Ordering Lunch for One / AI-assisted product development
Privacy by design for AI: seven principles you can check
Useful work. Sensible boundaries. Evidence you can question.
Imagine asking an AI assistant to turn meeting notes into a client proposal. It asks to connect your email. All of it. Somewhere nearby, a reassuring phrase appears: “Privacy by design.”
What should that promise change about the software? And how can you check it before handing over somebody else’s information?
At Mevanu, privacy by design means building protection and human control into every stage of a task—and making those decisions open to scrutiny. Here is a practical way to apply that standard when choosing an AI tool or building one with an agent.
Start with one task and three questions
- Name the job. “Draft a proposal from these selected notes” gives a tool a clearer boundary than “help with my business.”
- Trace the information. What does the tool read, where does it send it, and what does it keep?
- Ask for a check. How can you establish that the promised boundary holds—including when you refuse access?
Begin with invented notes while you examine the settings and documentation. You can investigate the workflow without using a real client’s records as the experiment.
What privacy by design actually means
Privacy becomes a design requirement early enough to shape how a system works. Protective defaults matter too: the starting settings should limit collection to the job’s needs.
Switzerland’s Federal Data Protection and Information Commissioner explains this design-and-default approach in its technical and organisational measures guide, section 5.1. It calls for considering data protection during system design and keeping default collection and use to what the purpose requires.
The seven principles below are plain-language applications of the established Privacy by Design framework. Our AI examples illustrate how to question a product; they are not demonstrations of Brava’s current behaviour.
Seven principles, applied to a real kind of task
We will follow one fictional example: a consultant uses selected meeting notes to prepare a proposal. The names, project and records are invented.
1. Anticipate the risk
A meeting note may contain a useful deadline, an agreed budget and an unrelated personal detail. Before building the workflow, decide which information belongs in the proposal and what would happen if the wrong person received it.
Ask: Which mistakes or misuses did the team plan for?
Look for: A description of the risk, the safeguard chosen and the check used to challenge it. “The assistant should be careful” leaves an important engineering question unanswered.
2. Make the default protective
For this task, start with selected notes. Broader email access needs a reason tied to the job. A feature that genuinely needs more information should explain that need and the consequences before access is expanded.
Ask: What can the tool access before I change any settings?
Look for: Clear initial permissions and a useful, narrower route where possible. In a demonstration with invented data, decline an optional permission and see what happens.
3. Build the boundary into the system
Telling an AI “do not read other files” is an instruction. Restricting which files its tools can open is a system control. A model instruction can help, but the software’s access controls need scrutiny too.
Ask: What prevents the assistant from reading an unselected file or sending a proposal before approval?
Look for: A permission boundary and checks of both allowed and blocked actions. If you commission software, ask the developer to demonstrate the blocked case with harmless test files.
4. Keep the useful work
Privacy and usefulness deserve to be designed together. A proposal may need the project’s goal, scope and deadline while leaving out personal details unrelated to the work. Try that smaller input and check whether the result remains useful.
Ask: Could we complete this task with less information or narrower access?
Look for: A working example that preserves the important result. Replacing a name with “Client A” does not necessarily make a record anonymous: other details may still identify the person. The Swiss guide explains this distinction in sections 5.2–5.5.
5. Protect the whole journey
Follow the notes beyond the upload button. Consider the working copy, a model provider’s copy if one is sent, saved drafts, logs and backups. Logs record what the software did and may contain information that needs protection too.
Ask: Where does each copy go, how is it protected, and when is it deleted?
Look for: Defined retention periods—how long information is kept—and an explanation of deletion, including backups and exceptions. Removing a chat from view does not by itself establish that every copy has gone. The Swiss guide addresses access, encryption and destruction in sections 7 and 8.
6. Make the practices checkable
A privacy statement tells you what a company says it does. Settings, technical documentation and properly scoped review provide different kinds of evidence. Each helps answer a different question.
Ask: Who checked this claim, which version did they examine, and what remained outside their review?
Look for: A dated account of what was examined, what was found and what remains open. Public summaries can explain safeguards and limits while protecting customer information and sensitive security details. Independent review should examine the actual work and evidence; a second AI agreeing with the first is insufficient.
7. Keep the person in charge
You should be able to understand the choices that affect your information. For the proposal, that might mean choosing the notes, reviewing recipients and deciding whether to send. Accessible controls and intelligible language belong in the design.
Ask: Can I narrow access, withdraw a connection and stop future work? What happens to information already shared?
Look for: An explanation of each action’s actual effect. Stopping a task and deleting stored information are separate operations; neither should be implied by a vague button label.
Turn a reassuring phrase into a check
A useful answer connects a particular claim to a safeguard and evidence. The following are questions to investigate, not results from a tested product.
| Claim | Follow-up question | Useful evidence |
|---|---|---|
| “It runs on your laptop.” | Which parts run here, and does any content go to an external service? | A data-flow explanation identifying local and remote processing for the feature you use. |
| “Your data is encrypted.” | During transfer, in storage, or both? Who can access the readable content? | Documentation of the protection and access model. Encryption alone does not establish acceptable collection or use. |
| “You can delete your data.” | Which copies, on what timetable, and with what exceptions? | A retention and deletion policy that accounts for the relevant systems and copies. |
Copy a short privacy review note
Use this blank template for one feature in one tool. Put “unknown” beside an unanswered question and link to the source of each answer. Keep the note free of client records and other confidential content.
Task and feature:
Tool, plan and date checked:
Information needed / information excluded:
Access allowed / actions requiring approval:
Local processing / external recipients:
Training or other reuse of inputs and outputs:
Storage, retention and deletion:
What refusing or revoking permission changes:
Evidence checked, version and limits:
Open questions and next decision:
You can ask an agent to help read a provider’s public documentation:
For this feature and plan, find the provider’s official statements about access, external recipients, training or reuse, retention and deletion. Link each answer to its source and date. Separate documented policy from behaviour actually tested. Mark missing answers unknown. Do not upload my files, connect accounts or change settings.
Check the returned passages yourself. An assistant’s summary can help you find the relevant detail; it cannot certify the provider’s practice.
If you are building with an agent
Give the developer a concrete requirement: “The proposal feature may read only the selected notes and must not send the draft without approval.” Ask how the implementation enforces it and how a reviewer can check it.
Use invented inputs to check the ordinary path, an unselected file and a refused permission. Keep the expected result, observed result, software version and unresolved questions together. For consequential work, a suitably qualified reviewer should inspect the implementation and evidence independently of its author.
These checks are a starting point for a bounded feature; they cannot establish the privacy of an entire product. Our AI-generated code review guide develops the relationship between requirements, implementation and independent review.
If you offer an AI-assisted service, the same note can help you explain a proposed client workflow before requesting access. You retain responsibility for deciding what information you are authorised to use and for reviewing the deliverable. Establishing the boundaries early may reduce later rework; any time saved needs to include setup, checking and corrections.
What this asks of Mevanu
Brava is in development. Local-first and all seven privacy principles are design and launch requirements for us. Our public commitments give people a standard to question; they do not establish that every requirement has been met.
We need to connect each principle to an implemented safeguard, evidence of its behaviour and an honest account of what remains open. Ask us the same questions you would ask another provider.
For your next AI task, start with the first line of the review note: what information does this job actually need? That is a useful question before the first connection request appears.
Sources and optional technical notes
Research checked 8 November 2026. The Swiss FDPIC’s Guide to Technical and Organisational Data Protection Measures (PDF, 50 pages; 15 January 2024) supports design/defaults (§5.1), identification risk (§§5.2–5.5), access (§7), lifecycle protection (§8) and outsourced processing (§9.7).
The Ontario Information and Privacy Commissioner’s Privacy by Design factsheet (PDF, two pages; January 2018) sets out the seven foundational principles. The headings and AI examples here are our practical explanation of them. Neither regulator has reviewed or endorsed Brava or this guide.
A review should distinguish the written requirement, the system control, the result of a check and independent assessment. Use the right expertise for the data and consequences involved. This guide is a practical starting point, not a finding of legal compliance or a product certification.
Want help defining sensible data boundaries and review questions for your AI workflow? Talk with Mevanu.