Case Study: How an SLG Publisher Saved $1M a Year with AI Localization
An SLG publisher with $500M+ in annual revenue handed 80% of its translation workload to AI. Six months later: annual costs down 59%, turnaround 75% faster.
An SLG publisher with $500M+ in annual revenue, shipping across 12 languages and 30+ countries, ran a 14-person localization team that juggled six translation vendors.
They decided to hand 80% of their translation workload to AI. Six months later—
| Metric | Before | After | Change |
|---|---|---|---|
| Annual translation cost | $1.8M | $740K | -59% |
| Version update turnaround | 5-7 days | 1.5 days | -75% |
| Translation quality score | 3.8/5 | 4.3/5 | +13% |
| Post-launch translation bugs | 35/build | 8/build | -77% |
This article breaks down their entire transformation, step by step. For the underlying numbers behind these savings, see our game localization cost comparison.
Why Change
Company G's localization workflow (anonymized at the client's request) was the industry "standard setup"—designers write the text, the PM packages it, distributes it to six vendors, collects it back, runs QA, fixes bugs, and ships.
Of the 14-person team, four people worked full-time just "chasing vendor progress" and "handling delivery quality issues." At a retrospective, the CEO asked one question: "Is the localization team creating value, or just managing chaos?"
The root causes:
- Terminology inconsistency was a systemic problem. 28,000 glossary terms were scattered across the TM systems of six different vendors, with 1,200+ inconsistencies across vendors and 340+ conflicts within a single language.
- Translation quality came down to luck of the draw. Different builds in the same language were assigned to different translators, producing wild quality swings.
- Translation was the bottleneck for version updates. Development, QA, and operations all sat waiting on the 5-7 day translation cycle.
The Transformation Path: Four Phases
Phase 1: Assessment (Weeks 1-2)
A content-classification audit found that 83% of the text could be AI-driven:
- System prompts / UI copy (40%) → fully AI-suitable
- Routine NPC dialogue (28%) → AI + spot checks
- Live-ops event copy (15%) → AI + review
- Main storyline (12%) → AI first draft + full review
- Marketing materials (5%) → AI-assisted + human transcreation
In parallel, they cleaned up the glossary—importing, deduplicating, and detecting conflicts. This single step alone explained why players kept complaining that "the same thing has different names on different screens."
Phase 2: Pilot (Weeks 3-6)
They chose Japanese as the pilot language—because it's the hardest (honorific system, kanji readings, length control, cultural adaptation).
V1 results: An LQA score of 3.4/5, below the traditional vendor baseline (3.7/5). Reviewers flagged 62 issues. But the nature of the issues was different—the traditional vendors' problems were systemic (terminology inconsistency, missing translations), while the AI's problems were fixable (subtle honorific deviations, a few unnatural phrasings).
V2 results (after feedback training): An LQA score of 4.1/5, above the baseline. Issues dropped from 62 to 18. The key improvements: per-character honorific configuration, and review edits automatically fed back into the system.
Phase 3: Rollout (Weeks 7-12)
They expanded to 12 languages across three waves:
| Wave | Languages | Review Strategy |
|---|---|---|
| Wave 1 | JA, KO, EN | Full AI + full review |
| Wave 2 | FR, DE, ES, PT | Full AI + 30% spot check |
| Wave 3 | AR, TH, VI, ID, TR | Full AI + targeted review |
After each wave launched, they collected a round of player feedback and QA data, confirming there were no systemic issues before expanding further.
Vendors were cut from six to two (retaining core transcreation and emergency-response capacity). The team was reshaped from 14 people to nine—three moved into overseas community operations, and two left through natural attrition without backfilling. No layoffs.
Phase 4: Steady-State Operations (Week 13+)
The new workflow: the PM uploads text (5 minutes) → AI translates into 12 languages (2-4 hours) → the quality dashboard automatically flags risky entries → the review team handles the flagged content → the main storyline gets a full review → ship.
The biggest change wasn't cost—it was tempo. Translation was no longer the bottleneck for version updates. "Simultaneous global launch" went from a dream to reality.
Three Unexpected Wins
1. Cross-language consistency issues vanished. All 12 languages now translate from the same glossary and character profiles, and brand consistency was achieved for real for the first time.
2. Iteration speed became a competitive advantage. Competitors' overseas builds lagged 3-5 days behind; Company G achieved simultaneous global launches. In the SLG genre, launching an event a day late can cost 5% of paying conversion.
3. The team went from a cost center to a strategy team. Their time allocation shifted from "chasing progress + fixing bugs" (60%) to "localization strategy + cultural adaptation + community operations" (65%).
Signs You're Ready to Transform
- ✅ Annual translation spend exceeds $300K
- ✅ Managing more than three translation vendors
- ✅ Version updates are bottlenecked by translation delivery
- ✅ Translation quality swings widely
- ✅ Terminology consistency is a long-standing pain point
Reasonable Expectations
- 40-70% cost reduction
- 3-10x faster delivery
- Quality matches or beats your baseline within 3 months, and improves significantly after 6
- AI needs a 2-4 week "learning period"—it won't be perfect on day one
For how this shift fits into the wider market, read our 2026 game localization whitepaper.