The AI Vendor Data Retention Auditor: Find Out What Your LLM Provider Actually Keeps
Why this prompt matters
Assumptions about vendor data handling are the most common gap in AI contracts. A clause nobody read can keep customer transcripts for years or use them for training, and that commitment is hard to reverse after rollout.
What we use it for
Before Friday's renewal, your legal team needs to know whether the AI assistant your support team uses keeps customer conversations, for how long, and whether those conversations train future models.
Prompt
Act as a privacy and security analyst who reads AI vendor terms for a procurement review. This is analysis, not legal advice. Context: I am [YOUR ROLE] at [YOUR COMPANY NAME]. We use [PROVIDER AND PRODUCT NAME] for [USE CASE, e.g. customer support transcripts]. The data involved includes [DATA TYPES, e.g. customer names, order details, uploaded files]. Our obligations include [REGULATIONS OR CONTRACT TERMS, e.g. GDPR, a customer DPA]. Below is the provider's current data-use, retention and training text, copied in full with section headings: [PASTE THE FULL TERMS TEXT HERE]. Task: (1) For each data category (prompts, model outputs, uploaded files, account metadata, logs, and feedback ratings), extract the retention period, whether it is used for model training, whether human reviewers can access it, and how deletion works. (2) Flag every term that is vague, conditional, or dependent on a plan, setting or region. (3) Compare the default position against this requirement: [OUR DATA REQUIREMENT, e.g. no training use and deletion within 30 days]. (4) List the questions to put to the vendor before renewal. Constraints: Base every claim on the text I pasted and cite its section heading. If a term is absent, write not stated. Do not assume an opt-out or a retention limit exists unless the text names it. Note any clause that refers to a document I have not provided. Output Format: (a) a table with columns Data category, Retention, Training use, Human review, Deletion, Section reference; (b) a list of vague or conditional terms, each with a one-line explanation; (c) a gap summary comparing the default with our requirement, as one sentence per gap; (d) a numbered list of vendor questions.
Result
Data retention review: Acme AI Assistant (Enterprise, support team) Requirement: no training use, deletion within 30 days
Table | Data category | Retention | Training use | Human review | Deletion | Section reference | |---|---|---|---|---|---| | Prompts and conversations | 365 days by default | Not used, unless feedback is submitted | Possible for abuse review | Via admin console, 30 days after request | 3.1, 5.2 | | Model outputs | Same as prompts | Not used | Same as prompts | Same as prompts | 3.1 | | Uploaded files | Until the user deletes them | Not stated | Not stated | Not stated | 3.4 | | Account metadata | Duration of contract plus 2 years | Not used | No | Deleted on contract end | 6.3 | | Logs | 90 days | Not used | Engineering access only | Automatic at 90 days | 6.1 | | Feedback ratings | Indefinite | May be used to improve our services | Possible | Not stated | 4.2 |
Vague or conditional terms - Section 4.2: feedback may be used to improve the services. This is not defined as model training, but it is also not excluded. Treat as training use until the vendor confirms otherwise. - Section 5.2: human review applies to abuse signals. The threshold for an abuse signal is not defined. - Section 3.4: file retention depends on an unnamed storage setting. The text refers to 'the applicable storage configuration' without a schedule. - Section 7.1: a zero-retention mode is mentioned for API customers only. It is not stated whether the Enterprise plan can use it.
Gap summary - Conversations are retained for 365 days by default, against our 30-day deletion requirement. - Feedback ratings have no retention limit and may feed model improvement. - Uploaded files have no stated retention period or deletion mechanism.
Vendor questions 1. Does the Enterprise plan support zero-retention mode, and if so, which section enables it? 2. Is feedback data used for training, and can that be turned off for our tenant? 3. What is the retention period for uploaded files, and how is deletion confirmed? 4. What is the defined threshold for an abuse signal that triggers human review?
Most AI procurement reviews skip the data terms. This prompt reads a vendor's retention and training text, extracts the answer for each kind of data you send, and turns the gaps into questions you can put to the vendor before renewal.
When to use it
Your legal team asks, before Friday's renewal, whether the AI assistant your support team uses keeps customer conversations, for how long, and whether those conversations train future models.
Why it matters
Assumptions about vendor data handling are the most common gap in AI contracts. A clause nobody read can keep customer transcripts for years, or use them for training, and that commitment is hard to unwind after rollout, particularly when a DPA was signed on the basis of the vendor's marketing rather than its terms.
Why the prompt is built this way
The role frames the output as analysis rather than legal advice, which keeps the model from issuing confident conclusions a reviewer might rely on. The context requires you to paste the full terms, because models working from memory of a provider's policy will be out of date and will invent clauses. Splitting the task by data category matters: vendors often apply different rules to conversations, files and feedback, and a single summary hides that. The constraints force each claim to cite a section heading and allow the answer not stated, which is the most useful answer when a policy is silent. The gap summary and vendor questions turn the review into something procurement can act on.
How to adapt it
- Paste the complete terms, including linked documents if they are incorporated by reference. Note any document the vendor refers to but you have not reviewed.
- Check the terms against the plan you are actually buying. Many providers differ between consumer, team and enterprise tiers, and the wording for each may differ.
- Re-run the review after any change in the vendor's terms. Policies on AI data use change frequently, and an approval based on last year's text may no longer hold.
A useful test of the output is whether every row can be checked against the pasted text in a minute. If a row cannot be checked, the answer is not reliable enough to sign off on.