Persona5: The Phantom X (P5X) is the first mobile and PC RPG in the Persona series. Perfect World and SEGA co-operate it, Atlus supervises it, and it is available in Japan, across Asia and in global markets. Its international project team works across languages.
Two SEGA teams handle its data. Yusaku Abe's Business Analytics Section works with each project to help producers and decision-makers act on data, and Abe mainly covers mobile. Kazutaka Yorozu's Data Engineering Section builds and runs SEGA's data platforms for mobile, PC online and console games. Each team needed to solve a different problem.
The problems each team brought
Each follow-up could mean new SQL
Abe's section was formed by merging the team that handled online PC titles with the team that handled mobile, so it inherited more than one BI tool and analytics environment. With the BI tools the team used before, the typical flow was to write SQL to aggregate the data and then build a dashboard on the result. Analysts and engineers could write that SQL, and many live-ops members could not, so questions from live-ops started as requests to an analyst.
After analysts answered one request, live-ops members often asked to see it "on this axis" or "split by segment," and a new angle could mean another query. Changing the grain could become its own job, because counting unique users by day and by month sometimes took separate SQL.
With SQL, we sometimes had to rebuild everything from scratch each time, which was a heavy load for analysts.
The pressure came from the pace of live ops. After an update, the team is sometimes asked how it performed by the next day, and it needs a clear read by the start of the following week to plan the next move. When analysts had to put immediate data pulls first, the deeper investigations that improve the game sometimes had to wait.
A translated dashboard that arrived days late
P5X is run by a multinational team, and dashboards built in Japan were written in Japanese. Before overseas colleagues could use one, it had to be translated, and the translated version sometimes arrived three days or a week later. By then the state of the game might already have changed. Abe also points to a staffing problem. Very few people combine analytics expertise with strong language skills. A translator without an analytics background can lose the meaning, and asking analysts to translate costs time the team does not have.
Manual deletion was a heavy burden
Yorozu's team started from a different place. SEGA already ran a shared Databricks platform holding data for all of its mobile games, with a mechanism for aggregating data across titles, and the plan was to bring P5X into that setup. For P5X's global simultaneous launch, SEGA required a way to delete the relevant data, such as personally related information, whenever a player submitted a deletion request. Doing that by hand requires a great deal of care and places a heavy burden on the person responsible, and Yorozu wanted it automated.
Why SEGA chose Agentic Engine
Yorozu first tried Agentic Engine more than two years earlier. Game KPIs were available out of the box, and Yorozu could build charts right away without instruction, but SEGA had already adopted a different tool and passed. P5X's producer at the time reopened the evaluation after seeing the product and asking to hear more.
The existing tools could have covered P5X without much additional cost, so signing an additional contract needed a real reason. Yorozu's reason was the operations side. If live-ops members could check and visualize the data they needed on their own, analysts could spend their time on more advanced analysis. Abe's deciding factors were multilingual dashboards, the ability to look up a single player for customer support, and analysis models built for games, such as Events Analysis, Funnel Analysis and Retention Analysis.
How the rollout ran
Agentic Engine runs on servers SEGA provides. SEGA estimated the expected number of players and the log volume the game would generate and shared both with ThinkingAI. ThinkingAI proposed a server configuration, SEGA reviewed it, and the two sides worked through anything missing or concerning as they went. SEGA assigned one infrastructure-leaning member to server setup and one engineering-leaning member from its application side. After SEGA provided the servers, ThinkingAI built the Agentic Engine application and took responsibility for operating it.
SEGA's part took one to two months. Today Yorozu's team receives information from ThinkingAI, reviews it and makes decisions, and the internal effort is close to zero.
How the two teams work now
Analytics builds in models and lets live-ops change them
Abe's team now considers an analysis model first and uses SQL for work such as complex processing or checking one player's situation in detail. Analysts can switch an analysis model from daily to monthly or a period total in one click. A model an analyst sets up also becomes a sample that live-ops members open and explore from on their own.
Analysts define player segments with tags. Analysts tag the players in a segment, then filter or group dashboards and reports by that tag. Abe values that tag states update to each player's latest value while past states stay available, so the team can compare a segment at an earlier update with how it looks now. For new players, the team runs Funnel Analysis split by platform and by country to find where players get stuck. For player support, User Behavior Sequence shows what one player did before and after an issue, which Abe says would otherwise have cost about as much as building an entire log-analysis environment. Dashboards are still built in Japanese, and AI Translation shows user-entered content such as comments and report names in each viewer's language.
Data engineering uses the data where it already lives
The P5X environment is connected to SEGA's shared Databricks platform, so SEGA uses the data without moving it. Using the data in place avoids duplicate storage and its costs, and the team retrieves key P5X KPIs for its existing cross-title aggregation. Deletion of the relevant data is automated, so the team can act on a player's deletion request without doing it by hand.
The results
Live-ops members now check data and report KPIs in meetings themselves, a task analysts often handled before. Abe says many of the project's key members, from live-ops directors and leads to project owners, look at Agentic Engine every day, and analysts bring deeper investigations when a question needs one. The team has also become much faster at pinpointing problems, because funnels can be compared by platform and by country without rewriting a query. Abe also says ThinkingAI completed one requested product update in about two to three months.
On the platform side, ongoing effort is close to zero. With deletion handled automatically, Yorozu says the range of data the team can work with has expanded.
I'm very grateful that they automated such a sensitive task that nobody wants to do.
What the teams want next
Both teams want more from AI. Yorozu wants to ask questions and get answers in natural language, along with an official way to query Agentic Engine data from outside its interface, for example through something like MCP. Abe hopes for AI-driven summaries, insights and alerts, and for a dictionary in AI Translation so that game-specific terms and names stay untranslated.
Adapted from two interviews first published in Japanese on ThinkingAI's site, with Yusaku Abe and with Kazutaka Yorozu.
© ThinkingAI · Persona5: The Phantom X ©Perfect World Adapted from Persona5 ©ATLUS ©SEGA
“Operating Agentic Engine takes almost no effort.”