August 2026 Summaries

Figma’s native AI agent, introduced at Config, can connect to Lokalise through the Model Context Protocol (MCP), enabling designers to manage localization conversationally without leaving the design canvas. The approach complements Lokalise’s established Figma plugin: the plugin provides structured, repeatable workflows for exporting text, creating translation keys, assigning translation or review tasks, generating language-specific design copies, and resyncing updates, while the AI-agent connection supports more customized requests such as checking content drift, syncing variables, and monitoring reviews. Both routes use Lokalise’s localization infrastructure, including glossaries, translation memory, AI quality scoring, visual context, and human-review workflows, which aim to address consistency, scale, and quality limitations of translating directly in Figma. Users connect the AI agent by configuring Lokalise’s MCP server URL and authentication in Figma, then can use natural-language prompts or reusable Markdown-based AI skills to standardize tasks such as pushing source text, creating translation tasks, assigning reviewers, detecting changed strings, and pulling approved translations into designs.

AI translation can reduce localization costs and turnaround times, but wider enterprise adoption depends on demonstrating that quality meets defined standards through objective, reproducible evaluation rather than isolated human reviews. The proposed framework separates three functions: pre-production evaluation compares translation methods against the same human-reviewed reference translations using measures such as perfect match rate, Translation Edit Rate (TER), BLEU, and ChrF; in-production scoring determines whether individual strings require human review; and post-edit analytics tracks how much human reviewers change AI output over time. Human assessment remains important for meaning, tone, and brand voice, but samples can be subjective, inconsistent, and unrepresentative across content types. Reliable evaluation requires current, human-reviewed reference data matched to the relevant content and language pair, with at least 500 source-and-translation pairs recommended per pair. Lokalise argues that its personalized Custom AI Profiles, informed by approved translation data and contextual assets, can outperform generic AI and standard machine translation, citing internal customer examples and acceptance-rate claims, while its platform automates comparisons and presents both metric-based results and side-by-side outputs to support decisions about scaling AI into additional languages, workflows, and content types.