# AI workflow pilot — MarketplaceBeta

A planning worksheet, not a forecast of savings. Start with one repeatable task.

## 1. Define the job
- Workflow / business decision:
- Process owner and backup:
- Frequency and number of cases:
- Current steps, tools and total minutes per case:
- Current error/rework rate and how measured:
- Desired improvement:

## 2. Prepare the evidence
- Approved input sources and permissions:
- Fields required / missing fields:
- Customer or confidential data to remove:
- Who may access the data and where it may be processed:
- Retention / deletion requirements:
- Calculation formulas (run outside the language model):

## 3. Design the pilot
- Representative sample and comparison baseline:
- AI task: draft / classify / summarize / explain:
- Output schema and required source identifiers:
- Named human reviewer:
- Actions that always require approval:
- Missing-data and conflicting-source behavior:
- Stop conditions / rollback owner:

## 4. Evaluate each case
| Case ID | Baseline minutes | AI preparation + review minutes | Factual corrections | Accepted? | Failure / next check |
|---|---:|---:|---:|---|---|
| | | | | | |

## 5. Review the economics
- Total baseline effort versus total pilot effort:
- Include preparation, review, corrections and maintenance:
- Tool/API costs, integration costs and training effort:
- Quality, exceptions missed and customer impact:
- Estimated capacity released is not automatically cash saved or revenue earned:
- Decision: stop / revise / continue with a larger controlled sample:
- Owner, review date and success threshold:

Sources, playbooks and inquiry: https://marketplacebeta.com/ai