Live-ops AI: buy the plumbing, build the judgment

Your live-ops calendar is fuller than the team can staff, and the vendors have noticed. Metaplay ships an agent that writes code and queries the live game. devtodev's analytics comes with an AI analyst. AWS published reference code for agentic live ops and walked studios through it at GDC. The question landing on your desk is whether to buy one of these or build your own.

The team you would hand it to is not asking for any of it. In GDC's 2026 survey of more than 2,300 professionals, 52% said generative AI is harming the industry, and the share who said it helps fell to 7% from 13% a year earlier. Whatever you buy lands on people who expect it to make their work worse. That narrows the answer before price comes into it. AI belongs behind the game, in the pipeline that decides what ships, where nobody's craft is being replaced.

Split the decision the way the work splits. Half of live ops is plumbing: storing config, shipping it to the right players, targeting segments, running the A/B test, rolling back an event that misbehaves. The other half is judgment: which event to run, for which players, at what price, and whether the economy can take it.

Buy the plumbing. It is proven in shipped games, and the prices are public. As listed this month, Balancy starts at $499 a project plus a fee per thousand daily players, Heroic Labs prices its Satori live-ops product from $600 a month, and Metaplay runs from €195 to €1,985 a month before usage. A studio that builds its own config delivery and A/B engine is paying engineers to rebuild something it could rent, and it starts years behind on production hours. If one of these fits your backend and your data can reach it, buy it.

Build the judgment, because nobody can sell it to you. A platform will run any event you hand it. It cannot tell you which event your players will pay for without leaving, because that answer lives in your own history: what shipped, and who churned after it. When we compared the two platforms one studio was weighing this summer, neither simulated an event or predicted its lifetime value before it shipped. Balancy's AI was in closed beta. PlayFab's Copilot, announced at GDC 2024, showed no sign of leaving preview. Both ship what you configure and report on it afterwards.

That half is what we built for a top-20 mobile games studio. Every event is simulated against 25 player baselines, five stages of the game by five spend tiers, before anyone sees it. Difficulty is calibrated in a loop instead of guessed. A value model ranks candidate events in close to the order their revenue lift came in, a rank correlation of 0.77 against events that shipped, and the studio uses it as a ranking, never as a dollar forecast. Gates check every design against the economy's hard rules and report what they could not check, so a skipped check never passes for a clean one.

We built that studio a ship layer too, because they wanted every change to pass a prediction and a gate on its way to players. It is the half we would talk most studios out of building.

The cheap version of the judgment half is a language model with a good prompt. Early in that build, a model's estimate of player strength came back off by 30 to 100 times. We replaced the estimate with measured player data, kept the deterministic parts deterministic, and gave the model the jobs it does well: drafting the brief, reading the readout, writing the patch notes. Then we put a gate between all of it and the players, with a person approving what goes live.

Check the data before you sign anything, whichever way you go. At one studio we re-read 158 finished A/B tests straight from the experiment console's own results, and in 81 of the 131 filed as no result, a metric had moved. A platform's AI reads the same data your team filed. If the filing was wrong, faster tooling ships the wrong answer sooner.

Buy when a platform fits your backend and the problem is shipping faster. Build when the decision depends on your own history, or when the AI has to answer to your economy's rules before anything reaches a player. Most studios need both: a platform for the plumbing, and a layer of their own that decides what to ship and checks it on the way out.

We give that answer on the first call. When it is a platform, it costs us the build, and we will say which one. When it is a build, it is the AI Build Program, and the live-ops tooling is what one looks like finished. What changed in a live-ops team once the tooling landed is in Live ops after AI.