A group of former Google DeepMind researchers has designed an AI behavior engine, aiming to transform the basic nature of video games by making the behavior and interaction mode of non-playable characters, or NPCs, more dynamic.

There's no shortage of companies using AI to generate NPCs that feel more realistic, but Canada-based Artificial Agency, fresh out of stealth with $16 million in funding, is betting its behavior engine will help it stand out from the pack.

Traditionally, NPCs are guided by decision trees and pre-written scripts, which often limit the number of outcomes a player can experience. For example, most NPCs in games respond to player behavior with a few repetitive dialogues, which can often feel unrealistic and boring.

Artificial Agency's behavior engine tosses this framework out the window, making the game developer more of a stage manager. The engine makes developers give each NPC a set of motivations, rules, and goals to live by; this then dictates how the NPC will behave in response to the player. It can be plugged into existing video games or be the basis for completely new ones.

A startup originating from Edmonton, Alberta, is venturing into an increasingly crowded space. Its competitors include Inworld, which also offers AI-generated behaviors for NPCs, as well as Nvidia, which has been working on AI-powered NPCs for some time now.

Artificial Agency believes, on the other hand, that the way forward lies through integrating AI-generated NPCs with a video game's design.

"The conversations we have with these studios often aren't if, but when," co-founder and CEO Brian Tanner told TechCrunch. "This type of dynamic interaction and dynamic response that our system allows is going to be table stakes in the games industry just a few years from now."

The startup recently raised $12 million in a seed round co-led by Radical Ventures and Toyota Ventures, the founders told TechCrunch. It had previously raised $4 million in an undisclosed pre-seed round from Radical Ventures, bringing its total raised to $16 million. Other participants in the latest seed round were Flying Fish, Kaya, BDC Deep Tech and TIRTA Ventures.

Who wants AI NPCs?
Whether gaming studios will even adopt their AI technology is a huge question for many of these startups. Some worry that the big studios will develop the technology themselves, or be slow to add generative AI to their flagship games, in fear of hallucinations and how untested the technology still is.

While it wouldn't name them, Artificial Agency says it's working with "several notable AAA studios" to develop its behavior engine, and expects the technology to be widely available in 2025.

"When we were reaching out to game studios, some of them were starting to build some of these behaviors themselves, when in reality, they're just trying to build games," said Daniel Mulet, an investor with Radical Ventures. "Once you see like 20, 30 groups that are trying to build this themselves, there is an opportunity to build a platform and make it available to everyone."

Generally, game developers seem receptive to the use of generative AI in building games but are still a bit hesitant. Almost half of the 3,000 game developers surveyed by GDC and Game Developer for the 2024 State of the Game Industry report stated they employ generative AI in some aspect of their development process. This applies more often to repetitive tasks. However, 42% of respondents expressed concern over the ethics in the use of generative AI. Still, only about 21% of those surveyed expect that generative AI will have a positive impact on the industry.

Artificial Agency's founding team, with decades of experience in Google DeepMind, gave Mulet confidence that it can build a best-in-class tooling layer to improve how NPCs behave. DeepMind, after all, has a long history of developing the cutting edge in AI that can play games - it built AlphaGo, the first computer program to beat a world champion at Go.

At approximately that time, Tanner and his team found it's time to leave and start developing video game agents which could replace NPCs.

From NPC to co-op companion
In a demo of the technology the startup shared with TechCrunch, co-founder Alex Kearney created an NPC powered by the behavior engine in Minecraft (the startup would not reveal the games it's currently working on). The NPC, named Aaron, was instructed to be friendly and helpful, and was endowed with basic functions such as movement, opening chests, digging and placing blocks.

One time, Kearney's in-game character asked Aaron to fetch supplies for a terrifying mining expedition. Unscripted, the NPC went into a number of chests and retrieved armor, helmets, tools, and foodstuffs, bringing the supplies back to Kearney's character. And when Kearney asked Aaron if it had also brought back some bread, since it knew she was gluten-free, the NPC apologized and instead brought back cooked chicken.

The simple demo showed how the AI NPCs in Artificial Agency could not only talk but do complex things without being specifically instructed to do so. Aaron showed at least some level of awareness, and the NPC created a unique experience without ever writing a script or programming anything. At least, perhaps the technology will save game developers some time.

Will gamers pay the price for AI?
He estimated this five-minute demo would run about $1 in AI inferencing costs, but Tanner noted that it would have run $100 a year ago. Artificial Agency expects costs to continue coming down, both thanks to improvement in GPU efficiencies and AI model optimizations. It currently is using open source models, including Meta's Llama 3. A year from now, Tanner expects the five-minute demo to run one cent or less.

But whether it costs a penny or $100, who’s going to end up paying for these inferencing costs? Artificial Agency says AI NPCs probably won’t make video games more expensive for an end user, but Radical Ventures’ Mulet wasn’t so sure. He said his venture firm is confident game studios are willing to pay to license Artificial Agency’s technology, but once it’s deployed, it could result in a monthly fee for gamers.

"The fact that there's inference costs associated with running these systems means that it has to be a bit of a premium feature," said Mulet. "Will you, as a gamer, pay $2.99 a month or $12.99 a month? That's a little bit early to tell."