What We Learned Building AI Platforms for Netflix and Zaland

The email from Netflix came in late 2018. They wanted to talk about optimizing their acquisition campaigns across European markets. Not just A/B testing or basic automation—they wanted AI that could adapt campaigns in real-time based on performance data.

We said yes before fully understanding what we were getting into.

That decision led to five years of building AI-powered marketing platforms at Campanja for some of the world's most demanding brands. We worked with Netflix, Zalando, SAS, and dozens of other companies operating at massive scale. We made a lot of things work. We also made a lot of mistakes.

Now that I'm building Cogny, I keep thinking about those lessons. Some of them still apply. But surprisingly, many of them are completely wrong for the world we're entering. Here's what we learned—and what we had to unlearn.

Lesson 1: Perfect Data is a Myth (And You Don't Need It)

What we thought: Before we can build effective AI, we need clean, perfectly structured data.

What we learned: You'll never have perfect data. Build systems that work with messy reality.

At Netflix, we spent three months just understanding their data architecture. Campaign data lived in one system, conversion data in another, customer data in a third. Attribution windows varied by market. Some tracking used cookies, some used first-party IDs, some used probabilistic matching.

Our initial instinct was to clean everything up first. Create a unified data model, standardize attribution, fix all the inconsistencies. We estimated it would take six months.

Netflix's response: "We don't have six months. Can you work with what exists?"

So we did. We built systems that could handle:

Missing data points (impute or skip)
Inconsistent time zones (normalize on the fly)
Multiple attribution models (weight by confidence)
Delayed conversions (update retroactively)

It was messier than we wanted. But it worked. And it shipped in weeks, not months.

The surprise: The AI actually performed better with imperfect data that covered more scenarios than with perfect data from limited sources. The models learned to handle ambiguity, which made them more robust in production.

What's different now: Modern LLMs are even better at handling messy data. They can reason about missing information, ask clarifying questions, and work with whatever context they have. The obsession with perfect data was a limitation of older ML approaches, not a fundamental requirement.

Lesson 2: Optimization Without Explanation is Dangerous

What we thought: If the AI improves performance, clients will trust it.

What we learned: Black box optimization creates anxiety, not confidence.

For Zalando, we built a system that automatically optimized ad spend across channels and campaigns. It worked brilliantly—ROAS improved by 34% in the first month. We were thrilled.

The client was terrified.

Why? Because they couldn't explain to their CMO why the budget was being allocated the way it was. The AI was moving millions of euros around based on patterns humans couldn't easily see. When performance dipped for a few days (normal variance), panic set in. They almost shut down the entire system.

We learned to build explanation layers:

Show why the AI made each major decision
Highlight which data points drove the choice
Compare AI decisions to what humans would typically do
Provide confidence scores, not just recommendations

Once we added transparency, adoption accelerated. The AI wasn't just optimizing—it was teaching the team what good optimization looked like.

The counterintuitive part: Adding explanation actually improved the AI's performance. When we forced the system to articulate its reasoning, it caught edge cases and errors that would have otherwise slipped through.

What's different now: This lesson applies even more strongly to LLM-based agents. The ability to explain reasoning in natural language is a massive advantage. At Cogny, we don't just show what the AI recommends—we show its full reasoning process. Users can challenge it, ask follow-ups, and understand why.

Lesson 3: Real-Time is Hard (And Often Not Worth It)

What we thought: Marketing optimization should happen in real-time to capture every opportunity.

What we learned: Real-time is expensive, complex, and often unnecessary. Near-time is usually good enough.

At Campanja, we built a system for a travel client that optimized ad bids in real-time—updating every few seconds based on flight availability, competitor pricing, and conversion rates. The technical complexity was enormous:

Stream processing infrastructure
Sub-second latency requirements
Coordination across multiple ad platforms
Graceful degradation when systems lagged

It worked, but it cost a fortune to maintain. And when we looked at the actual impact, we found something surprising: 90% of the value came from optimizations that happened every 15 minutes. The real-time adjustments added maybe 5-7% improvement at 10x the cost.

We started recommending different refresh rates based on use case:

Real-time (seconds): High-stakes scenarios like breaking news or flash sales
Near-time (minutes): Most campaign optimization, inventory-based advertising
Batch (hours): Strategy adjustments, budget reallocation, reporting

What's different now: LLM inference times mean true real-time is still expensive. But "good enough" time frames have gotten much shorter. What used to require custom stream processing can now happen in a serverless function. The cost/benefit equation has shifted dramatically in favor of faster optimization.

Lesson 4: Humans Don't Want Full Automation (Yet)

What we thought: Marketers want to set strategy and let AI handle execution completely.

What we learned: People want to stay in the loop, even when automation works perfectly.

For SAS, we built a system that could run entire campaigns autonomously—from budget allocation to creative testing to bid optimization. In testing, it outperformed human-managed campaigns by 20-30%.

The marketing team hated it.

Not because it didn't work. Because it made them feel redundant. They wanted to understand what was happening, have input on decisions, and feel like they were actually doing their jobs.

We redesigned around a "approval workflow" model:

AI suggests major changes (here's what I recommend and why)
Human reviews and approves or modifies
AI implements and monitors
AI reports back on results

Paradoxically, this slower process led to faster adoption. The team felt in control, learned from the AI's suggestions, and gradually increased their trust and delegation.

The insight: People don't actually want full automation. They want augmentation. They want the AI to do the tedious parts and elevate their decision-making on the important parts.

What's different now: This lesson is even more relevant with conversational AI. The ability to have a dialogue with the AI—asking questions, challenging assumptions, iterating on ideas—makes augmentation feel natural instead of threatening.

Lesson 5: Integration Hell is Real

What we thought: Modern marketing tech stacks are well-integrated and API-friendly.

What we learned: Integration is always harder than it looks, and it never stops being hard.

Every single client had a different stack:

Different ad platforms (Google, Facebook, Amazon, local networks)
Different analytics tools (GA, Adobe, custom solutions)
Different data warehouses (BigQuery, Redshift, Snowflake)
Different CRMs (Salesforce, HubSpot, custom builds)

And none of them worked quite the way the documentation suggested.

We spent probably 40% of our engineering time on integration:

API inconsistencies
Rate limiting
Authentication quirks
Data format differences
Timezone handling
Currency conversion
Platform-specific limitations

Building a great AI algorithm was maybe 20% of the work. Making it work reliably across different tech stacks was 80%.

The hard truth: There's no way around this. If you're building marketing AI, you're building an integration platform first and an AI platform second.

What's different now: This is still hard, but the landscape has improved:

More platforms have better APIs
Standards like OAuth have become universal
Reverse ETL tools help with data sync
LLMs can help write integration code

At Cogny, we've focused on deep BigQuery integration rather than trying to connect to everything. Better to do one thing excellently than ten things poorly.

Lesson 6: Performance Drifts Over Time

What we thought: Once you optimize an AI system, it stays optimized.

What we learned: Marketing environments change constantly. Your AI needs to adapt or it becomes obsolete.

We built an audience targeting system for a fashion retailer that worked amazingly well. For about four months. Then performance started degrading. Not dramatically—just a slow decline that was easy to miss.

The cause? The market had changed:

Competitors launched similar products
Customer preferences shifted with seasons
New platforms gained traction
iOS privacy changes altered tracking

The AI was still optimizing based on patterns it learned months ago. It hadn't adapted to the new reality.

We built monitoring systems that tracked:

Model prediction accuracy over time
Feature importance drift
Population stability metrics
Business metric trends

When drift exceeded thresholds, we'd retrain models or alert humans to review strategy.

The realization: AI in marketing isn't a "set it and forget it" solution. It's a living system that needs ongoing attention, just like the campaigns it's optimizing.

What's different now: LLMs don't "drift" the same way trained models do, but they still need updated context. At Cogny, we continuously feed the agent fresh data and context, so its recommendations stay relevant to current conditions.

Lesson 7: Scale Changes Everything

What we thought: If it works at small scale, it'll work at large scale with bigger infrastructure.

What we learned: Scale introduces entirely new problems you didn't anticipate.

Running AI optimization for a client spending $100K/month in ads is straightforward. Running it for a client spending $10M/month is a different beast entirely:

Small scale problems:

Is the algorithm working?
Are the recommendations good?
Can we improve performance?

Large scale problems:

What happens if the system fails and burns through $500K before anyone notices?
How do we coordinate changes across 50 markets and 200 campaigns?
How do we prevent the AI from making correlated bets that create business risk?
How do we audit decisions when millions of micro-optimizations happen daily?

We had to build entirely new systems for:

Risk management and guardrails
Progressive rollouts
Automated rollbacks
Anomaly detection
Audit trails

The insight: At scale, reliability and risk management become more important than raw performance.

What's different now: This lesson still applies. At Cogny, we're building guardrails from day one—spending limits, approval workflows, change tracking—because we know our clients will scale up fast.

Lesson 8: The Best AI Teaches Your Team

What we thought: AI's job is to optimize campaigns.

What we learned: AI's real value is making your team smarter over time.

The clients who got the most value from our AI weren't necessarily the ones with the biggest budgets. They were the ones whose teams actively learned from the AI's decisions.

Netflix's marketing team would review AI recommendations even when they approved them automatically. They'd discuss patterns, test hypotheses, and gradually internalize the optimization principles.

Over time, they needed the AI less for routine decisions and more for complex edge cases. The AI had effectively trained them to think like it did for standard scenarios.

The counterintuitive result: The best outcome wasn't dependency on AI. It was a team that understood optimization deeply and used AI to push beyond human limits.

What's different now: Conversational AI makes this learning process much more natural. Instead of reverse-engineering decisions, you can just ask the AI to explain its thinking. At Cogny, we design interactions to be educational, not just transactional.

What We'd Do Differently

If I could go back and rebuild those systems with what I know now, here's what I'd change:

1. Start with conversation, not automation

We built optimization engines that took actions. We should have built assistants that suggested actions. The learning curve would have been gentler, adoption would have been faster, and trust would have been higher.

2. Invest more in explanation from day one

Every feature should have shipped with "why did you do that?" built in. We added it later as a response to client requests. It should have been core from the start.

3. Design for iteration, not perfection

We tried to build complete solutions. We should have built minimal viable AI that could be improved through use. The best insights came from production usage, not planning.

4. Focus on well-defined use cases

We tried to build general-purpose optimization platforms. We should have nailed specific, high-value use cases first, then expanded. Breadth is the enemy of depth.

5. Make the AI auditable by default

Audit trails, decision logs, and reasoning transparency should be table stakes, not nice-to-haves. At scale, you can't operate without them.

The Biggest Lesson: We Were Building the Wrong Thing

Here's the uncomfortable truth I've come to accept: most of what we built at Campanja is obsolete now.

Not because it didn't work. It worked great. But because the approach—custom ML models, hand-crafted features, narrow optimization algorithms—is being superseded by general-purpose AI that can reason, adapt, and communicate.

We spent months building audience segmentation models. Claude can do similar analysis in seconds by examining your data and describing patterns in plain language.

We spent years building campaign optimization systems with complex logic. Modern AI agents can understand your goals, analyze your data, and suggest optimizations through natural conversation.

The tools we needed five years ago aren't the tools we need now.

That's not a failure. We helped major brands leverage AI before it was accessible. We learned what works, what doesn't, and what people actually need from marketing AI.

But it means the lessons I'm most excited about aren't the ones about how to build better ML models. They're the ones about:

How to make AI trustworthy
How to keep humans engaged
How to balance automation with control
How to make complex systems understandable
How to scale AI responsibly

Those lessons are timeless. The technology changes, but the human challenges remain.

What This Means for Cogny

Everything I learned building AI for Netflix, Zalando, and others informs how we're building Cogny:

Transparency over black box optimization: Every recommendation includes full reasoning. You can challenge, question, and understand.

Conversation over automation: We're building an AI agent you talk to, not a system that acts autonomously. Humans stay in the loop.

Depth over breadth: We're focusing on doing marketing analytics and optimization exceptionally well, not trying to be everything to everyone.

Adaptation over static models: The AI continuously learns from your data and feedback, staying relevant as conditions change.

Education over dependency: We want teams to get smarter over time, not more dependent on AI.

These aren't just product principles. They're lessons learned from watching what actually works when you deploy AI at scale with real companies and real stakes.

The Honest Truth About AI in Marketing

After five years building AI platforms for major brands, here's what I believe:

AI will transform marketing. But not in the way most people think.

The transformation isn't about replacing humans with algorithms. It's about making sophisticated marketing accessible to smaller teams, faster iteration cycles, and better decision-making at every level.

The companies that win won't be the ones with the most advanced AI. They'll be the ones that best integrate AI into their team's workflow, maintain strategic thinking while automating execution, and learn faster than their competitors.

AI is a tool, not a strategy. A powerful tool, yes. But it still requires smart people making good decisions about what to optimize, who to target, and why you're in business.

Building for Netflix taught me that at scale, every decision matters. Building for Zalando taught me that explanation is as important as optimization. Building for dozens of clients taught me that there's no one-size-fits-all solution.

Now I'm building Cogny with all those lessons baked in. Not because we got everything right the first time—we definitely didn't. But because the mistakes taught us what actually matters.

And what matters is building AI that makes your team smarter, faster, and more effective. Not AI that replaces them.

---

About Tom Ström

Tom is CEO and co-founder of Cogny, where he's building AI-powered marketing automation. Previously, he co-founded Campanja, where he built AI optimization platforms for Netflix, Zalando, SAS, and other major European brands. He's spent over a decade at the intersection of marketing, data, and AI, learning what works at scale.

Want to see how we're applying these lessons?

Cogny brings AI agents into your marketing workflow with transparency, conversation, and control. We're not building a black box—we're building a teammate. Book a demo to see the difference.