# Data inventory for one AI workflow

Begin with source locations and owners. This worksheet does not authorize new access, copying, model training, or sending information to another provider.

- Workflow and intended outcome:
- Included case types and exclusions:
- Business owner:
- Technical owner:
- Person who can judge a correct result:
- Date / next review:

## 1. Identify the data's purpose

Record whether each source is needed to run a case, ground a decision, evaluate the result, or adapt a model. Training or fine-tuning is a separate design decision; it is not automatically required.

| Source / location                          | Purpose           | Authoritative owner | Required fields or content | Stable record identifier | Format / access method |
| ------------------------------------------ | ----------------- | ------------------- | -------------------------- | ------------------------ | ---------------------- |
| Incoming request or document               | Run a case        |                     |                            |                          |                        |
| Current business record                    | Ground a decision |                     |                            |                          |                        |
| Approved policy or reference               | Ground a decision |                     |                            |                          |                        |
| Representative cases and expected outcomes | Evaluate results  |                     |                            |                          |                        |
| Training examples, only if justified       | Adapt a model     |                     |                            |                          |                        |

## 2. Confirm access and handling

Complete with the relevant data and system owners. Record approval evidence separately from assumptions. Include restrictions on retrieved passages, derived records, evaluation copies, and logs where applicable.

| Source | Permitted users / services | Approved processing and storage locations | Approval evidence | Retention / deletion owner | Unresolved restriction |
| ------ | -------------------------- | ----------------------------------------- | ----------------- | -------------------------- | ---------------------- |
|        |                            |                                           |                   |                            |                        |

- How will the workflow preserve user or team access boundaries?
- What data can be minimized, masked, or omitted from tests?
- How will corrections and deletions reach downstream copies?
- Who must approve use with a different service or for a different purpose?

## 3. Check quality and updates

| Source | Missing / duplicated / ambiguous values | Which source wins a conflict? | Update frequency and mechanism | How stale or withdrawn content is handled | Owner / next action |
| ------ | --------------------------------------- | ----------------------------- | ------------------------------ | ----------------------------------------- | ------------------- |
|        |                                         |                               |                                |                                           |                     |

Do not assume a historical action was correct. Where records or reviewers disagree, resolve the policy or define escalation before using the example as a test answer.

## 4. Build a representative case collection

Define coverage for the intended first release before choosing a sample size. Include normal cases, consequential exceptions, and cases the system should decline or route to a person. Keep a separate set of cases for evaluation that are not repeatedly used to tune the implementation.

| Case reference                             | Source / format / team represented | Expected business outcome and reason | Who confirmed it? | Expected escalation? | Development or held-out evaluation? | Coverage gap |
| ------------------------------------------ | ---------------------------------- | ------------------------------------ | ----------------- | -------------------- | ----------------------------------- | ------------ |
| Normal case                                |                                    |                                      |                   |                      |                                     |              |
| Missing information                        |                                    |                                      |                   |                      |                                     |              |
| Ambiguous identity or conflicting sources  |                                    |                                      |                   |                      |                                     |              |
| Unreadable or unusual format               |                                    |                                      |                   |                      |                                     |              |
| Restricted record                          |                                    |                                      |                   |                      |                                     |              |
| Duplicate or repeated request              |                                    |                                      |                   |                      |                                     |              |
| Unsupported request / action outside scope |                                    |                                      |                   |                      |                                     |              |
| Workflow-specific exception                |                                    |                                      |                   |                      |                                     |              |

- Real operating conditions not represented:
- Synthetic examples used and their limitations:
- Reviewer disagreements and how they were resolved:
- Cases that must block release if they fail:

## 5. Make the next decision

| Gap | Effect on the selected workflow | Action: obtain access / repair data / clarify rule / narrow scope / stop | Owner | Due date | Evidence needed |
| --- | ------------------------------- | ------------------------------------------------------------------------ | ----- | -------- | --------------- |
|     |                                 |                                                                          |       |          |                 |

- Is an evaluation on approved examples possible now? Why?
- What remains unproven before production?
- Who accepts the limits of the first release?
- Next decision date:

Prepared by PlatformBridge. Companion guide: https://www.platformbridge.ai/guides/data-needed-for-ai