# AI Tool Privacy and Procurement: A Practical Review

> Check data flows, account types, model providers, retention, contracts and exit plans before approving AI tools.

Source: https://mentionx.ai/learn/ai-tools/ai-tools-privacy-procurement  
Author: Dipuck Jones, founder of MentionX  
Updated: 8 November 2026  
Level: Practitioner

**Approve AI tools for a defined data flow, account type and use case, not for a brand name in general. Check what is uploaded, which models and subprocessors receive it, how interactions are stored, who can access outputs and how the organisation exits. "Not used for training" is only one question and does not mean zero retention or automatic legal compliance.**

> **Key takeaways**
>
> - Map data classification, processing, storage, sharing and deletion.
> - Distinguish consumer, commercial, core-service and additional-service terms.
> - Review model and connector choices, not just the main vendor.
> - Keep approval scoped, dated and owned.

*Checked 8 November 2026 (CET) against Microsoft commercial Copilot privacy, Google work-account Gemini Notebook, ElevenLabs policy and Adobe Firefly documentation. Procurement methods are editorial recommendations, not legal advice. Ledgerline is fictional; all review records and decisions are illustrative.*

## Why is vendor-level approval too broad?

A vendor may offer consumer, business, enterprise and API products with different terms and controls. A team can use the same familiar interface while operating under a different account or feature configuration. Approval should identify the actual service and use, not merely the company name.

Describe the workflow first. Public campaign copy, internal product plans, interview recordings and customer financial records have different sensitivities. A tool suitable for one can be unsuitable for another without either conclusion being contradictory.

Map the processing route. Identify the application, selected model, connectors, storage, logs, review tools and output destinations. Optional web search, agents or partner models can introduce different terms or data handling from the core experience.

Keep procurement and operational review connected. A contract can permit a use that is still poorly configured, while a careful operator cannot override an unsuitable agreement. The approval needs both appropriate terms and a workable control design.

## What data inventory should marketing prepare?

List the minimum fields or files needed for the task. Classify public, internal, confidential and restricted information using the organisation's own scheme. Remove data that contributes no necessary value rather than sending a full export by default.

Identify people and rights affected. Customer notes, employee recordings and licensed reports carry interests beyond the marketing team's convenience. Check whether the intended processing and later use fit the permission or agreement actually in place.

| Inventory item | Review question |
|---|---|
| Source data | What exact information enters the workflow? |
| Purpose | Why is each field necessary? |
| Account | Which organisation and service terms apply? |
| Processing | Which models, regions and connectors are involved? |
| Storage | Where are prompts, files, outputs and logs retained? |
| Sharing | Who can access or redistribute the material? |
| Exit | How are access and retained copies handled? |

Use synthetic or minimised data during initial evaluation where possible. A tool does not need a real customer invoice to prove that a field mapping works. Test mechanics without creating avoidable processing obligations or exposure.

## What does "not used for training" leave unanswered?

Microsoft's checked commercial Copilot documentation states that prompts, responses and Microsoft Graph data are not used to train foundation large language models in that scope. It also explains that interactions are stored and subject to organisational retention and governance controls.

Those statements are compatible. Training use and storage are different questions. Ask about retention, deletion, administrator access, audit, incident handling and the effects of connected services. Do not translate a no-training statement into "nothing is saved" or "no person can ever review anything".

The same documentation explains that Copilot respects existing access permissions. That does not repair overshared files. If an employee can already access an overly broad folder, AI can make the information easier to discover. Review source access before connecting the workflow.

Agents and web search can have separate terms or processing routes. Check the actual feature rather than extending the core commercial assurance to every optional service, third-party tool or consumer product.

## How do account editions change Google research-tool handling?

Google's current work-account Gemini Notebook guidance distinguishes service and licence categories. Some accounts receive the product as a core service with enterprise-grade protections, while other editions use an additional-service arrangement. Feature access and applicable terms vary.

The checked guidance describes different data handling for these categories, including feedback-related review conditions in the standard arrangement and stronger protections for eligible core-service use. Do not tell every work-email user that they automatically have the same enterprise treatment.

Ask the administrator to confirm the actual edition and service status. A badge showing higher usage limits is useful for access checks, but it is not a complete procurement record. Keep the applicable agreement and administrative configuration with the approval.

Review sharing separately. A research notebook can contain multiple sources with different sensitivities. Permitting the tool does not mean every notebook should be broadly shared or that every output may be published.

## What regional and model-provider questions matter?

Ask where processing occurs and whether optional models change the commitment. Microsoft's checked documentation currently identifies an EU Data Boundary exception for Anthropic subprocessor models. That is a reason to inspect the selected configuration, not to make a blanket claim that every Copilot interaction stays within the EU.

Do not infer residency from the user's location, a currency setting or a vendor's headquarters. Read the applicable regional documentation and contract, and involve the appropriate privacy or legal owner when cross-border or regulated processing is consequential.

Check sensitive-data restrictions directly. ElevenLabs' current policy places specific conditions on sensitive bank-account, identification and health information, including prior written agreement in the described circumstances. A marketing operator should not assume that ordinary product access authorises those uploads.

Adobe Firefly's FAQ also distinguishes Adobe's own model statements from the broader product environment. Keep partner-model choices within the review. A familiar application can expose different processing and rights questions when the selected model changes.

## What should a practical approval record contain?

Record the use case, permitted data classes, account type, model or feature scope, approved integrations, retention requirements, access controls, human review and owner. Add the date checked, supporting documents and unresolved conditions. Approval should be understandable to an operator, not only to procurement.

Define prohibited uses explicitly. For example, an image-concept workflow might permit approved public brand assets while excluding customer records and identity cloning. The boundary prevents a general tool approval from being stretched into unrelated sensitive work.

Set change triggers. New models, connectors, data classes, destinations, subprocessors or regional requirements may require review. A renewal is a useful checkpoint, but waiting for the annual renewal can be too slow when the workflow changes materially.

Plan exit before adoption. Identify how approved outputs, source records and useful templates can be retained appropriately, how access is revoked and what deletion or retention evidence is available. Do not promise export or deletion capabilities without checking the actual service and contract.

## How should you review logs and retained copies?

Ask where copies appear beyond the main application. A workflow may retain an uploaded file, a model interaction, an automation execution record, a review attachment and a destination asset. Deleting one visible chat does not establish that every relevant copy has been removed.

Create a processor-and-storage table with the purpose, owner, retention rule and deletion or access-revocation route for each location. Record what is confirmed in documentation or contract and what still requires vendor clarification. Do not fill an unknown retention period with an assumption based on another product from the same company.

Check whether support, debugging or feedback workflows introduce additional handling. Operators should know which material may be submitted and which channels are prohibited for sensitive content. A troubleshooting request should not become an unreviewed transfer of the original customer dataset.

Finally, test the practical exit process with non-sensitive material. Confirm that the organisation can retain the records it needs, remove unnecessary access and identify unresolved deletion questions. The goal is an evidence-based exit plan, not an unsupported promise that every copy can disappear instantly.

> **Pro move:** Put a short permitted-use label where operators start the task: approved account, allowed data, forbidden data and owner. A long procurement document helps little if the person uploading a file cannot find the boundary.

## Worked example: Ledgerline

Ledgerline wants an illustrative approved stack for creative concepts, internal research and narration. It does not issue a blanket approval for all marketing data. The following record is illustrative and does not assert that these vendors are suitable for every organisation.

| Use case | Illustrative allowed inputs | Condition before use |
|---|---|---|
| Firefly concept images | Approved public brand assets | Actual model and relevant terms checked |
| Gemini Notebook research | Permitted internal documents | Account edition, service status and sharing confirmed |
| Commercial Copilot analysis | Minimised approved workbook | Source permissions and feature scope reviewed |
| ElevenLabs narration | Approved script and authorised voice | Commercial-use eligibility and identity rules satisfied |

Customer bank details are excluded from these ordinary marketing workflows. If a proposed task genuinely needs restricted information, it goes through a separate specialist review rather than being squeezed into an existing approval.

### Walkthrough

1. Marketing lists the exact inputs and removes fields unnecessary for each task.
2. IT confirms the actual organisational accounts, licences, administrators and connected features.
3. Privacy and procurement review the applicable terms, processing routes, retention and regional conditions.
4. The workflow owner records allowed and prohibited uses in a short operator-facing label.
5. The team tests with synthetic or minimised data and checks sharing, logs and output handling.
6. Approval is dated and scoped, with review triggers for new models, connectors, data classes or destinations.

The record makes uncertainty visible. An unresolved contract or regional question remains a hold, not a footnote hidden behind a general statement that the vendor is enterprise-ready.

## Common mistakes

- **Approving a vendor instead of a use.** Scope the actual service, account and workflow.
- **Equating no training with no storage.** Review the full data lifecycle.
- **Assuming a work email guarantees enterprise terms.** Confirm edition and service status.
- **Ignoring inherited access.** AI does not fix overshared source folders.
- **Making blanket residency claims.** Optional models and features can change the route.
- **Forgetting exit and change review.** Adoption creates an ongoing operational responsibility.

## Try it yourself

Select one proposed AI workflow and prepare a data-flow sheet. List inputs, sensitivity, account type, model, connectors, storage, logs, output destinations and owner. Remove unnecessary fields and identify which claims still need official or contractual confirmation.

Write a permitted-use label and a prohibited-use list. Add review triggers for a new model, new connector and more sensitive data. Ask a colleague to decide whether a sample upload is allowed using only that record.

Deliver the review sheet, unresolved questions and exit checklist. For jurisdiction-specific issues, use qualified advice and the existing [EU AI Act lesson](https://mentionx.ai/learn/governance/eu-ai-act-applies) as related education, not as a substitute for the applicable legal review.

## Quick quiz

**1. Does a no-training commitment imply zero retention?**

*Answer:* No. Training, storage, access and retention are separate questions.

**2. Why check the selected model and optional features?**

*Answer:* They can introduce different providers, terms, processing routes or regional conditions.

**3. What should trigger a fresh review?**

*Answer:* Material changes such as new data classes, models, connectors, destinations or applicable requirements.

## What should you check before moving on?

Confirm approval names the actual account, data flow, permissions and unresolved conditions. Continue with [rights, disclosure and editorial review](https://mentionx.ai/learn/ai-tools/ai-tools-copyright-disclosure-review), which addresses different questions from privacy and procurement.

Previous: [AI Agents for Marketing: Permissions, Approvals and Safety](https://mentionx.ai/learn/ai-tools/ai-tools-agent-approvals-security) · Next: [AI Content Rights, Disclosure and Human Editorial Review](https://mentionx.ai/learn/ai-tools/ai-tools-copyright-disclosure-review)

Or return to the [AI Tools for Marketers track](https://mentionx.ai/learn/ai-tools/ai-tools-marketing-for-marketers).

Related: [Evaluate AI Marketing Tools With a Task-Based Scorecard](https://mentionx.ai/learn/ai-tools/ai-tools-selection-evaluation) · [AI Integrations: Reliable Data Contracts and Automation](https://mentionx.ai/learn/ai-tools/ai-tools-integrations-automation-contracts) · [AI Content Rights, Disclosure and Human Editorial Review](https://mentionx.ai/learn/ai-tools/ai-tools-copyright-disclosure-review)

## Common questions

**Does no training mean no retention?** No. Training, storage, administrator access, retention and deletion are separate questions.

**Does a work email guarantee identical enterprise terms?** No. Confirm the actual account edition, service status and applicable agreement.

**What should trigger procurement review?** Material changes to data, models, connectors, destinations, processors or applicable requirements.

## Sources

- [Data, Privacy, and Security for Microsoft Copilot](https://learn.microsoft.com/en-us/copilot/plan/microsoft-365-copilot-privacy)
- [Use Gemini Notebook with a work or school Google account](https://support.google.com/notebooklm/answer/16337734)
- [ElevenLabs Prohibited Use Policy](https://elevenlabs.io/use-policy)
- [Adobe Firefly FAQ](https://helpx.adobe.com/firefly/web/get-started/learn-the-basics/adobe-firefly-faq.html)
- All accessed 8 Nov 2026 (CET).

## Frequently asked questions

### Does no training mean no retention?

No. Training, storage, administrator access, retention and deletion are separate questions.

### Does a work email guarantee identical enterprise terms?

No. Confirm the actual account edition, service status and applicable agreement.

### What should trigger procurement review?

Material changes to data, models, connectors, destinations, processors or applicable requirements.

---

Written and reviewed by Dipuck Jones for MentionX Learn. The HTML version of this page: https://mentionx.ai/learn/ai-tools/ai-tools-privacy-procurement