Welcome everyone. I'm Bijan, I run Gemma Analytics — we're a data and AI consultancy in Berlin. And this is Torben, who heads Data & AI at Natsana. We've worked together since Natsana's early days. Today we want to share what we've learned about using data at different startup stages — and how AI changes the game in 2026.

Quick intros. Gemma is a data and AI consultancy — about 20 people in Berlin, over 70 projects since 2020. We build the full data stack for companies at different stages. Natsana is one of our longest-standing clients — Torben, want to give the quick pitch?

Here's how we'll spend the next hour. First, a compact 20-minute presentation to give you a framework for thinking about data and AI at your stage. Then — and this is the part we're most looking forward to — an interactive session where we discuss your specific questions and challenges. But first: grab your phones. There's a QR code coming up.

Take 30 seconds. Three quick questions. We'll look at the results together and use them to tailor the discussion. No wrong answers — we just want to know where everyone is starting from.

Alright, let's talk about data. Almost every startup begins with spreadsheets. And that's fine — actually, that's exactly right. The question is: what comes next? And more importantly: when?

Spreadsheets are great — until they aren't. We see the same pattern everywhere. It starts fine. Then versions diverge. Then different teams have different numbers. And eventually something breaks silently and you make decisions on wrong data for weeks. The point isn't to avoid spreadsheets. The point is to know when you've outgrown them.

Here's a rough framework. Don't obsess over the exact stage boundaries — they're fuzzy. The key insight: each stage is triggered by a pain point, not by a funding round. Stage 1 is fine as long as everyone agrees on the numbers. You move to Stage 2 when different people start having different versions of the truth. Stage 3 is about combining data from multiple sources. Stage 4 is when data becomes a real competitive advantage — predictions, AI, the works.

This is the modern data stack in 2026. Data flows from your sources — Shopify, Stripe, whatever you use — into a central warehouse. It gets cleaned and modeled. And then people consume it through dashboards, alerts, AI assistants, or custom apps. The key message: you don't need to build all of this at once. Start with what hurts most and expand from there.

Three patterns we see all the time. Some founders invest too early — hiring a data engineer when they have 10 people and no product-market fit. Some invest too late — they find out their numbers are wrong during a board meeting. And some invest wrong — over-engineering, building it all internally, or treating data as a one-time project. The sweet spot? Start when decisions genuinely depend on data accuracy. Start simple. And think of it as an ongoing capability, not a project.

Now Torben is going to walk us through what actually happened at Natsana — not the theory, but the reality. What did you start with? What worked? What would you do differently?

[TORBEN'S NOTES: Walk through the Natsana data journey chronologically. Hit the key decision points — when you moved from spreadsheets, when you brought in Gemma, what the setup looks like now.]

[TORBEN'S NOTES: Be honest. The audience will learn more from your mistakes than your successes. The Power BI story is a great one — management went back to Excel because they knew it better.]

[TORBEN'S NOTES: This is the slide people will remember most. One or two things you'd definitely do differently, and one thing that was absolutely the right call. End with a punchy quote.]

Now the part everyone's been waiting for. AI. But we're not going to give you the hype version. We'll show you three concrete categories of how AI actually helps with data — from the boring-but-valuable to the genuinely new.

Three categories. Category 1: doing the same things faster. AI coding assistants make building pipelines, writing SQL, and setting up infrastructure 2 to 5 times faster. That's useful but not transformative. Category 2: the real unlock. Because development is so fast now, things that were "too expensive" are suddenly viable. Custom dashboards, one-off internal tools, bespoke analyses. Category 3: AI as part of the solution itself. Chat with your data, AI-generated insights, automated reporting. This is the future — but it requires good data infrastructure underneath.

Concrete numbers from our experience. Setting up a new data pipeline used to take 2-3 days. With AI coding assistants, we do it in a few hours. Complex SQL that used to take hours of iteration — the AI gets you 80% of the way in minutes. This is Category 1: pure acceleration. It's not glamorous, but it dramatically lowers the cost of building data infrastructure.

This is the category that excites us most. Because AI makes development so fast, things that were "not worth building" are suddenly feasible. Need a quick internal tool for a specific workflow? Build it in an afternoon with Streamlit or Lovable. Have a one-off analysis question? Ask an AI assistant — it takes 30 minutes instead of a day. Want to automate that weekly manual process? The ROI calculation just changed completely.

Category 3: AI is in the solution itself. Chat with your data — ask questions in plain language and get answers from your warehouse. AI-generated insights — automated summaries that tell you what changed and why. Smart document processing — AI reads invoices, contracts, or tenders and extracts what you need. But here's the catch, and it's important: all of these work best when your underlying data is clean and centralized. Put AI on top of messy data and you get confidently wrong answers. That's why the earlier stages matter.

[TORBEN'S NOTES: This is the crown jewel story. The Power BI dashboard nobody used → custom AI-enhanced dashboard that management actually adopted. Walk through the before and after. If you can show a screenshot with anonymized numbers, that would be incredibly powerful.]

[TORBEN'S NOTES: These can be ambitious — if you're prototyping something, say "we're building X." As long as there's a good feeling about where it's heading. 2-3 examples of high-value AI use cases you're currently working on.]

Practical advice. Step 1: get your data house in order. Not perfect — just clean enough and in one place. Step 2: use AI to build that infrastructure faster. Step 3: look for those Category 2 wins — what's a manual process that could be automated in an afternoon? Step 4: experiment with AI-in-the-product, but start small. And the golden rule: start with a business problem, not with a technology.

Let me wrap up with five things to take away from today.

Five takeaways. One: spreadsheets are fine until they break — learn to recognize the signals. Two: always start with the pain point, not the technology. Three: AI fundamentally changes what's economically viable — things that were too expensive are now possible. Four: clean data is non-negotiable if you want AI to work. Five: don't overthink it. Start, iterate, improve.

Alright, that's the end of our prepared part. Now it's your turn. We're here for 40 more minutes and we want to hear from you. What's your biggest challenge? Where are you stuck? What excites you? No question is too basic — Torben brings the in-house perspective, I bring the consultant perspective. Let's go.