The Most Honest AI Adoption Story in Gaming
GamesIndustry.biz ran a story about InnoGames using AI to keep Sunrise Village alive. Mishka Katkoff had Thomas Lehr, the director of engineering who built the system, on the podcast to go two levels deeper than the press version. What came out was more useful, more honest, and in a few places, a direct contradiction of the article that started the conversation.
The decision wasn't a pivot. It was a forced bet.
Drop the “win-win” framing. Sunrise Village had hit a ceiling. Lehr was upfront that the revenue numbers in the original piece, sourced from SensorTower, undercounted reality since they miss direct-to-consumer and browser revenue. But the actual number doesn't change the structural problem: a 25-person team on a game that had plateaued for years, with no path back to growth.
The options on the table were 1) a skeleton crew, 2) outsourcing, or 3) AI. Lehr killed the first two fast. A skeleton crew without tooling just slow-bleeds a game to death; you can't produce enough content to retain players, so you're choosing a slower version of the same outcome. Space Ape sold us the “Lean Live Ops” narrative back in 2018. That story didn’t last the test of time.
Outsourcing brought management overhead and a loss of ownership that InnoGames wasn't willing to trade for. AI was the only option that didn't involve managing the decline.
Key takeaway #1: If you’re sitting on a plateaued title, the AI question only gets interesting once you've ruled out the boring alternatives. Most studios never get that far. They jump straight to the AI pitch deck without doing the elimination work first.
25 people became 3. Nobody got laid off.
Three people now run Sunrise Village. And the other 22 didn't leave the company; they moved to a new title, Cozy Coast. In real life, revenue didn't hold flat either, despite what the article implied. It kept declining, just much more slowly than before the AI transition.
“At least it slowed the decline drastically,” Thomas Lehr said, on what AI actually did to the revenue curve
Sunrise Village is not an AI growth story. Downloads have halved. Revenue is down by a fifth. But it is a tale of profitability. The team size went from 25 to 3. The game went from deep red to green. And that green is what allows InnoGames to invest in new games.
This is the honest version of the InnoGames story: a 90% headcount reduction on one title that slowed, but didn't stop, an existing decline. That's not the same as “AI saved the game.” It's “AI bought the game a longer, cheaper death, and freed up 22 people to go build something with upside.” Both things can be true, and the second one is the better business decision.
The actual unlock was a logging exercise.
Before any AI got involved, an engineer from InnoGames' AI task force sat with the Sunrise Village team and watched how the work actually happened. The single highest-leverage discovery: a game designer was hand-typing design definitions into Confluence, exporting them to Excel, converting to CSV or JSON, then importing into Unity. Days of manual, error-prone busywork for every content update.
That's the whole insight. One observed workflow, one custom tool built in an hour, one drastic time cut. No AI lab, no platform rollout, no quarterly AI strategy memo. AI changed the cost of building the automation, not the desire to automate.
Key takeaway #2: If you're running a studio and wondering where to start, stop touring AI tools and start watching your own team work. The bottleneck is rarely creative. It's almost always some unglamorous data-shuffling task nobody bothered to fix because fixing it used to take an engineer three weeks they didn't have.
They stuck with GPT-4o for a year and a half deliberately
This is the most contrarian thing Lehr said, and it's worth sitting with. GPT-4o was, by his own description, already an outdated model when we talked. InnoGames kept using it anyway, because every model swap means re-testing an entire production pipeline from scratch. Different models, even different versions from the same vendor, behave differently enough that a swap risks the reliability you spent months earning. They only moved off 4o because OpenAI deprecated it. Not because a better model showed up.
Compare that to how InnoGames treats coding models, where they jump to the newest Anthropic release on day one. The difference: coding work tolerates variance because a human reviews every output before it ships. A content pipeline running semi-autonomously across thousands of assets does not tolerate variance. In other words, stability beats capability the moment a workflow runs without a human checking every step.
Key takeaways #3: Instead of constantly switching tools to whoever sits at the top of the LLM leaderboard, ask which parts of your pipeline can absorb model drift and which can't.
Skill erosion is real, and Lehr didn't dodge it
Thomas admitted outright that engineers, including himself, are losing the ability to hand-write code in specific languages. His framing: that's fine if the eroding skill isn't the one that matters. Knowing Rust syntax is not the same skill as knowing how to architect a system. The job shifts from laying bricks to designing the building. He's optimistic that it's a net gain for people willing to move up a level.
Where he didn't have a clean answer was the cost of that shift. He described a real, current AI fatigue: more output requires more decisions, more context-switching between agents running in parallel, and people finishing the day more exhausted, despite or because of the AI doing the typing.
Key takeaways 6: CTOs should consciously decide which skills they're letting erode and which ones they're doubling down on, because AI will make that decision for them by default, and it usually picks wrong.
The honest read on the 84%
I pushed him on the 84.2% GamesIndustry.biz stat, industry professionals saying no amount of AI-generated content is acceptable at any stage of development. I asked him directly what happens to Western game development if that sentiment holds.
His answer was blunt: studios that don't adopt will get outcompeted, full stop. AI content isn’t automatically better. But the studios using it correctly will out-iterate everyone else on cost and speed, and the market doesn't reward craft that it never gets to see.
Key takeaways 5: The real stakes of the resistance are survival. Not a culture war over authenticity. A speed and cost war that most of the resisters are not currently equipped to win.
Why this conversation matters more than the article that started it
InnoGames talking about this openly is rarer than it should be. Most studios are either overselling AI to investors, hiding it from players, or don't understand the technology well enough to discuss it past the marketing language. Lehr did neither. He corrected the press numbers on the record, admitted the parts that didn't fully work, and described a fatigue problem he's still trying to solve.
That's the bar. Not hype, not denial. A grounded account of what the technology actually does to a real pipeline, run by a real team, with real tradeoffs that don't all resolve in AI's favor.
