Ubisoft just gave AI NPCs their own company. We think the useful lesson for indies is smaller than the headline. The version you can ship this year doesn’t need a studio. It needs a short list of moves you wrote yourself, and a model that only gets to pick one.
What Ubisoft announced
Ubisoft’s first half earnings release on November 19 named Motherbrain as “a dedicated subsidiary focused on AI for gameplay innovation.” Its current focus is “smart NPCs with memory and AI game masters that manage large-scale systemic worlds.” It aims to serve creators “within Ubisoft’s Creative House ecosystem and beyond,” and to partner with “selected technology actors.”
Motherbrain is the team behind Teammates, which Ubisoft calls its first playable generative AI experience. The name had already surfaced in a trademark filing in September. The work goes back further, to an effort called Project NEO NPC that gave characters a backstory and let them answer players like a chatbot.
The same day, Ubisoft confirmed it was laying off 166 workers at its Montreal and Quebec City studios. Headcount fell by close to 1,100 over the half. Ubisoft had about 15,500 people on September 30, down from 20,700 in September 2022. Those two facts will be read together, whether Ubisoft likes it or not.
What Teammates actually does
Ubisoft revealed Teammates on November 21, 2025. It’s a short first person shooter level with a voice assistant named Jaspar and two squadmates, Sofia and Pablo. You speak to them. Jaspar can highlight enemies, fill in story details or pause the game, and the squad takes spoken orders about cover and attacks. A few hundred players tried it in a closed playtest.
The most telling detail comes from narrative director Virginie Mosser. “Instead of fixed lines, we create these kinds of fences that let NPCs improvise within the world but stay within the boundaries of the lore and motivations we have given them,” she said. That’s the whole design problem in one sentence. The writing doesn’t disappear. It turns into rules.
The cheap version already won a prize
This week, at the Unity AI Game Jam in Busan, a team called We from Another Star won the OpenAI Special Prize with Dungeon Tycoon. Players build a dungeon, and a party of AI heroes argues tactics and fights the Demon Lord on their own judgment.
The trick is almost parsimonious. The team pre-built every action a hero could take, and the OpenAI API only chose which one to run. The whole jam cost them $0.26 in API fees out of a $500 testing credit. Most of us have spent more than that on a vending machine snack we regretted.
That’s the pattern we’d copy. The model doesn’t write code, move characters or invent abilities. It reads a small slice of game state and returns a choice from a fixed menu. Your code still runs the animation, the damage and the cooldowns, so a bad answer can only ever be a legal move.
The middle version is small models and reflexes
The other data point arrived on November 18. Will Wright’s studio Gallium picked Artificial Agency’s Behavior Engine for MyMems, a social simulation built from players’ memories. Gallium is targeting a live community release in 2027.
Artificial Agency’s pitch is about economics as much as charm. It combines small, game specific models with a reflex layer that handles immediate reactions, and it saves deeper model reasoning for decisions that need it. The stated goal is lower latency and lower inference costs. Put simply, the character reacts instantly and thinks only when it matters.
You can get close to that today without a partnership. Plain state machines or behavior trees can handle the reflexes. The language model only steps in for the rare moments where a scripted reply would feel wooden, like a merchant remembering you haggled last time.
Where it goes wrong
The backlash is not hypothetical. Six European game worker unions, including Britain’s IWGB and France’s STJV, published a manifesto urging players to “support games made without GenAI.” In the GDC survey, 55 percent of respondents said generative AI is having a negative effect on the industry. We covered that mood in our post on the GDC survey.
Steam adds paperwork. Valve calls AI content created while the game runs “Live-Generated,” and it asks what guardrails stop that AI from generating illegal content. If the AI service costs money per request, Valve expects you to manage player access and charge through a Steam supported method, such as microtransactions, a subscription or a DLC unlock.
Then there’s optics. A studio that cuts staff and opens an AI subsidiary the same week invites an uncharitable reading. An indie team doesn’t have Ubisoft’s scale, but it does have a store page and a comments section. If players think your talking NPC replaced a writer, the feature becomes the story. Our post on Big Walk’s AI disclosure shows how fast that conversation moves.
What we’d build
Our view is that the garrulous, open ended AI NPC is still a big studio luxury, and a mercurial one. The menu approach is a design choice, not a compromise. It keeps the game readable, keeps the bill tiny and keeps your writer in charge of what characters can do.
If you want middleware instead of rolling your own, our AI NPC dialogue tools guide compares Convai, Inworld, Neocortex, Tryll and Lingotion on free plans, costs and commercial rights. Tryll Engine runs the model on the player’s PC, so there’s no per request server bill. Neocortex lets NPC actions call your own code, which fits the menu approach neatly.
Tips for shipping a cheap AI NPC
- Write the moves first. List every action an NPC can take, in code, before you add any model.
- Make the model choose, not create. Ask it to return one action ID from your list, and reject anything else.
- Send less state. Pass only what the decision needs, like health, distance and the last two player actions.
- Keep reflexes scripted. Dodging, fleeing and greetings belong in a state machine, not a prompt.
- Cache and limit calls. Decide on a timer or on key events, never every frame.
- Write the fences. Give each character lore, motives and things they will never say, then test against them.
- Plan an offline fallback. If the API fails, the NPC should pick a sensible default and keep playing.
- Track cost per session. Log tokens per playtest so you know your real number before launch.
- Fill in Steam’s survey honestly. Mark live generated content and describe your guardrails in plain words.
- Say it on your store page. One short line about what the AI does, and what people wrote, heads off most guessing.
