Security, data privacy and the production checklist

0:00 Security, privacy, production

Your AI feature works, the demo impressed everyone, and launch is next week. Then the security team asks: what data leaves our systems, where is it stored, for how long, and who can see it? If you can't answer with evidence, launch slips. In this lesson you'll learn the data privacy questions to settle with every provider, the technical controls that protect you, and a production checklist that gets you through review.

0:31 Why it matters

Why does this matter? When you call an AI API, your users' content leaves your systems and goes to a third party processor. That's normal, but it's a responsibility. Think of sending documents to an external translation agency. You'd want a contract, you'd want to know how long they keep copies, where their office is, and you'd black out anything they don't need to see. AI providers deserve exactly the same diligence.

1:02 Four questions per provider

Ask four questions of every provider, product and plan. Training: business APIs from Anthropic, OpenAI and Google say they don't train on your API data by default, but consumer apps and free tiers can differ, so read the terms for the exact product. Retention: providers typically keep inputs and outputs for a limited period for abuse monitoring, and offer zero data retention arrangements for eligible customers. Some features store data by design, like files, batches and caches, and OpenAI's Responses API stores responses unless you set store to false. Location: where processing happens. And agreements: a data processing agreement, subprocessors and certifications.

1:46 Legal context (not legal advice)

Now the legal context, and this isn't legal advice. GDPR and UK GDPR require a lawful basis, data minimization, transparency, processor contracts and safe international transfers. Saudi Arabia's personal data protection law, the UAE's federal law and free zone regimes like DIFC and ADGM add their own requirements, plus sector rules in banking and health. Pakistan has been developing data protection legislation, so check its current status. And the EU AI Act brings phased obligations, like transparency for some AI interactions. Involve your privacy and legal team early.

2:24 Technical controls

Next, technical controls. Minimize: send only what the task needs. Redact personal data before sending where you can. Keep keys on the backend. Defend against prompt injection by treating user content and documents as untrusted, and never letting model output trigger sensitive actions without validation. Handle outputs safely: never render them as raw HTML. Prevent abuse with authentication, rate limits and moderation, and pass a pseudonymous user id where supported, like OpenAI's safety identifier. Log carefully. And keep humans in the loop for consequential decisions.

3:01 Simple example: gym reminders

A simple example of minimization. A gym wants AI to write personalized renewal reminders. The first draft of the code sends each member's full record: name, date of birth, address, phone, payment history and medical notes. The reminder only needs a first name, the membership type and the renewal date. Cutting the payload to those three fields removes most of the risk, cuts tokens, and makes the privacy review a two minute conversation.

3:33 Hands-on: redaction layer

The lesson includes a simple redaction layer. It replaces emails, phone numbers, card numbers, Pakistani CNIC numbers and Emirates ID numbers with placeholder tokens before sending text to a model, and can restore them afterwards if needed. But be honest about its limits. Regular expressions miss names, addresses and free form identifiers, and can over match. Add a named entity detection step, test on real samples, and treat redaction as one layer among several.

4:05 Business example: Lahore fintech copilot

Now a realistic business example. A Lahore fintech built a support copilot that drafts replies for agents. They redact CNIC, card and phone numbers before every call. They chose a provider with a signed data processing agreement and zero data retention on the relevant endpoint. Drafts are never sent automatically, and agents see a label saying the draft was generated with AI. Logs keep metadata for ninety days and redacted text for fourteen. And they run injection tests every quarter using real attack patterns from their inbox. Their security review passed in one cycle, because every item had evidence.

4:48 Production checklist

Finally, the production checklist. Quality: an eval per task with agreed thresholds and per language results. Reliability: error handling, backoff, circuit breaker, tested fallback. Cost: ledger, budgets, cost per outcome. Security: backend keys, secret manager, scanning, injection tests, output sanitization. Privacy: terms reviewed, retention configured, redaction, updated privacy notice. Observability: traces, dashboards and alerts. Operations: versioned config, tested rollback, an owner and a runbook. And compliance: disclosure where required and human oversight for consequential uses.

5:21 Common mistakes

Common mistakes. Sending full records when a few fields would do. Assuming not used for training means not stored. Logging raw prompts with personal data indefinitely. And rendering model output as HTML, which can open you to script injection. Each of these is a common finding in security reviews, and each is easy to fix before launch.

5:46 Answering data questions

What should you do when a customer asks what happens to their data in your AI features? Have a short, accurate answer ready, agreed with legal: which providers process it, for what purpose, how long it's kept, where it's processed, and how they can request deletion. Put it in your privacy notice and your help center. Being able to answer clearly and quickly is one of the strongest trust signals you can give business customers.

6:19 Disclosure and advertising rules

A quick word on disclosure and marketing rules, since many of you will generate customer facing content. Where AI generated content could mislead people, be transparent about it, and follow your market's advertising rules: sponsored posts need clear labels like ad, and claims about health, performance or prices need evidence, whether a human or a model wrote them. Regulators in the UK, the EU, the Gulf and the US all expect marketers to stand behind their claims. Put these rules into your prompts, your validators and your review steps.

6:58 Deeper: the fintech review

Let's deepen the Lahore fintech copilot example. Before launch, the security reviewer asked for evidence on every checklist row. The team provided a data flow diagram showing that CNIC, card and phone numbers were redacted before any provider call, the signed DPA and the zero retention configuration, screenshots of the AI label agents see, log retention settings, and results from their quarterly injection tests using real attack patterns. One gap surfaced: names in free text were not redacted. They added a named entity step and a test set of fifty real, anonymized messages. The review passed in one cycle, and the reviewer reused their checklist as the bank's template for the next AI project.

7:48 Watch me do it: redact() and restore()

Watch me do it. Let's run the redaction layer. Five patterns: email addresses, phone numbers with international prefixes like plus nine two, plus nine seven one or plus nine six six, card numbers, Pakistani CNIC numbers in the five dash seven dash one format, and Emirates ID numbers starting with seven eight four. Redact loops over each pattern, and for every unique match assigns a numbered placeholder such as PHONE one, stores the original in a mapping, and replaces it in the text. Restore swaps the placeholders back, for example after the model drafts a reply that must include the customer's email. I run it on, call Aisha on plus nine seven one fifty, one two three, four five six seven, or her email, about Emirates ID seven eight four and so on. The output replaces the phone, email and ID with placeholders. But Aisha's name is still there, which is exactly why the lesson adds a named entity step.

8:58 Recap + try this now

Quick recap. Ask every provider about training, retention, location and agreements. Know your legal context and involve legal early. Minimize, redact, protect keys, defend against injection and handle outputs safely. And ship with a checklist backed by evidence. Try this now: complete the production checklist for one AI feature, with a link to evidence for every row, add a redaction layer, and review your provider's current data terms and retention settings.