Four and a half days was the time from idea to live for our first Grid Pulse event in 2023. One full day went to aligning on the event framing, another to writing and editing player-facing copy, half a day to building trigger logic in remote config, and another half day to checking the economy. At launch, one-third of the event window was already gone. The next brief was half-written on my desk.
By early 2026, a standard event took under four hours. The important part is what changed, and what we deliberately left alone.
What One Event Actually Requires
Player-facing copy is not the difficult part of LiveOps event writing. It is usually the final and fastest step. The difficult part is building the brief that comes first.
A publishable brief defines the mechanic, whether that means a challenge, cooperative, timed, or solo reward structure. It also defines trigger conditions, such as a session count threshold, days since install, or calendar date; the rewards and an economy check; duration; and push notification timing. The economy check must confirm that the reward does not undercut existing premium items or flood a currency the player is already close to capped on. Each field is a decision, and each decision depends on where players are in progression at that point in the live service cycle.
With two active games and different player bases, that context varied by title. A reward that worked on Grid Pulse was not automatically right for Scout Horizon. Their progression curves and IAP conversion windows differ. Starting a brief from scratch meant holding all of that context before writing the first line of copy.
Where the Two Days Went
Our 2023 workflow was serial. I started an event idea after reviewing weekly session data. Selin reviewed its economy impact. Once she cleared the reward structure, Kerem built the trigger condition in remote config and flagged technical constraints. I wrote the player-facing copy from the finalized decisions. Burak then checked the config parameters before we pushed live.
No one was sitting idle. Each person had other work and picked up the event when it reached their queue. The handoffs added hours because the event was never anyone's only task, not because an individual step was slow.
The sequence also ran backward often. A reward could look safe alone, then conflict with progression once Kerem set the trigger thresholds. That loop was not process failure. It was the process catching a real issue. It did, however, add another day when it happened, and it happened often. Two days was the estimate when everything moved forward. Three was the more common reality.
The initial event brief was the bottleneck. Before the serial handoff could start, someone had to spend an hour or more with session data to produce a structured artifact everyone could review.
What the AI Draft Provides
Bold LiveOps reads our anonymized session event stream and produces a structured event brief draft. It proposes a mechanic type from recent session depth data, a trigger condition with reasoning, a reward structure calibrated to current progression, and first-pass player-facing copy.
The draft is not a finished event, and we do not treat it as one. It is the starting artifact, generated in minutes rather than an hour or more. A reviewer checks the recommendation against context the data cannot contain, then approves, edits, or rejects it with a note.
When the draft is reasonable, review takes one hour. When it needs changes, the changes appear as structured edits instead of a negotiation over what belongs in the brief. Reviewing is faster than generating, even when the edits are substantial. Most of the time saved comes from removing the blank-page problem.
The Approval Step We Left Unchanged
Every event on our games receives human sign-off before publishing. This is not a temporary safeguard that we plan to remove as the system matures. It is the design we selected from the beginning.
The draft reports what the session data suggests. A human decides whether that suggestion fits this particular week and everything happening around the game. A technically sound draft can still be declined because of an outside-the-game community conversation, a new content release that would split attention, or a gut read that players saw the mechanic two weeks ago and need variety. Session frequency data contains none of those signals.
We use approval first because the judgment layer should be explicit and attributed, not because we distrust draft quality. Whoever approves an event owns its outcome. That accountability matters to a four-person team.
What Changed in Output Volume
During the six months after full activation of AI event drafting on Grid Pulse and Scout Horizon, cadence rose from approximately two events per game per month to an average of 1.8 events per game per week. A four-person team could not sustain that output manually. We would have needed a dedicated LiveOps producer, which was not realistic at our stage.
The increase matters because event frequency directly affects mid-term retention. Internal Bold Games measurement across both live titles showed that players encountering an active event within their first ten days had materially higher session frequency over the following two weeks than players who did not. At two events per month, engaged daily players faced long stretches without a new reason to return. Those gaps affected day-14 and day-21 retention numbers, which later compound into IAP conversion.
What This Does Not Replace
We are not saying AI drafting removes LiveOps judgment. We are saying it removes the mechanical starting work. A brief can fit the session data and still be wrong for a particular point in a game's lifecycle. For example, recommending a cosmetic reward event during a week with a promotional discount could send a mixed-value signal to players deciding whether to spend. The data does not know about that offer. We do.
The change from two days to under four hours came from automating session cohort pattern extraction, structured brief construction, and first-pass copy. Judgment still requires the same time because it depends on information outside the data. It simply no longer comes after two days of mechanical preparation.