Nanny Agency JobsBefore an agency shares candidate or household information with an AI vendor, it needs a concrete account of where that information will go, what the service will do with it and what happens afterward. A reassuring sales answer is a starting point for questions, not a substitute for the relevant documents and a configuration you can inspect.
Use this guide to prepare operational diligence and identify unresolved issues for the appropriate adviser. It does not interpret a contract, determine your legal obligations or certify a vendor as safe. Begin with fictional test records while real-data questions remain open.
Describe the exact data journey you are considering
Name the task, the staff who will use the service and the information it would receive. “Our candidate database” is too broad. Separate résumé text, recruiter notes, reference contact details, identification documents and family information. Ask which categories are actually necessary for the proposed task.
Then map the route: where the information starts, how it enters the tool, what output returns and where that output is stored. Include attachments, integrations and any manual copying. A browser-based assistant and an embedded database feature may involve different routes even if they use a similarly named model.
The UK ICO's recruitment procurement questions address processing responsibilities, transparency and limits on personal information. That is UK regulatory guidance; the applicable requirements for your agency need separate assessment. ICO: AI recruitment data protection considerations.
Ask who receives information and for which purposes
Ask the vendor to identify the service operator and any other providers involved in processing the submitted material. Request the documentation applicable to your actual product and account type. A general website statement may describe a different service from the one being offered.
Useful questions include whether submitted information is used to train or improve models, whether that use can be controlled and whether feedback buttons change the treatment of a conversation. Ask separately about inputs, outputs, uploaded files, logs and support interactions. An answer about one category should not silently stand for all of them.
Request a clear explanation of who can access information for support or operations and under what conditions. If the vendor says data is not used for training, continue asking about storage, human access and onward processing. Those are different questions. Record any answer the vendor cannot yet substantiate.
Examine retention, correction and removal separately
Ask how long each relevant data category is retained and what controls your agency has. Distinguish deleting a visible conversation from removing uploaded files, derived indexes, operational logs or backup copies. Ask for the documented behavior and any exceptions rather than assuming a delete button resolves every copy.
Consider corrections as well as deletion. If a candidate updates availability, how will the change reach a searchable copy or previously generated summary? Who is responsible for identifying outdated outputs already exported to another system? The vendor may control only part of that journey.
Ask what happens when a staff account is closed, an integration is disconnected or the agency ends its subscription. Include the time and format available for exporting necessary work. Do not discover at cancellation that essential review records are available only inside a departing employee's account.
Turn broad assurances into inspectable evidence
Use this original question register to avoid accepting incomplete answers:
- Question and the exact product or feature it concerns.
- Vendor answer, date and named contact.
- Supporting document, version and relevant section.
- Configuration or demonstration still needed.
- Limitation, exception or unanswered point.
- Agency owner and the decision waiting on the answer.
Ask how permissions are enforced, how access can be reviewed and how a suspected incident is reported. Request the relevant security documentation and have someone competent assess its scope. The existence of a report or badge does not establish that it covers your particular feature and proposed use.
For contract questions, collect the actual terms, data-processing documents and proposed amendments. Bring unclear responsibilities or consequential gaps to the agency's qualified adviser. Do not ask a general AI assistant to declare the contract compliant or replace that assessment with an automatically generated score.
Test a fictional vendor answer for missing detail
Fictional example: A vendor tells an agency, “Your information is private and is not used for training.” The agency is considering a feature that searches uploaded candidate résumés. It records the answer but asks what happens to the uploaded files and the search index when a résumé is corrected or removed.
The vendor supplies retention documentation for chat messages but not for the search feature. The agency marks the search-specific question unresolved. It continues the demonstration with invented records and does not treat the chat policy as an answer about candidate-file storage.
A later demonstration shows a deleted fictional document disappearing from normal results. That is useful observed behavior, but it does not establish the handling of logs or backups. The agency asks for the remaining documentation and records precisely what the demonstration proved.
This approach avoids both extremes: rejecting a service because one answer needs clarification, or approving every data use because a broad assurance sounded satisfactory.
Make a bounded decision and keep the record current
When the questions are resolved sufficiently for the intended use, document the approved task, information categories, configuration and accountable person. Keep restrictions visible to staff. Approval for drafting generic messages should not become unreviewed permission to upload references or connect an entire database.
Set a review trigger for changed terms, new integrations, added features or a different use of candidate information. Keep vendor notices somewhere the responsible person actually checks. When a material answer changes, revisit the affected decision rather than relying on an old screenshot or sales email.
The final question is practical: can your agency explain this specific use of information from start to finish, with evidence for the important answers? If not, identify the gap before widening access.
Sources
ICO: AI recruitment data protection considerations, checked October 27, 2026. The register, fictional vendor exchange and operational questions are original NAJ tools. They do not constitute contract or legal review.