Privacy & AI Playbook
Transcript
Made automatically from the video's sound, so it may contain mistakes. Click a line to play from there.
Coming off our sixth explainer on AI, privacy and consent, today we're pulling the whole shebang together into a practical, no-nonsense launch playbook. We've designed this specifically for you product managers, founders, developers, and marketers out there.
But a quick heads up before we dive in. While we're going deep into these frameworks, this is general information, current as of November 2026, and absolutely not legal advice. Right?
Let's build something compliant. One, baseline docs, two, QA testing, three, AI launches, and four, red lines. Let's kick things off with section one, baseline docs, which covers artifacts one and two.
Your absolute first line of defense before a single line of code is pushed to production is Artifact 1, your data map or record of processing activities. Basically, this turns GDPR Article 30 into a living breathing table for your team. Now, you might have heard a rumour about a GDPR exemption for companies with under 250 employees.
Well, let me be crystal clear. That carve out almost never applies to regular SAS data processing. Why? Because the exemption completely evaporates the second your processing becomes more than just occasional.
According to official UK ICO and French CNIL guides, you have got to track your data categories, your purposes, your Article 6 lawful basis, your time to live, and your vendors. Keeping this as an active living document is gonna save you countless headaches down the road.
Moving on to Artifact 2, which is your privacy notice. Over in Europe, GDPR articles 12 to 14 demand a plain English, easily accessible notice. That means absolutely no dense legalese, especially if kids might be accessing your service.
But look across the Atlantic and California's CCPA takes a different approach, specifically mandating a notice at collection. Plus, if you cross the threshold of buying, selling, or sharing the personal information of 100,000 or more California residents, boom, you immediately trigger strict requirements to provide a highly visible do not sell or share my personal information link.
Knowing exactly which threshold you hit is literally half the battle. QA testing. This covers artifacts three and four. Let's jump into our active QA environment.
Artifact 3 is consent verification. And this actually plays out as a literal bug testing sequence right in your browser's DevTools. Step one, under ePrivacy, Article 5.3, absolutely zero non-essential tags can fire before the user actually clicks.
Step 2. Following the EDPB Cookie Banner Task Force mandate, you must have completely symmetrical reject all and accept all buttons on layer one, so no sneaky dark patterns hiding the reject button.
And step three, for the US, your site must honor the W3C global privacy control. In plain English, if a browser sends an HTTP header with sec GPC set to one, you must automatically opt that user out of data sharing. It's binary, folks.
Either it works or it's a bug you need to fix immediately. Now, Artifact 4 covers your vendor and subprocessor review. You simply cannot just blindly trust third parties anymore. Under GDPR Article 28, you need strictly enforced data processing agreements.
Sending data outside the EU? Well, then you're relying on standard contractual clauses, specifically decision 2021 914. If you're using US vendors, you must independently verify their active certification on data privacyframework.gov.
And here is a massive one. You absolutely must explicitly lock your large language model APIs to zero training tiers, so those vendors do not train their models on your user's personal data. That is absolutely crucial.
Alright, section three, AI launches, covering artifacts five and six. Now we're escalating to the really high stakes phase. Artifact 5 is your DPIA screening. Under GDPR Article 35, before you write a single line of code for a risky new feature, you've got to screen it using the WP248 revision 19 high-risk criteria.
If your feature hits two or more of those, like say processing sensitive data or using innovative tech, you must run a formal data protection impact assessment. And here is the hard stop. If a high risk remains that you just cannot mitigate, Article 36 clearly dictates that you must stop and consult the supervisory authority before shipping anything.
You cannot just wing it here. For Artifact 6, AI risk triage, you need to map these exact EU AI Act compliance dates directly against your product roadmap right now. Article 50 transparency rules actually went live back in August 2026, so if you're using chatbots, users must know they're interacting with a machine, and any deep fakes or synthetic media must be clearly labelled.
Looking ahead, if your AI touches employment sorting or critical infrastructure, Annex 3 high-risk rules will apply from December 2027. And if it's embedded into regulated physical products, Annex 1 rules hit in August 2028.
Finally, we reach section 4, red lines. This is exactly when to drop the DIY approach and call real legal counsel. If you hit any of these exact triggers we're looking at right now, you cross a hard boundary where startup hygiene ends and existential threats begin.
Drop everything and call a qualified lawyer. These triggers include meeting mandatory Article 37 DPO thresholds, processing Article 9 special category data, like health or biometric info, or processing children's data under 13, which triggers copper in the US.
It also absolutely includes building Annex III high-risk AI, making Article 22 significant automated decisions, scraping the internet for data, or facing a live security breach, because that last one starts a brutal 72-hour countdown to notify regulators under Article 33.
If you hit these lines, you need professional counsel right away. So, the ultimate question to leave you with as we wrap up this entire Explainer series is this.