🗂️ The only layer you can actually own

Every PM eventually notices the same thing: the model changes every six months, the UI changes every quarter, the prompt gets rewritten every sprint — but the data quietly compounds. Logs pile up. Labels get better. The golden set gets sharper. That's the layer that keeps working for you after the hype cycle moves on.

Which is why data strategy is not a backend topic. It is a PM topic. You decide what gets captured, what gets labeled, who owns it, how long it lives, and whether a vendor is allowed to train on it. If you don't, someone else will — and their decisions will be optimized for their economics, not yours.

This lesson gives you the PM-level mental model for data. What to collect. How to label it without drowning. When to buy a public dataset, when to hand-collect, when to let the model generate its own. And the boring-but-critical stuff: PII, retention, audit trails, vendor clauses. By the end you should be able to draw your feature's data flow on a whiteboard and answer three questions — what do we own, what could a competitor copy, and what would regulators want to see?

The one line to remember:
If your feature improves when you ship more usage, you have a data moat. If it doesn't, you're renting one from a vendor.

🔄 The data pipeline, end to end

Every AI feature that gets better over time runs the same loop. A user does something. You log it. A human (or an LLM judge) labels it. The best examples become a golden set. That golden set drives evals and, eventually, fine-tuning. Skip a step and the loop breaks.

User interaction question + click thumbs + text Logged event input + output PII stripped Labeled human or judge pass / fail / why Golden set curated truth 50–500 rows Evals + fine-tune model gets better The compounding loop — every user makes the next user's answer better Skip any box and the loop breaks — usage stops improving the product

Three things go wrong in real products. One — teams log the output but forget the input, so when a bad answer shows up nobody can reproduce it. Two — feedback is captured but nobody reviews it, so thumbs-down data rots in a BigQuery table. Three — the golden set is never formally promoted; it lives in a shared spreadsheet and nobody trusts it. PMs own all three problems.

📦 What to actually collect

When a PM says "we need more data" that's not a strategy — it's a wish. Break "data" into the six concrete artifacts below. Each has a different collection cost, a different shelf life, and a different regulatory profile. If your current feature doesn't produce all six, that's your backlog.

⌨️
Inputs
What the user typed, tapped, or uploaded — plus the context (page, time, session). Free to collect. Useless without the matching output.
💬
Outputs
Exactly what the model returned — tokens, latency, which version, which prompt. Log every single one, even on failure. You'll need them to reproduce bugs.
🏷️
Labels
A human (or trusted judge) saying pass, fail, or partial, with a reason. The most expensive item per row — and the only one that converts raw logs into a golden set.
👍
Feedback
Thumbs, edits, "copy to clipboard", time-on-answer. Implicit signals are 10× more abundant than explicit thumbs. Use both, weighted.
🔥
Failure cases
Refusals, timeouts, hallucinations, user retries. Store these in a separate, priority table. The next model version lives or dies on how you handle these.
🧭
Metadata
Model version, prompt version, retrieval doc IDs, user cohort, geography, language. The boring column everyone forgets — until you need to slice a regression.
PM call: if you could only instrument three of the six in sprint one, pick inputs, outputs, and metadata. Everything else can be back-filled from those three. Labels and feedback arrive in sprint two. Failure cases surface themselves once the other four are in place.

🇰🇪 Scenario — BoraPay Nairobi, fraud explanations

📱
Real-world scenario
PM Amina has 10 weeks to ship an AI that explains why a transaction got flagged — and no PII is allowed to leave BoraPay servers.

Amina leads fraud UX at BoraPay in Nairobi. Today, when the risk engine flags a send — say KSh 15,000 to a new till — the user sees "Transaction declined. Please try again." and rings the call centre. 60% of those calls are legitimate users who just want to know why. Her CEO wants an AI that writes a short, honest reason inside the app, in Swahili or English, in under two seconds.

Amina lists what she needs: a label taxonomy of ~20 fraud reasons (unusual hour, new recipient, device-switch, velocity, sanctions list, etc.), the real user questions people send to the call centre, and the historical outcomes — was the flag confirmed fraud, false positive, or never resolved? The engineering team offers her a vendor that trains on customer data. She says no — DPA regulation and the CEO's own line both forbid it. Every label, every input, every output has to stay on BoraPay servers.

By week 3 Amina has a 12-week label backlog on her hands. By week 6 she realizes the reasons the risk team uses internally don't match the way users ask questions. The data work is the product work. The model choice will take one day; the data pipeline will take ten weeks.

🧮 Buy vs collect vs generate

Three ways to get training data. Most teams mix all three. The mix you pick shows up in your margins, your defensibility, and your legal review — so make it an explicit call, not a default.

Dimension Buy (public / licensed) Collect (your own logs + labels) Generate (synthetic, LLM-written)
Quality Mixed. Broad but not your domain — public fraud sets won't know Kenyan till codes. Highest, if your labeling discipline is real. Matches your users exactly. Fine for templated tasks. Dangerous for edge cases — synthetic data mirrors the generator's blind spots.
Cost Flat licence fee or free. Predictable. High. Labelers at ₹30–₹80 per row, or $0.50–$3 in KE/NG markets. Near-zero per row — but you pay in review time. Bad synthetic data is invisible until eval.
Defensibility Zero. Your competitor can buy the same CSV. High. This is the moat. Every day of usage widens the gap. Medium. The seeds matter — whoever has the better seed prompts gets better synthetic data.
Privacy / legal Easy — already cleared. But licences usually forbid resale or commercial fine-tuning. Hard. PII, retention, DPDP/GDPR/DPA. You own the legal exposure entirely. Easy — no real users. Unless the seed prompts leaked real PII, which happens more than vendors admit.
Speed to first usable batch 1–2 weeks. 6–12 weeks from zero. Faster if you already have logs. Days. Fastest — but quality plateaus quickly.
When to pick Cold start, generic task (sentiment, summarization), internal tool. User-facing, regulated, anything where "our data" is part of the pitch. Augmentation — fill long-tail gaps your real data misses. Never 100% of a training set.

Where each source sits on volume vs quality

Public datasets are huge but noisy. Your hand-labeled golden set is tiny but pristine. Synthetic is cheap and infinite but brittle. Your own production logs are the interesting box — high volume and, with effort, high quality.

↑ High quality Low quality ↓ Low volume High volume Hand-labeled the golden set 50–500 rows, evals live here Your production logs the compounding asset label a slice, eval on the rest Synthetic LLM-generated fill gaps, don't train on alone Public datasets CSVs, HuggingFace cold start only — not a moat
PM heuristic:
spend the first sprint in the bottom-right (public + synthetic) to get to a working prototype. Spend every sprint after that moving rows from the right two boxes into the top-left — that migration is your data strategy.

The compounding curve

Labeled data does not grow in a straight line. It grows slowly while you're figuring out the taxonomy, then accelerates once labelers and the review process click. The chart below is what Amina's pipeline looks like if she actually ships the loop.

BoraPay labeled rows, by quarter Q1 0.5k Q2 2.5k Q3 6k Q4 12k Q5 20k Purple tip = new this quarter · bar = cumulative · competitor starting today is already 5 quarters behind
What Amina does with this: she protects Q1 and Q2 from scope creep. The reason the curve bends in Q3 is almost always that the team finally nailed the label taxonomy in Q1–Q2. Ship a taxonomy review in week 3, not week 9.

🤖 The AI connection

How real AI teams think about data

Every serious AI team separates three artifacts in their head: the raw event log (everything, cheap storage), the labeled training set (a subset, with ground truth), and the golden set (tiny, sacred, never used for training — only eval). They version all three and refuse to launch any model change that hasn't been run against the current golden set. That discipline is the difference between "we shipped an LLM feature" and "we ship an LLM feature every month and it keeps getting better." raw-logs labeled-set golden-set PII-scrubbing data-contract retention-policy

🎯 Mini challenge — draft Amina's data plan

20 minutes. No code. Use a whiteboard or a Notion page.

1. Draw the pipeline

Draw Amina's five-box flow: user question → logged event → label → golden set → eval/fine-tune. At each arrow, name the owner (PM, eng, ops, legal) and the artifact produced (a table, a taxonomy, a review doc). If an arrow has no owner, that's where the pipeline will break.

2. Draft the label taxonomy

List 10 reasons BoraPay might flag a transaction. Group them into 3 categories (behavioural, network, policy). For each, write the one-sentence explanation you'd want the AI to give a user. This is your first golden-set seed.

3. Decide the buy/collect/generate mix

For three layers — the reason taxonomy, the user questions, the historical outcomes — pick one of buy, collect, or generate. Write one sentence of why and one of what you'd give up. That tradeoff paragraph is your data strategy in miniature.

🎤 Interview questions — what a recruiter will probe

EasyWhat's the difference between raw logs and a golden set?
+

Raw logs are every input and output the system has ever produced — huge, noisy, unlabeled, often millions of rows. A golden set is a tiny, curated, human-labeled slice (usually 50–500 rows) that represents the task well enough to serve as ground truth. Raw logs are for debugging and future labeling; the golden set is what you run every model or prompt change against before it ships. You never train on the golden set, only evaluate against it — otherwise it stops being an honest measure.

EasyWhat are the six artifacts a PM should make sure get collected?
+

Inputs (what the user sent), outputs (what the model returned), labels (human verdict + reason), feedback (thumbs, edits, implicit signals), failure cases (refusals, hallucinations, retries), and metadata (model version, prompt version, cohort, locale). Miss any one and something breaks: no inputs, no repro; no labels, no evals; no metadata, no regression slicing.

MediumHow do you handle PII in the data pipeline?
+

Four moves, in order. One — strip or hash PII at the edge, before the event hits your log store. Phone numbers, names, bank account numbers, national IDs go through a scrubber. Two — segregate what can't be scrubbed (the actual message body) behind tighter access controls and short retention windows. Three — write the retention policy into a data contract: who can read this, for how long, and for which purpose. Four — build auditability in from day one. A regulator asking "show me every person who accessed customer 12345's conversation last quarter" should be a query, not a project. PII isn't a backend concern — if it leaks, the PM will be in the incident review.

MediumWhen would you choose synthetic data over real user data?
+

Three cases. Cold start — you have no users yet and need prompts to debug the prototype. Long-tail gaps — your real logs have 95% "where is my money" and almost nothing about sanctions-list hits, so you synthesize the missing 5% to make evals meaningful. Sensitive edge cases — you need to test refusal behaviour around suicide, abuse, or fraud without exposing real users in the set. What you never do is train 100% on synthetic data for a user-facing feature; synthetic data mirrors the generator's blind spots, and those blind spots compound silently.

HardYour vendor wants to use your data to train their model. How do you negotiate?
+

Start from the asymmetry: they get a free, one-way option to improve their product using your users' behaviour. You get nothing back — their next-gen model helps your competitors too. So the answer is almost always no, and the negotiation is about how to say no cleanly.

Four concrete asks: (1) no-train default — opt-out of training on your data baked into the contract, not a toggle buried in a console. (2) Data residency — specify the region (for BoraPay, it's Kenya); reject anything that ships data outside without an explicit clause. (3) Retention window — 30-day log retention on the vendor side, not 12 months "for abuse prevention." (4) Audit right — you (or a third party) can audit that (1)–(3) actually happen. If the vendor refuses (1) or (4), walk. If they agree to (1) but jack the price 3x, run the unit economics on collecting your own data instead — at high volume, you'll often find you can afford it. The negotiation isn't about the price; it's about who owns the compounding asset. Never give that away for a discount.

🧠 Quick check — 5 questions

Question 1 of 5
Amina has logs of inputs and outputs but no labels. What is she missing from the compounding loop?
Question 2 of 5
Which of these is the *only* data source that compounds as a moat over time?
Question 3 of 5
What is the golden set used for?
Question 4 of 5
BoraPay's legal team says no PII can leave Kenya. Which option is Amina ruling out?
Question 5 of 5
A vendor offers a 40% discount if they can train on your data. What's the right PM response?
0
/ 5
Nice — your data instincts are sharpening.

Review any miss, then move on to evals.

⚠️ Four mistakes that kill data strategies

Logging outputs but not inputs
You'll see a terrible answer and have no idea what the user asked. Every bug becomes unreproducible. Always log the pair plus the prompt version.
Treating labeling as a one-off sprint
The taxonomy changes, the product changes, user language changes. Labeling is a standing process with an ops owner, not a Q2 project.
Giving vendors default training rights
The default checkbox on most LLM APIs is "may use to improve service." Flip it off in writing. Your moat isn't their training corpus.
Treating the golden set like a dataset
If you fine-tune on the golden set, you've just deleted your ruler. Keep it locked in a separate table; promote new items in, never out of evals.

✅ Five things to take to your next standup

  • Data, not model, is the only layer in the AI stack you can durably own — compounding is the whole strategy.
  • Collect six artifacts, not "data" — inputs, outputs, labels, feedback, failure cases, metadata. Missing any one breaks a downstream step.
  • Buy for cold start, collect for moat, generate to fill long-tail gaps — never 100% of one source.
  • The golden set is tiny, sacred, and never trained on. Fine-tune on the labeled set; evaluate on the golden set; debug on the raw logs.
  • PII, retention, and vendor training clauses are PM decisions — not legal's. If you don't draft them, someone else's defaults become yours.
Africa & India's Best AI Academy
🤖 AI-first 🎓 Placement Assistance

You now know what data to own.
Next — build the evals it powers.

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 →