AI can generate the first 70% of a playable. The final 30% decides performance
AI playable production compresses the blank page fast, but pacing, feedback and fail states still decide whether a build actually converts.
AI playable production is rapidly improving the blank-page problem. It can draft assets, code common interactions, and translate a reference into a functioning prototype in a fraction of the time a manual build once took. At GDC this week, where AI-assisted production has been the dominant conversation track across sessions on procedural content and live-ops tooling, that compression was treated less as a novelty and more as a baseline expectation. That is meaningful. It is not the finish line.
Playable generation gets you a working build, not a working ad
A tool can now produce a playable that runs, responds to input, and resembles the reference brief. What it cannot yet reliably produce is the specific sequencing that makes a build persuasive rather than merely functional: the exact beat at which the first meaningful choice arrives, how quickly failure teaches the mechanic instead of just registering as a loss, and whether the call to action lands at a moment the player actually wants it rather than a moment the template scheduled it. Those decisions sit in the roughly 30% of the work that generation does not do for you, and it is disproportionately where performance is won or lost.
The number is deliberately rough. No vendor has published a rigorous measure of what share of a playable's build effort AI tooling now handles, and the true figure will vary by genre and studio. The useful claim is directional: the easy majority of the mechanical work is compressing fast, while the small remainder that decides whether the build actually persuades anyone is compressing much more slowly, if at all.
The remaining work is experiential
Performance depends on details that are easy to under-specify in a brief and easy for a generation tool to guess wrong: when the player first acts, whether success feels authored rather than automatic, how the game recovers from confusion instead of simply resetting, and whether the CTA arrives at a moment of earned desire. A generated prototype will usually get a plausible version of each of these. Plausible is not the same as tuned, and the gap between the two is exactly the gap that separates a build that tests adequately from one that becomes a durable concept family.
Use AI to widen the search, not to multiply cosmetics
The greatest value from generation tools comes from testing more distinct hypotheses, not from generating dozens of cosmetic siblings off one working build. Teams should spend the production time they save on better observation of how each variant actually plays, sharper variant design that isolates a single mechanic change, and harder editorial choices about which concepts deserve a human pass before they go anywhere near a media budget. Used this way, generation raises the number of real experiments a team can run in a quarter. Used the other way, it just raises the number of files.
What GDC's AI conversation is missing
Much of this week's GDC discussion has framed AI production as a speed story, and speed is real. What gets less airtime is that speed without a strengthened editorial layer just moves the bottleneck from production to review, since someone still has to judge which of twenty AI-assisted builds is worth a media test. Studios investing in generation tools without also investing in the judgment that decides what to do with the output are likely to end up with faster pipelines and the same win rate they had before.
That judgment layer does not need to be large to be effective. A single experienced reviewer who can watch a generated build for two minutes and identify exactly which beat is off, an opening that arrives a second too late, a fail state that punishes rather than teaches, will catch more of the value in the remaining 30% than a larger team without that specific skill. The scarce resource in an AI-accelerated pipeline is not headcount. It is the trained eye that can tell a working interaction from a persuasive one on first viewing.
The UA Ledger view
Automation raises the floor. It does not remove the need for a creative director who can feel the difference between a working interaction and a persuasive one, and that distinction is exactly where the value of a human editorial pass still lives, however good the first draft gets.
