The number that explains why this exists
Testing every combination of maps and modes in Battlefield V by hand would take roughly half a million hours, about 300 work-years, by EA's own estimate. That's the figure EA's research division, SEED, cites to explain why it started building reinforcement-learning (RL) bots for testing in the first place. It's a useful starting point, because it frames these tools correctly: not as a replacement for judgment, but as a way to brute-force coverage that no studio could staff by hand.
EA's bots have been in real AAA titles for years
SEED has deployed RL testing agents in actual production titles, including Battlefield V and the Dead Space (2023) remake. In Battlefield 2042, the team used curve-based pathing to improve AI-driven vehicles, cutting mean total path time by 39% and reducing stuck incidents. Separately, SEED found that imitation-learning agents, trained by watching human demonstrations, reached similar success rates to RL agents in 20 minutes of training versus 5 hours. A more advanced version, called CCPT, blends curiosity with imitation learning to explore 3D spaces near a demonstrated path and has found glitches that baseline bots missed entirely. None of this replaces scripted bots, which SEED says still carry most of AAA testing; RL requires deep machine learning expertise and training runs that can take days.
Ubisoft's bug-predictor predates the AI boom
Long before generative AI, Ubisoft's research group La Forge built Commit Assistant, later renamed Clever-Commit, which studies a decade of historical bug patterns to flag risky new code commits before they're merged. Ubisoft estimated the tool could free up to 20% of a programmer's time by catching regressions early. It was detailed publicly in February 2019, around when Mozilla partnered with Ubisoft to bring the tool into Firefox's own development workflow, aiming to catch an estimated three to four out of five bugs before they shipped, while expanding its language support to Rust, C++, and JavaScript.
Venture capital is betting the same way
Big publishers aren't the only ones betting on this. Copenhagen-based modl.ai raised an €8.5 million Series A in November 2022, led by Griffin Gaming Partners with Microsoft's own M12 venture fund participating, specifically to build AI bots for automated playtesting and debugging. Its early client list included King, N3twork, and Nitro.
What none of this replaces
The pattern across all of these tools is narrower than the "AI will test your game" pitch suggests. They're good at pathing, regression risk-scoring, and overnight soak-testing of mechanical systems, the kind of coverage problem that really is a math problem. None of them are built to judge whether a mechanic feels fun, whether a new player finds the UI confusing, or whether a bug is one a real stranger will actually hit and care about. Even EA's own research frames RL as something that augments scripted and manual testing, not something that replaces it. That's not a knock on the tools. It's just an honest read of what they're actually for.
Coverage math is solvable with bots. Knowing what a bug actually feels like to a player still isn't. IndieBug covers that half.