Operations
When processes need AI agents
A selection framework for choosing the least complex automation that can complete the job safely.
What agentic process automation means
Agentic process automation adds model-driven interpretation and decision-making to a business process while keeping the process layer responsible for state, approvals, policy, and auditability. It can complement workflow automation and RPA; it does not make those reliable mechanisms obsolete.
- The process still has an owner, trigger, case state, and outcome
- Agents handle bounded judgment or unstructured information
- Deterministic services execute stable calculations and rules
- Humans retain authority over exceptions and high-impact actions
RPA, workflows, copilots, and agents
RPA reproduces known actions across desktop or legacy interfaces and is well suited to repetitive rule-based work. Workflow automation coordinates defined steps and systems. A copilot proposes or summarizes while a person remains the actor. An agent can select actions and tools within a governed process.
- RPA: stable interfaces, repetitive actions, structured inputs
- Workflow: known sequence, branches, SLAs, and ownership
- Copilot: human initiates and approves most meaningful actions
- Agentic automation: variable path with verifiable, bounded decisions
Processes that fit agentic automation
The strongest candidates combine meaningful variation with a measurable outcome. Examples include interpreting documents, investigating exceptions, selecting among approved procedures, and preparing decisions for review. The process must expose reliable data and tools; otherwise the agent only automates uncertainty.
- Inputs include language, documents, or incomplete context
- The correct path varies but stays within known operational boundaries
- The result can be checked against records, rules, or human review
- A manual path already exists for unusual or high-risk cases
Keep the execution layer deterministic
Use agents to interpret, plan, or recommend; use APIs, workflow engines, and RPA to perform repeatable side effects. This hybrid keeps business records in their source systems, reduces prompt-dependent behavior, and lets teams replace a reasoning component without redesigning the entire process.
- Agent: classify an invoice exception and propose the next action
- Workflow: route the case, enforce SLA, and record ownership
- RPA or API: enter approved data into a legacy system
- Human: approve a payment, policy exception, or irreversible action
Approval is a process state
A reliable approval step pauses before the side effect, exposes the proposed action and evidence, accepts approve, edit, reject, or reassign input, and resumes from persisted state. A message saying the agent is unsure is not a human-in-the-loop design.
- Define which actions always require approval
- Show source records, assumptions, confidence, and expected impact
- Set a timeout, owner, delegation rule, and manual recovery path
- Record both the agent proposal and the human decision
A practical implementation sequence
Start with one process and one bounded decision. Document the current happy path and exception path, establish a baseline, keep the existing execution system, introduce the agent in shadow mode, and expand authority only after reviewed outcomes demonstrate value.
- Baseline volume, handling time, error rate, and escalation rate
- Run the agent without acting and compare recommendations to reality
- Allow low-risk proposals before allowing execution
- Roll back by routing all work to the existing deterministic process
Measure outcomes, not autonomy
Success is not the percentage of a process handled without people. Measure verified completion, error severity, manual rework, exception resolution, cost per successful case, latency, and policy violations. A workflow that escalates the right cases can be more valuable than one that maximizes straight-through processing.
- Verified completion and first-pass acceptance
- Manual intervention, rework, and recovery time
- Cost and latency per successful case
- Severe errors, unauthorized actions, and policy exceptions
Choose the automation mode
Map process predictability, exception frequency, error cost, data structure, and approval requirements to the right level of autonomy.
Use the framework above as a review sequencePrimary sources
Documentation reviewed for this guide
Claims about framework behavior are tied to current first-party documentation. The design recommendations are our synthesis of those documented capabilities.
- 01
- 02
- 03
- 04AnthropicBuilding effective agents ↗
- 05MicrosoftMicrosoft Agent Framework overview ↗