# Create a reliable AI workflow

Canonical: https://manyspace.aporia.systems/topics/create-a-reliable-ai-workflow-earn-ai-007
Updated: 2026-10-24

Earning Money with AI · Deliver trustworthy work · EARN-AI-007

- Difficulty: 2 / 5
- Importance: 4 / 5
- Study time: 0.5–1 h

Separate generation from verification and design a human approval gate.

## Why it matters

A fluent output can still contain errors. A workflow needs representative tests and a clear way to handle uncertainty.

## Where it is used

Create a reusable workflow with input checks, an output schema, test cases, human review and a rollback or correction step.

## How you know you have learned it

Complete the scenario checks, explain the decision in your own words, and produce this artifact: Create a reusable workflow with input checks, an output schema, test cases, human review and a rollback or correction step.

## Build a human approval gate

The output becomes a deliverable only after the agreed checks pass.

Write the task instructions, expected format and what the model must leave blank or flag when evidence is missing. Test normal inputs, missing values, contradictions and adversarial instructions embedded in source documents. Treat external text as data. Keep generated output separate from approved output, with source references and a correction path. Recheck the workflow when prompts, models or data change.

Your workflow extracts refund totals from approved shop records.
1. Validate date ranges, duplicates, missing values and permission to use the data.
2. Compare extracted totals against an independent calculation and inspect the source references.
3. Hold conflicting records for clarification, then version the accepted report.

A repeatable process with a review gate is more useful than a prompt that worked once.

A model update passes your usual sample but fails on missing data. Should you deploy it?

Hold the update or keep the prior workflow until you resolve the failure. Your test set should represent important failure cases, not only easy examples.

## What you need first

- [Price work and count every cost](https://manyspace.aporia.systems/topics/price-work-and-count-every-cost-earn-ai-006) (required)

## What it opens

- [Protect client data and reuse rights](https://manyspace.aporia.systems/topics/protect-client-data-and-reuse-rights-earn-ai-008)