AI Operators Need SOPs, Not Prompts
An AI workflow SOP is a one-page operating record for one bounded job: inputs, permissions, steps, acceptance evidence, stop conditions, human approval, owner, version, and review date. Copy the template below before you tune another prompt.
AI
5 min
An AI workflow SOP is a one-page operating record for one bounded job. It defines what enters the workflow, what the system may do, what acceptable output looks like, when it must stop, who approves consequential actions, and when the record is reviewed. The prompt belongs inside that system; it is not the system.
Author synthesis: this is a reusable blank template, not a report of a customer deployment, benchmark, or measured outcome.
Copy this one-page AI workflow SOP
Workflow name, owner, version, and review date. Name one accountable operator and the record that replaces this version.
Decision and boundary. State the job, the user or business decision it supports, what stays out of scope, and why a model is needed instead of fixed rules.
Trigger, inputs, and prohibited data. Define when the run starts, approved source material, required context, retention limits, and data the workflow must never receive.
Steps, tools, and permissions. List the smallest sequence, the allowed read and write tools, account scope, and actions that remain unavailable.
Acceptance test and evidence. Define a passing example, a failing example, deterministic checks first, the evidence saved for review, and any uncertainty that remains.
Stop conditions, retry ceiling, and escalation. Name the failures that stop the run, the maximum retries, the rollback or safe state, and the person who receives control.
Human approval. Mark the exact checkpoint before an irreversible, external, high-risk, financial, legal, privacy-sensitive, or customer-facing action.
Change log and recheck. Record the model or workflow configuration, reviewer, material changes, known failures, review cadence, and next review date.
Run the pre-live review
Boundary: use the simplest reliable pattern. If fixed rules can complete the job, keep the model out of the control path.
Permission: verify the exact data, tools, account scope, write access, and human approval point before a live run.
Evidence: run known-good and deliberately failing cases. Save the checks, outputs, reviewer judgment, and limitations instead of only the score.
Control: confirm the stop condition, retry ceiling, safe state, escalation owner, and rollback before the first production run.
Maintenance: schedule review when the model, tools, data, permissions, failure modes, or operating context changes.
Source notes
OpenAI’s practical guide to building agents supports explicit instructions, tools, guardrails, failure handoff, and risk-based human intervention.
Anthropic’s Building Effective Agents supports starting with the simplest pattern, using environment evidence, human checkpoints, and explicit stopping conditions.
NIST AI RMF Core supports documented roles, periodic review, traceable measurement, monitoring, and accountable risk decisions. NIST does not prescribe this page’s template.
The Measurement Layer supplies the first-party rule to choose acceptance criteria before a run and retain failed cases and configuration beside the result.
Where this fits
Use the AI agent evaluation metrics that matter to choose the operating scoreboard, and the AI agent incident review template when a failure escapes. Continue with the AI agents for operators hub.
