# AI project risk and data-handling checklist

Template v1 - 2026-10-18 - MiniAI Labs - https://miniailabs.in/resources/templates/
Free to use and adapt for your own projects. A working checklist, not legal advice: which laws apply depends on your data, your users and where they are. Have your legal or privacy advisor confirm section 2.

Fill it in before the build starts and update it before launch. Every No or blank is either a task to finish or a risk that someone accepts in writing in section 8.

Project: ____________________   Owner: ____________________   Last updated: ____________

## 1. Data inventory

| Data | Comes from | Personal data? | Sensitive? (health, financial, children, biometric, government ID) | Needed for the task? | Processed where (system, provider, region) | Kept for |
|---|---|---|---|---|---|---|
| | | Yes / No | Yes / No | Yes / No | | |
| | | | | | | |

- [ ] Every row marked Needed: No is removed before it reaches the model, the logs or the test set.

## 2. Rules that apply

- Where your users are: ____________________
- Laws your advisor says apply (for example India's DPDP Act, the EU or UK GDPR, US state privacy laws, HIPAA for US health data): ____________________
- Contracts that restrict this data (client agreements, data processing agreements): ____________________
- Confirmed by ____________________ on ____________

## 3. Model and service providers

For each provider that receives your data, record what its current terms say and the date you read them.

| Provider | What it receives | Used for training? | Retention | Region | Terms or setting checked (link and date) |
|---|---|---|---|---|---|
| | | | | | |

- [ ] Provider accounts are in your organisation's name.
- [ ] Settings that turn off training or extended retention are on where the provider offers them, and the setting is recorded above.

## 4. Access and secrets

- [ ] Each person and service has the least access that works.
- [ ] Keys and tokens live in a secrets store, never in code, prompts or documents.
- [ ] Access for the build team ends on a set date: ____________

## 5. Logs and records

- [ ] The logs kept for debugging are listed, with what each contains and how long it is kept.
- [ ] Personal data is masked in logs wherever it is not needed.
- [ ] Someone can find and delete one person's data when they ask.

## 6. Risks specific to AI systems

Based on the OWASP Top 10 for LLM Applications 2025 (https://genai.owasp.org/llm-top-10/, checked 2026-10-18). Fill in a control and an owner for every row that applies.

| Risk | Ask yourself | Your control | Owner |
|---|---|---|---|
| Prompt injection (LLM01) | Could text the system reads, such as a web page, an email or a document, contain instructions it might follow? | | |
| Sensitive information disclosure (LLM02) | Could an answer reveal data this user should not see? | | |
| Improper output handling (LLM05) | Is model output checked before another system uses it? | | |
| Excessive agency (LLM06) | Does the system have more permissions or actions than the task needs? | | |
| System prompt leakage (LLM07) | Would it matter if a user read the instructions? Keep secrets out of them. | | |
| Misinformation (LLM09) | What happens when an answer is confidently wrong, and who would notice? | | |
| Unbounded consumption (LLM10) | Is there a cap on usage and spend, per user and per day? | | |

## 7. People in the loop

- [ ] Actions that change records, move money or contact customers have an approval step or a limit: ____________________
- [ ] Users are told they are dealing with an AI system wherever that matters to them.
- [ ] There is a way to reach a person, and someone staffs it.

## 8. Before launch

- [ ] Test set agreed and passed. Agreed by: ____________________
- [ ] The system can be switched off quickly. People allowed to do it: ____________________
- [ ] Incident contacts on both sides: ____________________
- [ ] Accepted risks, signed by the owner:

| Risk accepted | Why | Review date | Signed |
|---|---|---|---|
| | | | |

## 9. System card (one page, kept with the system)

| Field | Entry |
|---|---|
| System | |
| Purpose | |
| Users | |
| Data in | |
| Data it must never receive | |
| Actions allowed | |
| Actions that need approval | |
| Hand-off to a person | |
| How users are told it is AI | |
| Providers, and whose accounts | |
| Evaluation (test set, pass rule, last run) | |
| Owner | |
| Next review | |

Framework reference: the NIST AI RMF Core functions Govern, Map, Measure and Manage (https://airc.nist.gov/airmf-resources/airmf/5-sec-core/, checked 2026-10-18). Using this checklist is not a certification or an audit.