AI’S HOURS-SAVED PITCH IS UNDERSELLING THE REAL
CHANGE TO PRODUCT WORK By Lukas Offerhaus, Product Head for Heroes of History at InnoGames
A
I now automates most of my routine writing: the weekly management reports, project
updates for Slack, Jira bug tickets built out of player reports, and more. These wins save me a few hours a week, but if I had stopped there, I would have missed out on AI’s biggest unlocks. At Heroes of History, we are currently testing a rework of the first 10 minutes
of gameplay. To get a read on the first week of data, I prompted my AI agent: “How is our new tutorial performing?”. The result, in the form of a visual web page: a 51-step funnel with the drop-off rate for every step, the median seconds players spend on it, and the drop-off rate per second. Two steps account for more than 40% of churned players. Another step only lasts for two seconds, but 8% of new players never complete it. “Would you like me to break it down by traffic source next?”, the agent asked. No data analyst ticket, no queue, no waiting. These types of results are why I no longer treat AI as something I
consult occasionally in a browser tab. I now use it as the hub where all my work starts. Some tasks it takes over with minimal steering. For others, it assembles the context I need to do the job myself, or it structures a problem I don’t yet know how to approach. The mechanism is unglamorous: richer context in, better output out. Most disappointing AI results get blamed on the model when they’re in fact context problems. Claude Code, the agent I work
with, is wired into our data warehouse, Jira, and Discord, plus a dozen other systems, and it reads a set of local documents describing our economy, design, and live operations. The magic isn’t clever prompting (even though that helps, too); it’s plumbing - setting up your workspace to have the richest, most accurate context possible. While text output has been useful, being able to build throwaway
web pages, tools, and prototypes has been an even bigger win for me. Our marketing growth model is an interactive web page of scenarios that lets me tweak assumptions and see the model react live. Tools like this take minutes to build. Nobody maintains them, and most are dead by the end of the meeting they were built for, but they turn an abstract argument into something that can be understood intuitively. AI fetches and synthesizes well, but despite extensive
experimentation, I have not been able to get it to reliably interpret and judge (yet?). Ask how many players finished our last event and the answer is accurate and reliable. Ask what our retention curve means for the roadmap, and it comes back plausible and confidently argued, but completely untrustworthy. A few common best practices help you see through the AI’s misplaced confidence: word your prompts neutrally; challenge what comes back; ask for the opposing case; compare approaches; talk to your real colleagues. There’s a lot of talk about AI trivialising our work. I’m
observing the opposite: I move through more work per day, I operate with more complete information, and I spend more of my time on complex and high-impact problems that require genuine product judgment.
September/October 2026 MCV/DEVELOP | 19
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52
