Data Protection Impact Assessment (DPIA) for Retrieva
| Controller | Codea Solutions (sole proprietorship, eenmanszaak), Prinses Beatrixstraat 5, 5953 LL Reuver, the Netherlands, Chamber of Commerce (KvK) 42180669 |
| Service | Retrieva, a study app with AI (retrieva.nl, retrieva.eu) |
| Privacy contact | info@retrieva.eu (English) / info@retrieva.nl (Dutch) |
| Version | 1.2, 26 October 2026 (see §12 for what changed) |
| Related documents | Privacy statement and terms version 2026-10-26.2, the record of processing activities and the data breach procedure |
| Status | Adopted by the controller; legal texts reviewed by a lawyer (see §11) |
| Next review | No later than 25 October 2027, or earlier on any change listed in §10 |
This DPIA follows Article 35(7) GDPR. It contains:
- a systematic description of the processing operations and their purposes;
- an assessment of necessity and proportionality;
- an assessment of the risks to data subjects;
- the measures to address those risks.
The Dutch version is the original. This English version has the same content.
1. Why a DPIA?
A DPIA is required when processing is likely to result in a high risk. The Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) and the European Data Protection Board (EDPB, guidelines WP248) list criteria. If processing meets two or more of them, a DPIA is generally required. Five apply to Retrieva:
| Criterion (WP248) | Applies to Retrieva? |
|---|---|
| Vulnerable data subjects | Yes. The target group starts at the first year of secondary school (about age 12). |
| Innovative use of new technology | Yes. Generative AI creates explanations and questions, and assesses open answers. |
| Evaluation or scoring | Yes, limited. The AI assesses answers to practice questions. This is feedback for the pupil only and has no consequences (see §6). |
| Data on sensitive or highly personal matters | Possibly. Photos of homework may accidentally contain names or other data of third parties. |
| Transfers outside the EEA | Yes. Hosting, AI and backups (also) run through the United States. |
Not applicable: systematic monitoring, large-scale processing (not yet at launch), matching or combining datasets from different sources, and preventing data subjects from exercising a right or using a service.
Conclusion: a DPIA is required. This assessment covers private (consumer) accounts, for which Codea Solutions is the controller. If a school uses Retrieva, the school is the controller and Codea Solutions the processor; see §10 for what is then needed in addition.
2. Description of the processing
2.1 What Retrieva does
A pupil or student creates subjects and adds study material: a photo, a PDF, a Word or PowerPoint file, or typed text. From this, Retrieva produces:
- explanations, practice questions, flashcards, summaries and a study plan up to the test;
- step-by-step help with homework;
- an assessment of open answers.
Whatever was answered wrong comes back until the pupil knows it. At the pupil's request, a parent or guardian can receive a weekly summary containing totals only.
There is a free 7-day trial. After that, a paid subscription is available: Complete (one person) or Family (three separate accounts), monthly or yearly.
2.2 Purposes
- Providing the study service: account, study content, AI features, planning and progress.
- Payment, invoicing and keeping the legally required records.
- Preventing misuse and protecting minors: content moderation, limits on AI use, and a limited number of trial accounts per connection.
- Security, availability and recovery: backups, a log of administrator actions, monitoring.
- Communication:
- service emails (confirmation, trial, subscription, invoices);
- reminders and the weekly summary for parents, only if the user switches them on;
- news, only to people who have or had a subscription.
- Measuring whether promotion works, in a limited way and without tracking technologies (see §4.3).
- Improving the service through "Is this wrong?" reports and reasons given when cancelling or deleting an account.
2.3 Data subjects
| Group | Notes |
|---|---|
| Pupils and students | The core group. From about age 12; under 16 with permission from a parent or guardian. |
| Adult users | For example for a driving theory test, a professional exam or a course. |
| Payers | 18 or older. Often a parent on a Family subscription. |
| Family members | Invited by the payer. Each has their own account. |
| Parents or guardians receiving the weekly summary | A third party, only after their own confirmation (double opt-in). |
| Third parties in uploaded photos | For example the name of a classmate or teacher on a worksheet. Not intended; see R2. |
| Website visitors | Counted without a cookie and without lasting personal data. |
2.4 Personal data
| Category | Data | Source |
|---|---|---|
| Account | name, email address, password (scrypt hash only; older accounts still bcrypt); when logging in with Google: name, email and Google ID | User, Google |
| Study content | subjects, questions, answers, notes, flashcards, tests, planning, results, photos and files | User, AI |
| Settings | language, theme, daily goal, reminders, parent's email address | User |
| Usage | days and hours of activity (hourly data kept 90 days), AI actions and AI costs per day | System |
| Payments and invoices | plan, amount, date, status, Mollie reference, invoice details; for bank transfers also the name and account number shown on the bank statement | User, Mollie, bank |
| Consent records | time and version of acceptance of the terms and privacy statement, age confirmation, waiver of the right of withdrawal | User |
| Origin | channel, campaign and ad from the link, language page, country (from the browser language), sign-up method | The link with which the app was opened |
| Security | HMAC of the IP address at sign-up (7 days); IP address only briefly in working memory | System |
| Content moderation | category and time of a hit, never the content | OpenAI moderation |
| Reports and reasons | "Is this wrong?" (with the account), reason for cancelling (with the account), reason for deleting (anonymous) | User |
| Email log | address, subject, type, delivered or not; no content | System |
Special categories of personal data (Art. 9 GDPR) and criminal-offence data are not requested and not needed. Users are asked not to upload such data: in the terms, in the privacy statement and when uploading. See R2 for what happens if they do anyway.
2.5 Systems and parties
| Party | Role | What they receive | Where | Transfer basis |
|---|---|---|---|---|
| Replit, Inc. | Processor: hosting of the app, the database (SQLite, with PostgreSQL for durable storage), uploaded files and server-side backups | All data in §2.4 | European Union (Replit, Inc. is based in the US) | Replit's general terms and privacy policy; Replit only offers a separate data processing agreement on more expensive plans |
| Anthropic, PBC (Claude) | Processor: AI for explanations, questions, summaries, study plans and assessment; the analyst in the admin panel receives totals only | Text and images the user submits for an AI feature, with language and level; no name or email address | United States | SCCs in Anthropic's data processing agreement (part of its commercial terms) |
| OpenAI (agreement with OpenAI Ireland Ltd.) | Processor: moderation of input before the AI; only when photo questions are switched on, also creating an image for a question | Input for AI features (text, photos, files) without name or email; for images only a description written by our AI | Ireland / United States | SCCs in OpenAI's data processing agreement |
| Google (Google Ireland Ltd. / Google LLC), via Google Workspace | Processor: all outgoing email (info@retrieva.nl and info@retrieva.eu); Google Drive of the Workspace account for the encrypted backups and the invoice PDFs | Emails (address, content of service emails); backups encrypted only (AES-256-GCM, key not held by Google); invoice PDFs | Ireland / United States | EU-US Data Privacy Framework (Google LLC is certified) and SCCs in the Google Workspace data processing terms |
| Mollie B.V. | Independent controller for payment data | Name, email, amount; bank and card details go directly to Mollie | Netherlands | n/a (EEA) |
| Google (Sign in with Google) | Independent controller | Only if the user chooses to log in with Google | Ireland | n/a |
| Monitoring service, Meta | Not recipients of personal data: the monitoring service only checks whether the site is reachable; from Meta we only fetch totals for our own ads | Nothing about individuals | n/a | n/a |
No Meta Pixel, no Conversions API, no third-party analytics. By default, Replit injects its own statistics script into published pages. It has been switched off in the deployment, and our Content Security Policy would block it in any case.
2.6 Data flow (simplified)
Pupil's browser ──https──▶ Retrieva server (Replit, EU) │ ├─▶ database and files (Replit, EU) │ ├─▶ moderation (OpenAI) ──▶ AI answer (Anthropic, US) ──▶ back to the pupil │ ├─▶ email via Google Workspace (service email, invoice, parent's weekly summary) │ ├─▶ daily backup, encrypted ──▶ server + PostgreSQL (Replit) + Google Drive (Workspace) │ └─▶ Mollie (only when paying; bank/card directly at Mollie) └── only functional storage in the browser (login cookie, settings); no tracking cookies
2.7 Retention periods
| Data | Period |
|---|---|
| Account and everything the user created | Until the user deletes it, or after 24 months without activity, with an email warning about 30 days in advance. A running paid subscription is never removed. |
| After deletion | Removed from the database and disk immediately. If an administrator deletes the account, a snapshot is kept for 30 more days. |
| Backups | One per day, encrypted, 10 days at most, in all three locations. |
| Payments and invoices | 7 years (tax retention obligation). Payments are unlinked from the account when it is deleted. |
| Reason for deletion | Anonymous, 24 months at most. |
| Email log | 6 months. |
| Log of administrator actions | 12 months. |
| Hourly activity, content moderation hits | 90 days. |
| Payment attempts that were not completed | 30 days. |
| HMAC of the IP address at sign-up | 7 days. |
| Messages to us | As long as needed, then 2 years at most. |
| Login sessions | 24 hours at most, or 7 days with "Remember me". |
The app enforces these periods itself every night.
3. Lawfulness
| Processing | Legal basis (Art. 6 GDPR) |
|---|---|
| Account, study content, AI features, settings, planning | Performance of a contract (b) |
| Payments, invoices, withdrawal | Contract (b) and legal obligation (c, tax retention) |
| Consent records, age confirmation, waiver of the right of withdrawal | Legal obligation and legitimate interest (proof) |
| Content moderation, limits on AI use, trial accounts per connection | Legitimate interest (f): protecting minors and preventing misuse |
| Backups, log of administrator actions, security | Legitimate interest (f) |
| Origin at sign-up | Legitimate interest (f). Nothing is stored on the device (see §4.3). Users can object. |
| Reminders, weekly summary for parents | Consent (a). Can be switched on and off in Settings, with an unsubscribe link in every email. |
| News by email | Legitimate interest plus Art. 11.7(3) of the Dutch Telecommunications Act (own customers, own service). Only to people who have or had a subscription; can be switched off at sign-up, in Settings and from every email. |
| "Is this wrong?", reason for cancelling | Legitimate interest or consent; always voluntary |
Minors. The account is based on the contract, so the age-16 threshold of Art. 8 GDPR does not apply to the service itself. It does apply to processing based on consent: reminders and the weekly summary. At sign-up, a separate checkbox has users confirm that they are 16 or older, or that a parent or guardian has given permission. For processing based on consent, Retrieva goes further (since version 1.1):
- When someone switches on reminders, the app first asks whether they are 16 or older. If they are younger, the reminders only start after a parent or guardian has confirmed by email (double opt-in, the same system as for the weekly summary).
- The weekly summary only goes to a parent or guardian who confirms it themselves by email.
The controller considers these reasonable efforts within the meaning of Art. 8(2) GDPR, because:
- reminders and the weekly summary are limited in nature;
- the weekly summary goes only to a parent who confirms it themselves by email;
- a parent confirms from their own email address before anything based on consent starts;
- the parent receives no questions, answers or notes;
- everything can be switched off with one click;
- on a Family subscription, the parent pays and manages the accounts.
Only someone aged 18 or over can take out a paid subscription.
4. Necessity and proportionality
4.1 Purpose limitation and proportionality
Each processing operation belongs to a purpose in §2.2. Study content is not used for advertising, not sold and not used to train AI models. Under the API terms of Anthropic and OpenAI, they do not use API data for training.
4.2 Subsidiarity
The AI features are the core of the service and cannot work without sending the study material. An AI service within the EU was considered. Anthropic (Claude) was chosen for the quality of its explanations, combined with data minimisation (no name or email) and contractual safeguards (SCCs). Replit was chosen for hosting, with the servers, the database and storage in the European Union.
4.3 Data minimisation
- The AI receives only the study material plus language and level, never a name or email address.
- Moderation runs without a name or email. For a hit we keep only the category and time, never the content.
- Inappropriate uploads are deleted immediately and are not included in backups.
- IP addresses are not stored. Only an HMAC (which cannot be reversed) is kept for 7 days to limit sign-ups per connection.
- Origin travels only in the link to the app. Nothing is stored on the device: no cookie and no local storage. For a referring site we keep only the domain name. A Meta tracking ID (fbclid) is removed from the address immediately.
- Visitors are counted with a daily hash that means nothing the next day.
- Country is derived only from the browser language, never from the IP address.
- The payer of a Family subscription sees only the name and email address of family members, not their work.
- The parent's weekly summary contains the time practised (also per subject), the number of questions and percentage correct, the topics that are still difficult and the tests in the next two weeks. No questions, answers or notes.
- The reason for deleting an account is stored anonymously.
4.4 Accuracy
AI can make mistakes. This is stated in the terms, on the website and in the app. Therefore:
- Retrieva first shows what it read from a photo.
- Questions go to a review screen first.
- A question without an explanation is rejected.
- Users can report mistakes with "Is this wrong?".
- Users can always change an AI assessment themselves.
4.5 Storage limitation
See §2.7. Everything is removed automatically, including from backups (after 10 days).
4.6 Transparency
- The privacy statement, cookie statement, terms and company details are complete in Dutch and English. French, German and Spanish show the English text, with a notice that where versions differ, the interpretation most favourable to the consumer applies.
- The country and transfer basis are stated for each processor.
- AI transparency: users know that content is generated by AI (Art. 50 AI Act).
4.7 Data subject rights
- Access and portability: in the app, a zip with all data linked to the account (JSON, with CSV, ICS and Markdown) and the uploaded files; separate parts also as Word or PDF. Only the password and secret keys are left out.
- Erasure: by the user in the app, immediately.
- Rectification: name, email and password in the app.
- Objection: to origin measurement and news, by email or via the unsubscribe link.
- Withdrawal of consent: in Settings or via the unsubscribe link.
- Requests by email: response within 2 working days, handled within one month at the latest.
- Parents can make requests on behalf of their child; we verify that they are the parent.
- No decisions with legal effects (Art. 22 GDPR). AI assessments are feedback for the pupil only.
5. Transfers to the United States
| Party | Mechanism | Additional safeguards |
|---|---|---|
| Replit | Replit's general terms and privacy policy | https; passwords hashed; backups encrypted with our own key; uploads accessible to the owner only |
| Anthropic | SCCs (Anthropic data processing agreement) | No name or email; no training on API data; only what is needed for one answer |
| OpenAI | SCCs (agreement with OpenAI Ireland Ltd.) | No name or email; moderation without the content being kept by us; images only from an AI-written description and only when photo questions are on |
| Google (Workspace) | DPF (Google LLC) and SCCs (Workspace terms) | Backups encrypted with AES-256-GCM; Google does not hold the key. Email is transactional. Workspace account protected with two-step verification |
Transfer assessment. We consider the risk of access by US authorities to this data to be low. It concerns study material and learning data with no political, financial or special-category sensitivity. Content goes to the AI services without identifying data. Backups are stored at Google in encrypted form only. The live database at Replit is not encrypted in this way. It is stored in the European Union, but Replit is a US company and can be asked for access under US law. Without a data processing agreement with standard contractual clauses (which Replit only offers on more expensive plans), this is the largest residual risk. It is accepted, with the safeguards in the table.
An appeal against the EU-US Data Privacy Framework is pending before the Court of Justice of the EU. If the framework falls, the SCCs remain as the legal basis.
6. Automated decision-making and AI
- The AI creates study content and assesses open answers to practice questions. This is feedback for the pupil only: no school grade, no admission and no other decision with legal or similarly significant effects follows from it. Art. 22 GDPR therefore does not apply.
- The pupil can change any assessment.
- AI Act. For private use, Retrieva does not fall into the high-risk category. The transparency obligation (Art. 50) does apply and is met. If a school uses Retrieva to evaluate learning outcomes or to place pupils, this may fall under Annex III, point 3 (education). A separate assessment is then required (§10).
7. Risks to data subjects
Scale: likelihood low, medium or high; severity limited, significant or serious. Residual risk is the risk after the measures.
| # | Risk | Likelihood | Severity | Measures | Residual risk |
|---|---|---|---|---|---|
| R1 | Unauthorised access to study content and photos through a stolen password or a flaw in the separation between accounts | Medium | Significant (photos of minors, personal notes) | scrypt hashes; limits on login attempts; session cookie httpOnly and not sent to other sites; uploads only for the logged-in owner (a link gives no access); automated test of separation between accounts; "Log out everywhere" and a notice on every new admin login; Sign in with Google available | Low |
| R2 | Third-party data or special-category data in uploaded photos (classmates' names, health) | Medium | Limited to significant | Instruction in the terms, the privacy statement and at upload to cover such data; sent to the AI without the user's name or email; photos for the owner only; deletion always possible; retention periods; no training on content | Low to medium (accepted; depends on user behaviour) |
| R3 | AI provider retains or uses input longer or differently than expected | Low | Significant | Data processing agreements with SCCs; under their terms no training on API data and only temporary retention for safety; no identifying data sent; only what is needed for one answer | Low |
| R4 | Access by authorities outside the EEA (US) | Low | Significant | SCCs or DPF per party; backups encrypted with our own key; minimisation towards the AI; nature of the data (study material) | Low to medium (accepted; see §5) |
| R5 | Incorrect AI explanation or assessment, so a pupil learns something wrong | Medium | Limited | Clear notice that AI makes mistakes; showing first what was read; review screen; no question without an explanation; "Is this wrong?"; assessment can be changed; no consequences outside the app | Low |
| R6 | Minors and inappropriate content or misuse (sexual material, hate, self-harm, using the AI for things other than learning) | Medium | Serious | Moderation of all input before the AI; inappropriate uploads deleted immediately and kept out of backups; for self-harm a notice with help resources, counted without the account; AI can be switched off for an account after repeated hits; safety rules in every AI prompt; usage limits | Low |
| R7 | Data breach via backups (Drive or PostgreSQL) | Low | Serious | AES-256-GCM with our own key, not held by Google; Workspace account with two-step verification; 10 days at most; administrators only | Low |
| R8 | Keeping data too long | Low | Limited | Automatic deletion per §2.7, every night, including backups; warning before deletion after inactivity | Low |
| R9 | Profiling or unwanted marketing | Low | Limited | No tracking cookies, pixels or third-party analytics; origin not stored on the device and limited to domain or campaign; news only to (former) paying customers and can be switched off at sign-up; unsubscribe link in every email; no profiles | Low |
| R10 | Weekly summary gives a third party (parent) insight into learning behaviour | Low | Limited | Only if the user sets it up; the parent confirms themselves (double opt-in); totals only; can be switched off by user and parent | Low |
| R11 | Misuse by an administrator (insider) or a compromised admin account | Low | Serious | Admin panel with password, emailed code and authenticator app; notice on every new login; actions logged (12 months); one administrator; two-step verification on all linked supplier accounts | Low |
| R12 | Loss of data (availability) | Low | Significant | Daily backup in three locations (server, PostgreSQL, Google Drive); restore tested, with an automatic safety copy and a per-table check; availability monitoring; export for the user | Low |
| R13 | Payment data | Low | Significant | Bank and card details only at Mollie; we keep only amount, status and reference; invoices 7 years as required by law | Low |
| R14 | Rights cannot be exercised | Low | Limited | Export and deletion by the user in the app; response within 2 working days; parents can make requests for their child | Low |
| R15 | Credentials or keys leaking (API keys, sessions) | Low | Serious | Keys in a protected secrets store, not in the code (checked, including the history); the source code is not public; admin sessions stored as hashes only; breach procedure with steps to revoke keys | Low |
8. Measures in summary
Technical
- https everywhere, with strict security headers (HSTS, CSP allowing only our own scripts, X-Frame-Options DENY).
- Passwords as scrypt hashes; backups encrypted with AES-256-GCM and our own key.
- Uploads accessible to the owner only; automated tests for separation between accounts, misuse, and complete export and deletion.
- Moderation of all AI input; limits on AI use per account (per day, per 30 days, per minute, concurrent requests), and a cost alert by email on the total AI costs per day.
- Minimisation towards the AI (no name or email); no tracking; origin not stored on the device.
- Automatic retention periods, including in backups.
- Daily backups in three locations, with a tested restore process.
Organisational
- Data processing agreements with Anthropic, OpenAI and Google Workspace; Replit under its general terms and privacy policy.
- Record of processing activities; data breach procedure (notification within 72 hours; to schools within 48 hours).
- Two-step verification on all admin and supplier accounts.
- Lawyer's review of the legal texts (October 2026).
- This DPIA is reviewed every year, and earlier when things change.
9. Conclusion
After the measures, the residual risks are low. For R2 (third parties in photos) and R4 (transfers to the US) the residual risk is low to medium. Both are deliberately accepted, for the reasons in §5 and §7.
No high residual risk remains. Prior consultation of the Dutch Data Protection Authority (Art. 36 GDPR) is therefore not required. The processing may proceed as described.
10. When this DPIA must be revisited
Review on:
- a new AI provider, a different AI model with different processing, or new AI features being switched on (such as photo questions for everyone);
- a move to another hosting provider, or to a Replit plan with a data processing agreement. Hosting is already in the EU; a data processing agreement with standard contractual clauses would lower R4 further;
- use by schools. The school is then the controller, and the following are needed:
- a DPIA by or with the school;
- a data processing agreement based on the Dutch Model Data Processing Agreement 4.0 (Privacyconvenant);
- an assessment under the AI Act if Retrieva is used to evaluate learning outcomes (Annex III, point 3). We will not take on schools before that assessment is done. We will check the start date of the high-risk obligations (2 December 2027 according to the Digital Omnibus) in the Official Journal;
- new categories of data or data subjects, or new forms of marketing or measurement;
- a data breach with risk to data subjects;
- 25 October 2027 at the latest.
11. Record
| Prepared | 15 October 2026, updated 25 and 26 October 2026, based on Retrieva's working code and documentation |
| Legal review | The legal texts were reviewed by a lawyer in October 2026. The advice has been implemented, except the point about ending a yearly subscription early after a change in features (a deliberate choice). |
| Open actions | (1) Keep copies of the accepted data processing agreements of Anthropic, OpenAI and Google Workspace, with dates. (2) At the first school: the points in §10. (3) Check whether Art. 50(2) AI Act (machine-readable marking) applies to the images with questions; they are made with AI. |
| Adopted by | The owner of Codea Solutions, 25 October 2026 |
12. Changes
Version 1.2 (26 October 2026)
- Replit: the app, the database and storage run in the European Union (the region of the published app). Replit, Inc. remains a US company. On 26 October 2026 Replit confirmed that its data processing agreement does not apply to the Core plan; the DPIA already said so.
- Reference to privacy statement and terms version 2026-10-26.2.
Version 1.1 (25 October 2026)
- Replit: corrected. Version 1.0 mentioned a data processing agreement with SCCs; Replit is used under its general terms and privacy policy, as the privacy statement also says.
- Passwords: scrypt (version 1.0 said bcrypt).
- Reference to the current version of the privacy statement and terms.
- Fewer internal details: no names of keys, accounts, services or files.
- Reminders for under-16s: only after confirmation by a parent (§3).
- Weekly summary: the description now matches what it contains (§4.3).
- Export: a zip with all data and the uploaded files (§4.7).
- AI Act: no schools before the assessment is done; check Art. 50(2) (§10, §11).