September 2026 Summaries

Lokalise argues that enterprise adoption of AI translation is limited less by model capability than by organizations’ inability to verify quality on their own content, and presents its Proof of Value feature as a way to address this concern. The feature evaluates a RAG-powered Custom AI profile, a base AI profile, and Google Translate against a company’s approved translations, using measures such as exact matches, BLEU, chrF, and Translation Edit Rate to estimate similarity and editing effort. It separates reference data from contextual translation-memory data to prevent leakage, while allowing teams to refine AI context through glossaries, style guides, language selection, date filters, and specialized profiles for different content types. The article emphasizes that high-quality translation memories and references are essential, since flawed data can both mislead evaluations and reinforce poor translation patterns. It distinguishes controlled Proof of Value comparisons, intended to help select an AI approach, from production Analytics, which tracks ongoing editing and acceptance outcomes. Citing internal customer evaluations from April through August 2026, Lokalise reports that Custom AI profiles consistently exceeded base profiles in perfect-match rates, particularly for nuanced content with rich historical translation context, while noting that results vary by language, content, and dataset quality.