A PRD says what to build. A DRD says what "correct" means, where the data comes from, and who is allowed to touch it.
A Data Requirements Doc, or DRD, is the artefact that makes you useful to an ML team. It is not a long PRD. It answers a different set of questions, and it is the document engineers will actually open twice, which is more than can be said for most PRDs.
Here is the structure. Ten sections. Aim for four to six pages, no more.
| # | Section | What goes in it |
|---|---|---|
| 1 | The behaviour we are buying | One paragraph. What the feature does for a user, in their language. If you cannot write this without the word "AI", you do not have a feature, you have a technology. |
| 2 | Definition of correct | Three examples of excellent output, three acceptable, three unacceptable. Real examples, written out in full. This is the heart of the doc. |
| 3 | Failure taxonomy | The ways this can go wrong, ranked. Separate "annoying" from "unacceptable". Name the one failure that would make the CEO call you. |
| 4 | Data inventory | A table: source, volume, owner, freshness, quality notes, are we legally allowed to train on it. One row per source. |
| 5 | Sourcing plan | What we have, what we must create, what we will buy. With costs and dates. |
| 6 | Labeling plan | Who labels, against which guideline, at what volume, at what target agreement (kappa), with what QA sampling rate. |
| 7 | Architecture decision | Prompt, RAG, fine-tune, or a combination, justified by section 4. If your justification does not reference the data inventory, you guessed. |
| 8 | Evaluation plan | The golden set, how big, who wrote it, what score gates a release, and the one metric a customer would actually notice. |
| 9 | Privacy and risk checklist | Lawful basis, PII handling, retention, deletion, consent, third-party terms, bias review, human-in-the-loop points. |
| 10 | Monitoring and refresh | How we detect drift, who looks, how often we retrain or re-index, and the feedback signal we are capturing from production. |
Two sections carry disproportionate weight, and if you only do two, do these.
Section 2, definition of correct. Write the actual sentences. Not "the answer should be accurate and concise". Write the answer you want, word for word, for three real customer questions. Then write three answers that would be unacceptable and say why. An engineer can build against that. Nobody can build against an adjective.
Section 9, privacy and risk. This is where PMs cost their companies money by staying quiet. Two live examples that should be on your radar.
WHAT THE REGULATORS HAVE ACTUALLY DONE
Italy and OpenAI. In December 2024, Italy's data protection authority (the Garante) issued a €15 million fine against OpenAI, finding among other things that personal data had been used to train ChatGPT without an adequate legal basis and without proper transparency. It was the first GDPR enforcement action of its kind against a generative AI company. Worth knowing the sequel: the Court of Rome annulled that decision in March 2026. The fine did not survive, but the lesson did. Every AI company in Europe spent eighteen months and considerable legal budget on a question that a documented lawful basis would have answered on day one. The EU AI Act. In force since August 2024, with obligations phasing in. Following the Digital Omnibus on AI, which the Council finally approved on 29 June 2026, the high-risk obligations for standalone Annex III systems, which include recruitment, credit scoring, and education, now apply from 2 December 2027, and embedded high-risk systems from 2 August 2028. Transparency obligations under Article 50 still apply from 2 August 2026. If your product generates synthetic content for the EU market, labelling and machine-readable marking obligations land in December 2026. Check the current position on the European Commission's AI Act page before you commit to dates in a doc, because this has already moved once. India. The DPDP Rules, 2025 were notified on 14 November 2025. Consent Manager provisions come into force around November 2026 and the substantive compliance obligations from May 2027. Penalties reach ₹250 crore for failures of reasonable security safeguards. Breach reporting to the Board is required within 72 hours.
You are not expected to be a lawyer. You are expected to be the person who put the question in writing before the model was trained, because retraining a model to remove data is vastly harder than deciding not to include it.
YOUR MOVE THIS WEEK
Take the PRD for your current AI feature. Add sections 2, 4, and 9 to it. Nothing else. Then send it to your ML lead and ask "does this change anything for you?" The answer, in my experience, is yes, and usually within an hour.