"Enterprise-grade security" is not an answer. A URL with a last-updated date is. If you build products on third-party models, your privacy job is less about policing coworkers' browser tabs and more about reading vendor retention pages like contracts.
Disclosure: This piece is from the ToolPotion Media Team. ToolPotion is our product. Retention examples and breach stats below come from our AI tool data privacy framework.
The risk moved to the vendor side
Shadow AI is the default setting in a lot of companies. Reporting we cite in the framework notes 45% of employees as regular AI users on corporate devices (up from 15% a year earlier), with 67% of them working through non-corporate accounts (privacy framework). Shadow AI also climbed sharply as a non-malicious insider action in DLP datasets.
A policy that only says "do not use unapproved tools" fails before lunch. The workable move is triage at the vendor level: publish a short list of tools that clear a documented bar, then make those the easy path.
Breach economics make the same point. High levels of shadow AI added about USD 670,000 to average breach cost in the study we reference, against a global average around USD 4.44 million. Among organizations with AI-related incidents in that research, 97% lacked proper AI access controls, and 63% of organizations studied had no AI governance policy (privacy framework).
Five questions with the right shape of answer
Every question below wants an artifact you can save and re-check next quarter. Adjectives fail.
Does our data train the model, on which tier?
Example shape: OpenAI documents that API data is not used for training by default since 1 March 2023, separately from abuse-monitoring retention.How long are prompts and outputs retained?
Example shape: Anthropic deletes API inputs within 30 days in the policy language we quote, while keeping Usage Policy violation content longer (up to 2 years) and classifier scores for 7 years. Those are different clocks.Where is data processed and stored?
Look for region pins and whether "EU residency" is a real carve-out or marketing.What certifications are current, and for which services?
SOC 2 is not a synonym for "will not train on my repo." Read the scope.Who can compel access, and how are subprocessors listed?
You want a dated subprocessor list and a clear government-request path.
If a salesperson answers with a slide instead of a link, the vendor failed the question for now.
A four-tier trust model for busy teams
You do not have time for a 40-question questionnaire across every listing in a large directory. Use tiers:
- Tier A: dated public docs cover training opt-out, retention in days, and region options. OK for sensitive code with contract review.
- Tier B: solid on training and retention; weak on residency. OK for internal non-regulated work.
- Tier C: consumer product; assume prompts may be used to improve models unless proven otherwise.
- Tier D: unknown or contradictory docs. Block for work data.
Revisit tiers quarterly. Policies move.
What developers can do this sprint
- Put approved tools in the README and in MCP / assistant configs so the secure path is the default. See our MCP explainer.
- Redact secrets before prompts; add pre-commit checks for common key patterns.
- Prefer API tiers with explicit no-train defaults for customer data.
- For the highest sensitivity, keep inference local or VPC-hosted. Our local vs cloud guide covers when that pays off.
- Log model calls like you log production data egress.
Privacy theater to skip
- Collecting SOC 2 badges without reading scope.
- Banning ChatGPT while ignoring IDE plugins with the same data path.
- Trusting "we never store your data" when the abuse-monitoring clause stores it for 30 days.
- One-time legal review with no calendar reminder.
Practical takeaways
- Demand dated policy URLs, not adjectives.
- Separate training opt-out from retention.
- Tier vendors so engineers do not freestyle.
- Make the approved path the easy path in tooling.
Full checklist and vendor examples: Which AI Tools Can You Trust With Your Data?.
Concrete developer checklist (print this)
Before a new model vendor touches customer data:
- Save PDFs/HTML of training, retention, and subprocessor pages with today's date.
- Confirm which product tier you are on (consumer chat vs API vs enterprise).
- Disable training where the control exists; screenshot the toggle or contract clause.
- Set retention to the minimum that still supports abuse review if you must keep any.
- Route traffic through a gateway that redacts tokens and records model IDs.
- Add the vendor to your internal tier list with a review date 90 days out.
- Tell engineers in Slack/README which lanes are approved — silence creates shadow AI.
For MCP and plugin connectors, repeat the same review per server, not once per host app. A safe chat vendor plus an unsafe connector is still an unsafe path.
Pairing privacy with architecture
Privacy choices shape architecture. If Tier A vendors are expensive, invest in redaction and field-level filtering so more traffic can stay on Tier B. If legal requires residency, do not start a prototype on a global-only endpoint and "fix it later." Later becomes a migration with prompt logs you cannot delete. Bake the constraint into the first RFC.
Example: reading two clocks on one vendor
OpenAI's public API materials (as we quote in the framework) separate "API data not used for training by default since 1 March 2023" from abuse-monitoring logs that may hold prompts for about 30 days. Anthropic's materials separately describe deleting API inputs within 30 days while retaining certain policy-violation content much longer. If your threat model only asks "do you train on us?" you miss the retention clock. If it only asks "how long do you store?" you miss training. Ask both. Store both answers. Re-check when the vendor emails a policy update instead of assuming silence means stability.
ToolPotion Media Team


Top comments (0)