Key takeaways
- Ask the agent for a data map – one row per data type with source, destination and purpose – not for finished answers.
- Audit the release build; debug builds can include different SDKs and permissions.
- Collected, shared, optional and ephemeral have precise meanings in Google's definitions; the agent should flag edge cases rather than settle them.
- Features that send user content to an external AI model count as data leaving the device and belong in the map.
- Re-run the map whenever dependencies change, so the form and the app never drift apart.
What the form asks – and where an agent helps
Under App content in Play Console, the Data safety form asks per data type whether your app collects it, whether it shares it, for which purposes, and whether users can choose not to provide it. It also asks about encryption in transit and about deletion requests. Google makes you responsible for the complete picture, including what third-party libraries do; the Play Console Help page on the Data safety section defines every term.
The form itself is not the hard part – knowing the true answers is. In an app assembled over many agent sessions, nobody remembers that an analytics package arrived as a dependency of the crash reporter, or that the profile screen uploads photos to a moderation service. Reading an entire project, following every network call and listing the results with evidence is exactly what a coding agent is good at.
Step 1: let the agent build a data map
Ask for a table with one row per data type and these columns:
| Column | Example entry |
|---|---|
| Data type, in Google's categories | Email address |
| Source in the code | Sign-up screen, with file and line |
| Destination | Own backend, users endpoint; authentication SDK |
| Own code or SDK | Both |
| Purpose, as the agent reads it | Account management |
| Leaves the device | Yes |
| Confidence | High, or UNSURE with a reason |
The sources the agent should work from: the release runtime dependency tree from Gradle, the merged manifest of the built App Bundle, every HTTP client call in the code, each SDK's initialisation code and the collection settings it documents – for example whether analytics collection or advertising ID access is switched on. For Flutter, React Native or Capacitor apps, include the generated Android project; plugins bring in native SDKs and permissions you never see in Dart or JavaScript.
Talk to an AI Code Engineer about your Google Play Data safety answers.
In a free call we look at what you built, tell you honestly what stands between it and an approved release – and whether you need us at all.
Step 2: check the map against real traffic
Reading code misses things: SDKs that send data on first launch before your consent screen, server-driven features, endpoints assembled at runtime. So a person captures traffic from a fresh install of a debuggable copy of the release build through an intercepting proxy, walks through the main flows and saves the capture. The agent then reads the exported capture and compares the hosts and payloads with its map. Every destination missing from the map becomes an open question.
Our Data safety rejection guide lists the audit commands, proxy configuration and SDK disclosure sources in detail – it makes a good instruction file for the agent.
Step 3: the decisions that stay human
With the map finished, the answers are decisions rather than lookups. These are the cases the agent should flag and leave to you:
- Collected or processed ephemerally. Data held only in memory to serve one request does not count as collected; the moment it is logged or stored, it does.
- Shared or sent to a service provider. Passing data to a processor acting on your behalf is not sharing in Google's sense; passing it to an ad network for its own use is.
- Optional. Only when every user can avoid providing that data.
- AI features. A chat, voice or photo feature that sends user content to an external model provider means that data leaves the device. Whether the provider acts as a service provider depends on its terms – read them rather than assume.
- Encrypted in transit. True only if every endpoint, SDK endpoints included, uses TLS.
- Deletion. Answer yes only when users can really request it. Apps with accounts need an in-app path and a web link; see the deletion requirement and our playbook on building account deletion with an agent.
Step 4: filling in the form without surprises
Play Console can export the Data safety answers as a CSV file and import one back; an import replaces the current answers. A routine that works well with an agent:
- Export the current answers.
- Give the export and the approved data map to the agent and ask for a proposed CSV plus a plain-language list of every changed answer.
- Review the list of changes line by line, and diff the proposed file against the export.
- Import, then check the preview of the store listing section before saving.
Never import a file the agent changed without that diff. Then bring the privacy policy in line with the same map – data types, partners, retention and the deletion process. The agent can draft the changed paragraphs; the policy remains yours.
Keeping the answers true over time
Most Data safety problems start with an update, not with the first release: a new SDK, a plugin upgrade that adds a permission, a feature that sends one more field to the server. Add a CI step that records the release dependency list and merged permissions and fails a pull request when they change without an updated data map. Tell the agent in its memory file that any new SDK or permission means a map update in the same pull request. New permissions deserve a look at the sensitive permissions policy, too.
The iOS side of this work is the privacy manifest plus the App Privacy details – see privacy manifests with AI agents. Our AI Code Engineers run this audit as part of the Android submission service and go through each answer with the account holder.
Data safety checklist for agent-built Android apps
- The agent built a data map from the release dependency tree, merged manifest, network calls and SDK settings.
- Cross-platform plugins were audited in the generated Android project.
- A traffic capture from a fresh install was compared with the map, and every unknown host was explained.
- Each "collected", "shared" and "optional" answer was decided by a person using Google's definitions.
- Data sent to external AI services appears in the map and in the form.
- The encryption and deletion answers are true for every endpoint and every account type.
- The CSV the agent prepared was diffed against the export before import.
- The privacy policy names the same data types, partners and deletion process as the form.
- CI flags dependency or permission changes that arrive without an updated data map.
Frequently asked questions
Can an AI agent fill in the Data safety form for me?
It can prepare everything: the data map, a draft CSV and a list of proposed answers. The account holder should make the final answers and import them, because Google treats the form as the developer's own declaration.
My own code sends nothing. Can I answer "no data collected"?
Only if no SDK in the release build sends anything either. Analytics, crash reporting, ads and authentication libraries usually do, and their data counts as collected by your app.
Do internal testing builds need a Data safety form?
Apps that exist only on the internal testing track are exempt. As soon as you use closed or open testing or production, the form has to be complete and accurate.
Can you help us get the Data safety section right?
Yes. Book a free call and an AI Code Engineer goes through your build, SDKs and current answers with you. We can also handle the whole Play submission in your own account.
Free, 15 minutes, no obligation
Get your app from agent output to approved release.
Our AI Code Engineers take apps built with Claude Code, Codex, Cursor, Lovable, Base44 and co. through signing, CI/CD, store listings and review – in your own developer accounts.
AI tools, store policies and requirements change quickly – always check the current documentation of the tools and stores you use. This article is independent advice and not affiliated with or endorsed by Apple, Google, Anthropic, OpenAI or any other company or product mentioned; all trademarks belong to their owners.