions | Service or user permissions | Tool-specific agent permissions | Prefer role-specific identities and tools |
| Operating cost | Usually predictable | Model, retrieval and evaluation costs | Higher model, state, tracing and failure cost |
| Best fit | Stable repeatable process | Variable knowledge work with measurable completion | Work that genuinely needs role separation or parallel specialization |
When ordinary automation is the better answer
Prefer deterministic software when decisions depend on validated fields, regulatory rules, exact calculations, fixed sequences or high-volume repetition. Examples include schema validation, invoice-total calculation, entitlement enforcement and deployment policy checks.
Adding an agent to a stable path can reduce reliability while increasing latency and cost. A natural-language interface may still call the deterministic process without giving the model control of the process itself.
When a single agent is justified
A bounded agent can fit when the workflow includes incomplete language, variable documents, tool selection or exception handling that is difficult to enumerate. Good candidates have:
- a clear business goal and end state;
- accessible source evidence;
- tools with narrow contracts;
- a measurable baseline;
- low-risk or reversible early actions;
- an escalation path when context is missing.
Examples include triaging a service request, gathering evidence for a case, drafting a change plan or reconciling information across approved sources before a person decides.
When multiple agents add value
Use multiple roles when they need different instructions, context windows, identities, tool permissions or parallel work. An onboarding workflow might separate an HR-policy role, identity-provisioning role and task-coordination role. A supervisor can route work and retain the overall state.
Do not create multiple agents merely to represent an organization chart. Every handoff adds latency, cost and another failure boundary. First prove that a single agent or deterministic router cannot meet the requirement.
Calculate the business case
Measure the current process before selecting technology:
- completion time and human effort;
- rework and exception rate;
- error impact;
- throughput and demand variability;
- current software and labour cost;
- customer or employee outcome.
Then estimate model calls, retrieval, tool execution, observability, evaluation, support and human review. Include the cost of failures. A demonstration that completes one ideal scenario does not establish ROI.
Build a stop condition
Define when the agent must return control to a person or deterministic process. Common triggers include missing required data, conflicting sources, denied permission, excessive retries, high-impact action, low confidence in evidence or unavailable systems.
The stop condition is part of the architecture, not an error message added after testing.
Practical lab: compare both paths
The Enterprise AI Agent Integration Lab begins with one measurable workflow. Participants decide what remains deterministic, what the agent may interpret, which tools it may call and where human approval is required.
Frequently asked questions
Is every application that uses an LLM an agent?
No. A classifier, summarizer or chatbot may use a model without allowing it to control workflow execution or select tools.
Are multi-agent systems more capable than a single agent?
Sometimes, but not automatically. They can improve role separation and parallel specialization while also creating more handoffs, state, cost and evaluation work.
Should an agent replace an existing workflow platform?
Usually not by default. A workflow engine can retain deterministic state, retries, schedules and audit behavior while an agent handles selected interpretation or decision steps.
Explore Enterprise AI Agent Engineering or discuss a team use case.