What this lesson is

Compliance for an AI PM is not a law school course. It is a small, repeatable set of questions you ask before code ships and before a vendor sees your data. The questions don't change the law. They change whether your DPO walks into the room saying "we have a problem" or saying "we have a process."

Two things make AI features more interesting than the privacy reviews you've done before. One — AI eats free text. Users paste account numbers, phone numbers, passport scans, names of their accountant, into prompts that classical forms would never have collected. Two — your AI vendor is a data processor in the legal sense, even when they feel like a stateless API. What they do with your prompts after the response is sent — log, train, fine-tune, share with sub-processors — is your problem on paper.

So this lesson teaches awareness, not law. By the end you should be able to (a) draw the PII flow through any AI feature on a whiteboard, (b) read a vendor's data-use clause and know what to push back on, (c) tell DPDP from GDPR in the four ways that matter to a PM, and (d) recognise the five situations where you stop and email Legal before the next sprint.

The one-line takeaway:
a PM doesn't ask "is this legal?" — that's the DPO's job. A PM asks "is this legal-flagged?" and produces the artefact that makes the answer obvious.

The six dimensions every AI PM must know

"Privacy" is too big a word to be useful in standup. Break it into six concrete dimensions. For each one there is a question, an artefact, and a fix. If your feature is silent on any one of these, that silence is the bug.

🛡️
PII redaction
A detector strips names, account numbers, phone numbers, and IDs before the prompt leaves your servers. Logs store the redacted form plus a flag, never the raw value.
🌍
Data residency
Where the bytes physically sit. India-only? EU-only? Cross-border with safeguards? DPDP and GDPR both care; users care more than they say in research.
🤝
Vendor data-use
The clause in the contract that says whether your prompts can be used to train the vendor's next model. Default is often yes on free tiers and no on enterprise.
✅
Consent
The user agreed, in plain language, to the specific processing this feature does. "We use your data to improve services" is too broad to count under DPDP or GDPR.
📜
Audit logs
Tamper-evident records of who saw what, when, and which model version returned which output. Required for breach forensics and for regulator visits.
⏳
Retention
How long inputs, outputs, and labels are stored, and the automated deletion that proves it. "Forever" is not a retention policy.
How a PM uses these: at the start of a feature, draw a 6-row table. For each dimension, fill in what we do today and what the policy says we should do. The rows where those two answers differ — that is your compliance backlog. Everything else is fine.

DPDP vs GDPR — the four-page version, in six rows

You will hear "DPDP" in India and "GDPR" in Europe and they are not the same regime. But for the PM-level decisions you make every sprint, they overlap more than they differ. Here is the laminated comparison. Read it once; come back for the rows that match the geography you ship in.

Dimension DPDP (India) GDPR (Europe)
Jurisdiction Applies to processing of digital personal data of "Data Principals" inside India, plus offers of goods/services into India. Applies to processing of personal data of subjects in the EU/EEA, regardless of where the processor sits.
Consent Specific, informed, free, unconditional, unambiguous — and presented in a clear notice in plain language. Withdrawable. Specific, informed, freely given, unambiguous. Same shape — withdrawable, granular, no pre-ticked boxes.
Data residency Cross-border transfers allowed unless government notifies a country as restricted. Sectoral rules (e.g. RBI, IRDAI) may add stricter localisation. Transfers outside the EEA need an adequacy decision, Standard Contractual Clauses (SCCs), or another lawful basis.
Data-subject rights Access, correction, completion, erasure, grievance redressal, nominee. Acted on within a reasonable period. Access, rectification, erasure, restriction, portability, objection, automated-decision opt-out. Generally one month to respond.
Breach notification Notify the Data Protection Board and affected Data Principals. Timing per Board rules — assume hours, not days. Notify the supervisory authority within 72 hours of becoming aware; notify subjects if high risk.
Fines (ceiling) Up to ₹250 crore per instance, set by the Data Protection Board against the data fiduciary. Up to €20 million or 4% of global annual turnover, whichever is higher.
PM-level read:
the two regimes converge on the four things you will actually be asked about in a sprint review — do you have consent, do you know where the data sits, can you delete on request, and can you tell us within hours if something leaked. Build for those four; the rest is your DPO's craft.

Scenario — BoraPay's fraud explainer hits the DPO desk

📱
Real-world scenario
BoraPay Nairobi — PM Amina has 10 weeks, a shipping fraud explainer, and a DPO who wants a 1-pager by Friday.

Amina ships fraud UX at BoraPay in Nairobi. The fraud explainer she's been building (Lesson 09) goes from "Transaction declined" to a short, honest reason in Swahili or English. The model pass rate is comfortably over her gate. Canary is two weeks away. On a Tuesday she gets a calendar invite from Esther, the Data Protection Officer, titled "AI-feature 1-pager: PII flow + vendor clauses."

The ask is small but precise. Esther wants one A4 page that shows: which fields cross which servers, who the vendor is and what they may do with prompts, where the logs sit, and how a deletion request is honoured. No legal language. Just the picture and the answers. Without it, the launch gate fails.

Amina has a model that works and a vendor she likes. What she doesn't have is the artefact. The next four sections are how she builds it.

How PII actually flows — the strip Amina draws first

Esther wants to see the bytes move. So Amina starts with a strip diagram: where the user types, where redaction happens, where the prompt goes, where the response comes back, and where the log lives. If she can draw this and label each box with a server location, the rest of the 1-pager almost writes itself.

User input "Why was the KSh 15,000 declined?" PII detector strip name, phone, account, KRA PIN Redacted prompt "<NAME> sent KSh X to <TILL> 02:14 AM" Model vendor API no-train clause Response re-hydrated to user (Swahili) Log PII flag 90-day retain PII flow — every box names a server, a region, and a retention rule Boxes 1, 2, 3, 5, 6 sit on BoraPay servers in KE · only box 4 leaves · audit log is append-only with 7-day key rotation A deletion request walks this strip backwards: log row → response → input → confirmation to user within 30 days

Three things to notice. One — the redactor sits inside BoraPay, not at the vendor. Sending raw PII to the vendor "for them to redact" defeats the point. Two — re-hydration (putting the user's name back into the response) happens server-side after the model responds, never in the client. Three — the log stores the redacted prompt plus a PII-flag pointer, never the raw value. Forensics and deletion both rely on that pointer.

The PM's risk matrix — severity × reversibility

Compliance decisions are not all equally scary. The two axes that matter for a PM are severity if it goes wrong and reversibility once it does. Plot every privacy decision on this 2×2; the top-right is where you stop and email Legal before the next commit.

↑ High severity (regulatory or user harm) Low severity ↓ Reversible Irreversible → Slow down · review STOP · escalate to Legal PM call · document Watch list · monthly review Train on user data (vendor) Cross-border transfer w/o SCC Raw PII in logs Vendor lock-in (single API) Consent text too broad 90-day vs 30-day retention Eval golden set frozen w/ PII
Why severity × reversibility, not severity alone: a reversible bug is a backlog ticket. An irreversible one is a press release. "Vendor trained on our prompts" is the textbook irreversible-high-severity case — there is no take-back. Treat that quadrant as a hard escalation gate.

The 1-pager Amina hands to her DPO

This is the artefact Esther asked for. It is one page. It uses no legal language. It maps directly to the six trait cards and the flow strip you saw above. PMs in the cohort produce one of these per AI feature — it lives next to the PRD in Notion.

Notion · BoraPay · Risk & Compliance
AI Feature 1-pager — Fraud Explainer (v0.9 · canary in 2 wks)
1
PII flow
User input → redactor (BoraPay KE) → vendor API (EU region, no-train) → response re-hydrated server-side → log (BoraPay KE, redacted only). Diagram: see strip above.
2
Fields touched & classification
PII: name, phone, KRA PIN, account-tail. Sensitive: transaction amount, recipient. All redacted before vendor call. Five-field allow-list documented in flows/fraud_explainer_v1.json.
3
Vendor & data-use clause
Enterprise tier · "Vendor will not use customer data to train, fine-tune, or improve foundation models" · 30-day vendor-side log · sub-processor list reviewed quarterly.
4
Consent & user-facing notice
Plain-language banner on first use in EN/SW: what data, why, who sees it, opt-out path. Granular toggle for "use my prompts to improve our model." Default off.
5
Audit logs & retention
Append-only log: input-hash, redacted prompt, model version, response-hash, user-id, timestamp. Retain 90 days. Auto-delete job runs nightly. Forensics replay tested.
6
Deletion & data-subject rights
Erasure walks the flow backwards: log → response → input. Vendor deletion API hit on the same job. Confirmation to user within 30 days. Eval golden-set carve-out: see Hard Q below.
What changed in Esther's review:
the v0.9 above is the third draft. v0.7 sent raw PII to the vendor "because they have better redaction." v0.8 stored raw prompts in the log "for debugging." Both got rejected in 30 seconds — and that 30-second rejection is the whole point of the artefact.

The AI connection

Real AI teams operationalise compliance through three pieces of plumbing — a PII detector in front of the model, a contract clause that disables vendor training, and an append-only audit log keyed to model version. The PM owns the policy decisions; engineers own the implementation; the DPO signs off the artefact. pii-redactionno-train-clauseaudit-logdata-residencyretention-policy

🛑 Stop and escalate to Legal when…

  • A vendor's free or default tier may train on customer prompts. Do not ship before this is confirmed in writing.
  • Data crosses a border and you don't have SCCs or an adequacy basis on file.
  • The feature processes children's data, biometric data, or health data — these have stricter rules under both DPDP and GDPR.
  • You're asked to "log everything for debugging" — raw PII in logs is the single most common fine-trigger.
  • A user asks for deletion and you can't honour it because the data is locked inside an eval set, a fine-tune, or a vendor index.

Mini challenge — write the 1-pager for a feature you use

Pick any AI feature you actually use. Spend 30 minutes writing its compliance 1-pager.

1. Draw the PII flow strip in 5 boxes

User input → ? → ? → vendor → ? → log. Put a region label on every box ("US-East", "EU-West", "phone"). Circle every box that holds PII. If you can't fill a box, that gap is your finding.

2. Read the vendor's data-use clause

Open the privacy policy of the AI vendor behind the feature. Search for "train" and "improve." Highlight the one sentence that decides whether your prompts feed their next model. Three to five lines, max.

3. Plot three decisions on the 2×2

Pick three: PII in logs, training-data clause, retention window, cross-border, vendor lock-in. Drop each on the severity × reversibility 2×2. The top-right ones are the ones you would email Legal about today.

Interview questions

Easy What's data residency, and why does an AI PM care?
+
Data residency is the physical location where your data is stored and processed. An AI PM cares because (a) regulators care — DPDP can flag transfers to restricted countries, GDPR needs SCCs or an adequacy decision for non-EEA transfers, and sectoral regulators (RBI in India, CBK in Kenya) often add stricter localisation; (b) users care more than they say in research — "your data stays in India / Kenya / the EU" is a feature; (c) it changes vendor selection — a US-only inference endpoint may rule out a vendor for an India-residency-required feature. The PM's job is to know where every box on the PII flow strip physically sits, before procurement signs anything.
Easy What's the difference between an audit log and an analytics log?
+
An analytics log is for product insights — funnel events, latencies, A/B keys — and it is usually mutable and aggregated. An audit log is for forensics and regulator visits — it is append-only, tamper-evident, keyed to a user, a model version, and a timestamp, and it can be replayed to reconstruct exactly what happened in one transaction. AI features need both, but they are not the same store. Conflating them is a classic mistake — analytics logs get truncated for cost, audit logs must not. PMs flag the requirement at design time so engineering picks the right storage.
Medium A vendor wants to train on your prompts. What do you do?
+
Push back, in this order. One — ask if they have an enterprise tier with an explicit no-train clause; on most vendors that is a contract toggle, not a price increase, and it gets you 80% of the way there. Two — if no, ask whether you can opt out per request via an API parameter, and confirm it in writing not just in docs. Three — if both fail, escalate to Legal and Procurement; this is an irreversible-high-severity decision (top-right of the 2×2) and not a PM call. Never assume "they probably won't" — vendor terms change, ownership changes, the data leaves your control the moment it lands in their training pipeline. The PM artefact is one paragraph in the 1-pager naming the vendor, the tier, the clause, and the date it was last verified.
Medium Your engineer says, "let's just log the full prompt for debugging." How do you respond?
+
No, but offer the equivalent. The reason engineers want raw prompts is reproducibility — they need to recreate a bad answer. You can give them that without storing PII: log the redacted prompt, the redaction map (which placeholders mapped to which token classes), the model version, and the response. For escalations, allow a short-window (24-hour) raw-prompt store behind a separate access role, with every read itself audit-logged. That gives engineering 95% of the debugging power without the 100% of the regulatory exposure. Frame the conversation as "what problem are we trying to solve?" not "no" — engineers respect a PM who finds the smaller cut.
Hard A user requests deletion of their data. How do you handle prompts and responses already in your eval golden set?
+
This is the question that separates PMs who have actually shipped from PMs who have only read the policy. The honest answer is: a properly designed golden set never contained the user's PII to begin with — entries are redacted at curation time and re-checked before promotion. If you did your job, deletion of the user's account does not affect the golden set, because the golden set holds the shape of the question, not the user's identifying data.

If you didn't do your job — and the golden set has raw PII or quotes — you have two choices, and both have costs. One: purge those entries on deletion and re-curate replacements. This is the right answer under both DPDP erasure and GDPR Article 17, but it shrinks your eval coverage and you'll lose comparability with prior model versions. Two: argue the eval set is anonymised aggregate research data and qualifies for an exception. That is a Legal/DPO call, not a PM call, and you should not assume it will hold.

The senior-PM move is to prevent the question from arising. Three controls. (a) Redact at curation, not at use — the golden set never sees raw PII. (b) Tag every entry with the source user-id (hashed) so that on a deletion request you can verify there is nothing to remove, and prove it. (c) Treat the golden set as a regulated artefact in your data map — it shows up on the PII flow strip with its own residency and retention policy. The same logic applies, even more sharply, to fine-tune sets — once a user's data is fused into model weights, deletion is effectively impossible without re-training, which is why sensible AI PMs avoid fine-tuning on PII at all.

Test your knowledge

Question 1 of 5
Where should the PII detector sit in an AI feature's flow?
Question 2 of 5
Which is the cleanest definition of data residency for a PM?
Question 3 of 5
Under GDPR, what is the standard window to notify the supervisory authority of a personal data breach?
Question 4 of 5
Your AI vendor's free tier may use prompts to train future models. Your enterprise tier doesn't. Which is the PM-correct action?
Question 5 of 5
A user requests deletion of their data. Where is the hardest place to honour it from?
0
/5
You can map a PII flow.

Common mistakes first-time AI PMs make

Trusting "vendor probably won't train on it"
Defaults change. Ownership changes. The only safe artefact is a no-train clause in the contract, with a date you re-verify.
Logging raw prompts "for debugging"
Single most common DPDP/GDPR fine-trigger. Log redacted prompts plus a redaction map; gate raw access to a 24-hour window with its own audit log.
Treating the eval golden set as outside the data map
If it has user PII, it is in scope for deletion. Redact at curation; tag every entry with a hashed user-id so you can prove erasure later.
Writing consent text the lawyer wrote
"To improve services" is too broad to count under DPDP or GDPR. PM-write the plain-English version, then have Legal redline it — not the other way around.

Key takeaways

  • Compliance is six dimensions: PII redaction, data residency, vendor data-use, consent, audit logs, retention. Silence on any one of them is the bug.
  • DPDP and GDPR converge on four PM-level questions: do you have consent, do you know where the data sits, can you delete on request, can you tell us within hours if something leaked.
  • Draw the PII flow strip first. If you can't label every box with a region and a retention rule, you can't ship.
  • Treat the no-train clause and cross-border transfers as the top-right of the severity × reversibility 2×2 — escalate to Legal before code, not after.
  • Your job is awareness and the artefact (the 1-pager). Your DPO's job is the legal answer. Confusing the two is how launches slip.
Africa & India's Best AI Academy
🤖 AI-first 🎓 Placement Assistance

You now know what compliance asks.
Next — decide what to buy and what to build.

Join the Best AI Program: The AI-Era Agentic Engineer — 5 months, fully online, with placement assistance. Week 1 is free.

Explore The AI-Era Agentic Engineer →