Frequently Asked Questions

AI translation uses large language models (LLMs) and neural machine translation (NMT) engines powered by deep learning transformer architectures to convert text between languages with contextual understanding of tone and nuance — a major leap from older rule-based and statistical MT systems that produced literal, awkward output. Unlike earlier systems like pre-2016 Google Translate, modern AI engines (GPT-4, Claude, Gemini, DeepL) generate fluent, contextually appropriate translations that can rival human translators for many content types. At Into23, we use the term 'AI Translation with Verification' to describe our approach: leveraging the speed and scale of AI while ensuring every output passes through certified human review for guaranteed accuracy.

Neural Machine Translation (NMT) engines like Google Translate and DeepL are purpose-built systems trained specifically on parallel bilingual corpora — millions of aligned sentence pairs. They excel at producing fluent translations quickly and are optimized for throughput. LLM-based translation uses general-purpose large language models (GPT-4, Claude, Gemini) that were trained on vast amounts of multilingual text. LLMs bring advantages in handling context, ambiguity, creative content, and following specific instructions (e.g., 'translate this legal clause maintaining formal register'). However, LLMs can sometimes hallucinate or add information not present in the source. The optimal approach — which Into23 employs through Into23 Workspace — is to select the best-fit engine per language pair and content domain, combining the reliability of NMT with the contextual intelligence of LLMs.

AI translation can handle specialized content, but accuracy risks increase significantly for high-stakes domains like legal contracts, pharmaceutical documentation, and patent filings, where a single mistranslation can have serious regulatory or financial consequences. Modern AI engines perform well on general business content and marketing copy, but for highly specialized domains the risk of subtle terminology errors increases. This is why Into23 uses AI with Into23 Verify+: AI handles the heavy lifting for speed and cost efficiency, while certified domain-specialist linguists review every output for accuracy, terminology consistency, and regulatory compliance. For regulated industries, we recommend our Tier 2 (Into23 Verify+ at $0.08-$0.15/word) or Tier 3 (Enterprise AI Ops) service levels.

Into23 supports AI translation across 75+ languages, with quality varying by language pair: high-resource pairs (English to/from Spanish, French, German, Chinese, Japanese, Korean) produce consistently strong results, while medium-resource languages (Hindi, Arabic, Thai, Vietnamese) have improved dramatically with LLMs but may need more human oversight. Low-resource languages (many African, Southeast Asian, and indigenous languages) remain challenging for AI. Our Into23 Workspace platform provides access to 17+ MT/LLM engines, allowing us to select the optimal engine for each specific language pair and content type — for example, DeepL excels in European languages while specialized Chinese models outperform Western LLMs for CJK pairs.

AI translation with Into23 Verify+ is Into23's core methodology. It combines the speed and scalability of AI translation with mandatory human quality assurance. Every piece of content translated through our pipeline, whether via NMT engines or LLMs, passes through a certified human linguist who verifies accuracy, cultural appropriateness, terminology consistency, and brand voice alignment. This approach solves the fundamental tension in enterprise translation: businesses need the speed and cost efficiency of AI, but they cannot accept the risk of unverified machine output for client-facing, legal, or regulated content. Our verification layer includes LQA (Linguistic Quality Assurance) scoring, reviewer identification, and complete audit trails, giving enterprises the documentation they need for compliance and quality governance.

Modern LLMs are remarkably capable at understanding and preserving context and tone — a significant leap from earlier MT systems. When prompted correctly, models like GPT-4 and Claude can maintain formal register for legal documents, adopt conversational tone for marketing copy, or preserve technical precision for engineering documentation. However, AI still struggles with certain nuances: cultural idioms that don't translate directly, humor that relies on wordplay, brand-specific voice guidelines, and content where the intended audience's cultural context differs significantly from the source. Into23 addresses this through our engine selection process (choosing the right AI for each content type) and our Into23 Verify+ layer, where linguists ensure tone and cultural appropriateness meet the client's brand standards.

AI is transforming the role of human translators rather than replacing them — the industry is shifting toward 'human-in-the-loop AI' where technology handles scale and speed while humans provide judgment, cultural intelligence, and quality assurance. For high-volume, lower-risk content (internal communications, knowledge base articles), AI can handle 80-90% of the work with light human review. For high-stakes content (legal, medical, marketing campaigns), human expertise remains essential — but AI dramatically accelerates the process. At Into23, we invest in both our AI platform (Into23 Workspace) and our global network of certified linguists because both are essential to the future of language operations.

Prompt engineering for translation involves crafting specific instructions that guide LLMs to produce optimal translation output. Unlike NMT engines that simply take source text and produce a translation, LLMs can be instructed with detailed context: 'Translate this pharmaceutical label from English to Japanese, maintaining formal medical register, using approved terminology from the attached glossary, and preserving all numerical values exactly as written.' Effective prompt engineering can dramatically improve LLM translation quality by specifying domain, tone, target audience, terminology preferences, and formatting requirements. Into23's Into23 Workspace platform incorporates optimized prompt templates per industry vertical and content type, ensuring consistent, high-quality output across projects. Our prompt logs are included in every LQA report for full auditability.

Translation Memory (TM) is a database that stores previously translated text segments as source-target pairs, automatically reusing them when identical or similar content appears in future projects — reducing costs, ensuring consistency, and accelerating turnaround. When new content is submitted, the TM compares each segment against stored translations: a '100% match' (exact match) is applied automatically, while 'fuzzy matches' show similar segments with a percentage similarity score. TM technology has been the backbone of the localization industry for over 25 years. Into23 maintains client-specific TMs that integrate with both our human workflows and our AI translation pipeline through Into23 Workspace, meaning your TM investment compounds in value over time.

Fuzzy matching is the process by which a Translation Memory system identifies segments that are similar — but not identical — to previously translated content. The percentage score indicates how closely the new segment matches a stored translation. A 100% match means the segment is identical to one already in the TM. A 95-99% match (often called a 'high fuzzy') typically means minor differences like a changed number, date, or product name. A 75-94% match indicates more substantial differences that require careful review. Below 75%, the differences are usually significant enough that starting fresh may be more efficient. Industry-standard pricing models reflect these tiers: 100% matches are often free or heavily discounted, high fuzzies are discounted 50-70%, and lower fuzzies receive smaller discounts. Into23 applies standard fuzzy match discount tiers and provides transparent TM leverage reports so clients can see exactly how much value their TM investment delivers.

A Computer-Assisted Translation (CAT) tool is specialized software that helps translators work more efficiently. Unlike machine translation (which translates automatically), a CAT tool is a productivity environment that presents source text segment by segment, integrates Translation Memory suggestions, provides terminology databases, and offers quality assurance checks. Popular CAT tools include memoQ, Trados Studio, Phrase (formerly Memsource), and XTM. Modern CAT tools now integrate AI translation engines directly, allowing translators to review and refine AI output within the same interface. Into23 is technology-agnostic and works with all major CAT platforms. Our partnership with memoQ and integration with Custom.MT infrastructure means we can slot into existing client workflows without requiring tool migration.

A termbase — also called a glossary or terminology database — is a structured collection of approved terms and their translations across target languages. Terminology management ensures that key brand names, product terms, technical vocabulary, and industry-specific language are translated consistently across all content and all languages. Inconsistent terminology is one of the most common quality issues in enterprise translation. If your product is called 'SmartFlow Controller' in English, it must be rendered identically every time it appears in Japanese, Chinese, or German — not sometimes as 'Smart Flow Controller' and other times as 'Intelligent Flow Device.' Into23 builds and maintains client-specific termbases as part of our onboarding process. These termbases feed into both our human translation workflows and our AI engine prompts, ensuring consistency whether content is processed by a linguist or an LLM.

Translation converts text between languages preserving original meaning; localization adapts content for a specific market including cultural references, date/number formats, currency, and layout direction; transcreation recreates the message in the target language to achieve the same emotional impact — commonly required for marketing slogans and advertising campaigns. Into23 offers all three levels of service. Our AI translation pipeline handles translation and many localization tasks efficiently, while transcreation projects are led by specialist creative linguists with AI-assisted ideation support. The choice between these approaches depends on content type, audience, and risk tolerance.

A context match — sometimes called a 101% match, ICE match (In-Context Exact), or perfect match — goes beyond a standard 100% match by also verifying that the surrounding segments are identical. In other words, not only is the segment itself the same as a previously translated one, but the sentences before and after it are also the same. This provides the highest level of confidence that the existing translation is correct in context. Context matches are particularly valuable for software UI strings, where the same word might be translated differently depending on what appears around it (e.g., 'Save' as a verb vs. 'Save' as a noun). Most CAT tools and TM systems support context matching, and Into23 leverages this in our workflows to maximize consistency and minimize unnecessary review costs.

TM leverage refers to the percentage of a new translation project that can be matched against existing Translation Memory content. High TM leverage means a significant portion of the content has been translated before, reducing both cost and turnaround time. For example, if a 10,000-word document has 40% exact matches and 20% high fuzzy matches, only about 40% of the content requires full translation effort. Into23 provides detailed TM leverage reports at the quoting stage so clients can see projected savings before approving a project. Over time, as your TM grows, leverage typically increases — meaning each subsequent project becomes more cost-effective. This is one reason we encourage long-term partnerships: the TM becomes a valuable asset that compounds in value.

Alignment is the process of taking previously translated documents — where you have the source and target versions but no Translation Memory — and creating TM entries by matching source segments with their corresponding translations. This is commonly done when a client switches translation providers or adopts a new TM system and wants to preserve the value of past translations. The process can be done manually (a linguist matches segments one by one) or semi-automatically using alignment tools that use formatting and structural cues to propose matches. Into23 offers alignment services as part of our onboarding process, ensuring that clients who come to us with existing translated content don't lose the investment they've already made.

Modern translation workflows support an extensive range of file formats. Common document formats include Microsoft Word (.docx), Excel (.xlsx), PowerPoint (.pptx), PDF, and InDesign (.idml). For software and web content, we handle JSON, XML, XLIFF, PO/POT, YAML, HTML, and Markdown. Multimedia workflows support SRT/VTT subtitle files, and we can extract translatable content from design tools like Figma and Sketch. Into23's Into23 Workspace platform supports 50+ file formats for direct upload, and our API accepts structured content in JSON and XLIFF for automated pipeline integration. For formats that require special handling (e.g., embedded graphics in PDFs, or complex InDesign layouts), our DTP team ensures the translated files match the original formatting precisely.

Into23's Into23 Workspace platform provides unified access to 17+ machine translation and LLM engines through a single interface, including Google Cloud Translation, DeepL, Microsoft Translator, Amazon Translate, OpenAI GPT-4/GPT-4o, Anthropic Claude, Google Gemini, and leading Chinese AI models (Baidu Ernie, Alibaba Qwen, Moonshot Kimi, MiniMax). Each engine has different strengths: DeepL excels in European languages, Google handles the broadest language coverage, GPT-4 is strong for creative and contextual content, and specialized Chinese models outperform Western LLMs for CJK language pairs. Our engine selection algorithm recommends the optimal engine based on language pair, content domain, and quality requirements.

Best-fit engine selection is Into23's approach to choosing the optimal translation engine for each specific project based on language pair, content type, domain, and quality requirements. No single engine is best for everything. DeepL may outperform GPT-4 for German legal text, while GPT-4 may produce better results for creative Japanese marketing copy. Our Into23 Workspace platform maintains performance benchmarks across all supported engines, updated continuously based on LQA scores from verified projects. When a new project comes in, the system recommends the engine most likely to produce the highest quality output for that specific combination of parameters. Clients can also override the recommendation and select a preferred engine manually. This multi-engine approach is a key differentiator — rather than being locked into a single provider, you get the best tool for each job.

Yes — through Into23's partnership with Custom.MT, clients can train custom MT engines using their own bilingual data (minimum 50,000+ segment pairs required), producing output that matches their specific terminology, style, and domain vocabulary with less post-editing needed. Custom-trained engines are particularly valuable for clients with large, recurring translation volumes in specialized domains — pharmaceutical companies, legal firms, or technology companies with proprietary terminology. The training process involves preparing and cleaning bilingual data, training the engine, evaluating output quality against benchmarks, and iterating until performance targets are met. Into23 manages this entire process as part of our Enterprise AI Ops tier.

Dedicated MT engines (Google Translate, DeepL) are trained specifically on parallel text — millions of source-target sentence pairs. They're optimized for translation speed, consistency, and handling high volumes. LLMs (GPT-4, Claude, Gemini) are general-purpose models trained on vast amounts of text in many languages. Their translation capabilities emerge from this broad training rather than being their primary function. The practical differences are significant: MT engines are faster and cheaper per word, more consistent across similar content, and less prone to hallucination. LLMs are better at understanding context across paragraphs, following specific style instructions, handling ambiguous or creative content, and adapting tone. LLMs can also perform translation and other tasks simultaneously — for example, translating while also simplifying the language level or adapting cultural references. Into23 uses both types strategically through Into23 Workspace.

Adaptive machine translation refers to MT systems that learn and improve from translator corrections in real time. When a human post-editor corrects an MT suggestion, the system incorporates that correction and applies it to subsequent similar segments within the same project or session. This creates a feedback loop where the MT output gets progressively better as more corrections are made. Some CAT tools (like Lilt and memoQ with certain engine integrations) support adaptive MT natively. Into23's workflow incorporates a similar feedback mechanism: corrections made during our Into23 Verify+ step are logged and used to refine engine selection, prompt templates, and terminology enforcement for future projects. Over time, this means your translations get better and cheaper as the system learns your preferences.

Generic MT engines are trained on broad, general-purpose datasets and perform well across a wide range of content types. Domain-specific MT engines are fine-tuned on data from a particular industry or subject area — legal, medical, financial, technical — and produce significantly better results for content in that domain. The difference can be dramatic: a generic engine might translate a medical term incorrectly or use informal language in a legal document, while a domain-specific engine would use the correct terminology and appropriate register. Into23 addresses this through a combination of engine selection (choosing engines known to perform well in specific domains), custom glossary enforcement, and prompt engineering that provides domain context to LLMs. For Enterprise AI Ops clients, we can deploy fully custom-trained domain-specific engines.

Linguistic Quality Assurance (LQA) is a systematic evaluation of translation quality using the MQM (Multidimensional Quality Metrics) framework, which categorizes errors by type (accuracy, fluency, terminology, style) and severity (critical, major, minor), producing a weighted quality score on a 0-100 scale — with scores above 95 considered publication-ready. Each error is assigned a weighted penalty score calculated against the word count. Into23 generates LQA reports for every verified translation project, including error categorization, severity assessment, reviewer identification, and an overall quality score. These reports provide the objective quality data that enterprise clients need for vendor management and compliance documentation.

Machine Translation Post-Editing (MTPE) is the process where a human linguist reviews and corrects machine-translated output, typically 30-60% faster than translating from scratch and proportionally less expensive. There are two standard levels: Light Post-Editing (LPE) produces comprehensible, accurate output suitable for internal communications, while Full Post-Editing (FPE) produces publication-quality output indistinguishable from human translation — required for client-facing and regulated content. The key to effective MTPE is starting with high-quality MT output, which is why Into23's best-fit engine selection is critical: better raw MT means less post-editing effort, lower costs, and faster delivery.

Our verification process follows a structured workflow: First, content is translated by the selected AI engine (NMT or LLM) through the Into23 Workspace platform. The AI output is then routed to a certified linguist who specializes in the relevant language pair and domain. The linguist reviews every segment for accuracy, fluency, terminology consistency, cultural appropriateness, and adherence to client style guides. Corrections are made directly, and each edit is logged. After verification, an automated QA check runs for technical issues (tag integrity, number formatting, placeholder consistency). Finally, an LQA report is generated with the quality score, error breakdown, reviewer ID, and complete edit history. For Enterprise AI Ops clients, we add a second review layer and project manager oversight for additional quality assurance.

Back-translation is the process of translating a completed translation back into the source language, performed by an independent linguist who has not seen the original text. The back-translation is then compared with the original to identify any meaning shifts, omissions, or errors. This technique is commonly required in pharmaceutical and clinical trial documentation, where regulatory bodies (FDA, EMA) may require back-translations to verify that informed consent forms, patient-facing materials, and labeling are accurate. It's also used in legal contexts where translation accuracy has contractual or regulatory implications. Into23 offers back-translation as an add-on service for clients in regulated industries, with full documentation suitable for regulatory submission.

Enterprise translation programs should track several key metrics: LQA Score (average quality score across projects, targeting 95+), First-Pass Quality Rate (percentage of segments accepted without edits during verification), Turnaround Time (from submission to delivery), TM Leverage (percentage of content matched from Translation Memory), Cost Per Word (across different content types and language pairs), and Error Distribution (which error types are most common, indicating areas for process improvement). Into23 provides all of these metrics through our reporting dashboard, with the ability to filter by language pair, content type, engine used, and time period. For Enterprise AI Ops clients, we conduct quarterly business reviews analyzing trends and recommending optimizations.

The Multidimensional Quality Metrics (MQM) framework is an internationally recognized standard for evaluating translation quality, developed by the German Research Center for Artificial Intelligence (DFKI) and widely adopted across the localization industry. MQM provides a comprehensive error typology with eight top-level categories: Accuracy, Fluency, Terminology, Style, Design, Locale Convention, Verity, and Internationalization. Each category has detailed sub-categories (e.g., under Accuracy: mistranslation, omission, addition, untranslated). Errors are classified by severity — Critical (content is unusable or dangerous), Major (significantly impacts comprehension or usability), and Minor (noticeable but doesn't impair understanding). The framework allows organizations to customize which dimensions matter most for their specific use case. Into23 uses MQM as the foundation for all our LQA evaluations.

Consistency across large programs requires a systematic approach combining technology and process governance. Into23 employs several mechanisms: centralized Translation Memory ensures previously approved translations are reused consistently; client-specific termbases enforce uniform terminology across all languages; style guides document tone, formatting, and brand voice preferences; reference materials and approved translations serve as benchmarks for new linguists. On the technology side, our Into23 Workspace platform enforces glossary terms during AI translation, runs automated consistency checks across language versions, and flags deviations from established patterns. For Enterprise AI Ops programs, we assign dedicated project managers and lead linguists per language who maintain deep familiarity with the client's content and preferences over time.

Into23 protects confidential content through end-to-end encryption (TLS 1.3 in transit, AES-256 at rest), Alibaba Cloud infrastructure with APAC data residency options, strict need-to-know access controls with full audit logging, and NDAs with all linguists and staff. Our MT engine partner Custom.MT holds ISO 27001 certification, and Into23 holds ISO 9001 and ISO 17100 certifications, with our own ISO 27001 certification currently in progress. For clients with heightened security requirements, we offer dedicated infrastructure, VPN access, and custom data retention policies.

No — Into23 has explicit data processing agreements with all MT and LLM providers (including OpenAI, Google, DeepL, and Anthropic) contractually prohibiting the use of client content for model training. When we send content through our Into23 Workspace platform, we use enterprise API agreements that ensure your data is processed and discarded, never retained for training purposes. This is a critical distinction from consumer-grade tools like free Google Translate or ChatGPT, where user inputs may be used to improve models. Into23 can provide copies of our data processing agreements with each engine provider upon request.

Yes. Into23 maintains full compliance with the General Data Protection Regulation (GDPR) for all content involving EU personal data. Our compliance measures include: Data Processing Agreements (DPAs) with all clients and sub-processors; lawful basis documentation for all processing activities; data minimization practices (we only process the content necessary for the translation task); right to erasure procedures (client content can be permanently deleted upon request); breach notification protocols (within 72 hours as required); and regular privacy impact assessments. For clients subject to other data protection regimes (PDPA in Singapore, PIPL in China, APPI in Japan), we adapt our practices accordingly. Our legal team can provide detailed compliance documentation for any specific regulatory framework.

Into23 Workspace is hosted on Alibaba Cloud with primary infrastructure in the Asia-Pacific region. For clients with specific data residency requirements, particularly common in financial services, healthcare, and government sectors, we offer data residency options that ensure content is processed and stored within specified geographic boundaries. Through our Custom.MT partnership (ISO 27001 certified), we can deploy dedicated translation infrastructure in specific Alibaba Cloud regions. We also support on-premises deployment discussions for clients with the most stringent requirements. Data residency configurations are documented in our service agreements and can be audited by client security teams.

When translation content contains PII — names, addresses, identification numbers, health information, financial data — we apply additional safeguards. Our standard approach includes PII identification and flagging during project intake, restricted access controls (fewer linguists with higher clearance), enhanced audit logging, and accelerated data deletion after project completion. For clients in healthcare (HIPAA) or financial services, we can implement PII masking/pseudonymization before translation, where sensitive data elements are replaced with placeholders, translated, and then re-inserted in the target language. This ensures linguists never see actual PII while still producing accurate translations. All PII handling procedures are documented in our Data Processing Agreements.

Free online translation tools pose several significant security risks for enterprise content. First, data retention: most free tools retain submitted content indefinitely and may use it to improve their models — meaning your confidential contracts, product roadmaps, or patient data could become part of a training dataset. Second, no access controls: anyone with the URL can potentially access cached translations. Third, no audit trail: there's no record of what was translated, when, or by whom — a compliance nightmare for regulated industries. Fourth, no data processing agreements: you have no contractual protection for how your data is handled. High-profile incidents have occurred where confidential corporate information submitted to free translation tools was later discoverable through search engines. Into23's enterprise platform eliminates all of these risks through contractual protections, encryption, access controls, and audit logging.

Into23 holds ISO 9001 (Quality Management Systems) and ISO 17100 (Translation Services) certifications, with ISO 27001 (Information Security Management) certification in progress for 2026 completion. Our technology partner Custom.MT already holds ISO 27001 certification. Our infrastructure runs on Alibaba Cloud, which maintains SOC 2 Type II, ISO 27001, and numerous other certifications. We conduct regular security assessments, penetration testing, and vulnerability scanning. For enterprise clients, we can provide SOC 2 readiness documentation, security questionnaire responses, and arrange calls with our security team.

Our standard data retention policy retains Translation Memory and terminology data indefinitely (as these are valuable assets that improve over time), while source files and translated deliverables are retained for 12 months after project completion to facilitate revisions and reference. After the retention period, files are securely deleted. However, we fully customize retention policies based on client requirements. Some clients in regulated industries require longer retention for compliance purposes, while others require immediate deletion after delivery. All retention terms are documented in our service agreements, and clients can request early deletion at any time. Translation Memory data is always considered client property and can be exported in standard TMX format upon request.

AI Data Services are human-powered tasks required to train, evaluate, and improve AI systems — including LLM evaluation, RLHF annotation, AI red teaming and safety testing, and creation of annotated training datasets. Into23's AI Data Services division specializes in multilingual scenarios: evaluating AI performance across languages, providing human feedback in non-English languages, and creating training data for multilingual AI systems. Our team of trained annotators across 6 priority languages (English, Chinese, Hindi, Japanese, Korean, Arabic) supports AI companies building global products that need to perform reliably beyond English.

AI Red Teaming is the systematic practice of probing AI systems to identify vulnerabilities, biases, harmful outputs, and failure modes before they reach end users — through adversarial prompting that tests for toxic content generation, misinformation, bias, privacy violations, and jailbreak vulnerabilities. Into23's multilingual red teaming service is particularly valuable because safety issues often manifest differently across languages and cultures: a model that behaves safely in English may produce harmful content in Hindi or Japanese due to different training data distributions. Our red team annotators are native speakers trained in adversarial testing methodologies, providing culturally informed safety evaluation across 6 priority languages.

RLHF (Reinforcement Learning from Human Feedback) is the training technique that aligns large language models with human preferences by having evaluators compare pairs of model outputs and indicate which is better — more helpful, accurate, appropriate, or safer — then using this preference data to train a reward model that guides the LLM toward preferred outputs. RLHF is the technique behind the dramatic improvement in models like ChatGPT, Claude, and Gemini — transforming them from technically capable to genuinely helpful and conversational. Into23 provides multilingual RLHF annotation services where trained evaluators assess model outputs across multiple languages, ensuring AI alignment extends beyond English to all major deployment languages.

LLM Evaluation involves systematically assessing the quality of large language model outputs across multiple dimensions: factual accuracy, helpfulness, coherence, safety, instruction-following, and cultural appropriateness. Response rating is the specific task of scoring individual model outputs on these dimensions, typically using structured rubrics. This evaluation data serves multiple purposes: it identifies model weaknesses that need improvement, provides benchmarking data for comparing models, generates training signal for RLHF, and validates model performance before deployment. Into23's evaluation services cover both monolingual and cross-lingual scenarios — for example, evaluating whether a model's Japanese responses are as accurate and helpful as its English responses, or whether translated content maintains the quality of the original.

Into23 provides three tiers of transcription services: Orthographic Transcription (verbatim text capture of audio/video content), Tagged Transcription (with speaker identification, timestamps, and non-verbal cues), and Semantic Transcription (with additional annotation of intent, sentiment, and topic classification). Our annotation services extend to Audio Annotation (speech quality assessment, emotion and sentiment labeling, prompted speech collection for TTS training), and Image Recognition and Annotation (Visual Question Answering, OCR validation, and damage/defect detection for industrial applications). All services are available across our 6 priority languages, with expansion to additional languages on a project basis. These services power AI training pipelines for speech recognition, computer vision, and multimodal AI systems.

Multilingual AI evaluation introduces several layers of complexity beyond English-only assessment. First, evaluators must be native speakers with cultural fluency — you cannot reliably evaluate Japanese AI output with English-speaking annotators, even if they speak Japanese as a second language. Second, quality dimensions may manifest differently across languages: what constitutes 'natural' phrasing varies dramatically between languages, and cultural norms around directness, formality, and humor differ. Third, factual accuracy evaluation requires knowledge of local context — a model might generate information that's accurate for the US but incorrect for Japan or India. Fourth, safety evaluation must account for culturally specific sensitivities. Into23's multilingual evaluation teams are native speakers with domain training, ensuring that AI evaluation reflects genuine user experience in each target market.

Into23 Workspace is Into23's cloud-based translation management portal that gives enterprise clients unified access to 17+ MT and LLM engines through a single interface — upload files (50+ formats supported) or send content via REST API, select or auto-assign the optimal translation engine, and receive translated output with optional Into23 Verify+. The platform includes a real-time cost dashboard showing usage and spend across all engines, project management tools for tracking translation status, and automated LQA reporting. Into23 Workspace is designed for enterprise teams managing high-volume, multi-language translation programs who need full visibility into quality, cost, and turnaround metrics. Access is via web portal at intone.into23.com or through our API for automated pipeline integration.

Yes. Into23 Workspace provides a comprehensive REST API that enables automated translation workflows. Common integrations include CMS platforms (translating content automatically when new pages are published), product information management systems (keeping product descriptions synchronized across languages), customer support platforms (translating tickets and knowledge base articles), and CI/CD pipelines (localizing software strings as part of the build process). The API supports both synchronous (real-time) and asynchronous (batch) translation requests, with webhook notifications for job completion. We provide API documentation, SDKs for popular programming languages, and integration support during onboarding. For Enterprise AI Ops clients, our engineering team can build custom integrations.

Into23 Workspace's real-time cost dashboard provides complete visibility into translation spend across all engines, languages, and projects. The dashboard displays current month usage (words translated), cost breakdown by engine and language pair, quality scores from verified projects, and trend analysis over time. Clients can set budget alerts, compare cost-per-word across different engines for the same language pair, and identify optimization opportunities. For example, the dashboard might reveal that switching from Engine A to Engine B for Korean legal content saves 15% while maintaining the same quality score. All cost data is exportable for integration with internal financial reporting systems. The dashboard updates in real time as translation jobs are processed.

Into23 Workspace supports 50+ file formats for direct upload, covering the most common enterprise content types. Document formats include DOCX, XLSX, PPTX, PDF, RTF, and TXT. Web and software formats include HTML, XML, JSON, YAML, PO/POT, XLIFF (1.2 and 2.0), and properties files. Design formats include IDML (InDesign) and SRT/VTT for subtitles. The platform automatically detects the file format, extracts translatable content while preserving formatting and structure, processes the translation, and reconstructs the output file in the original format. For formats not natively supported, our team can configure custom parsers. API users can also send raw text or structured content in JSON format for maximum flexibility.

When Into23 Verify+ is enabled (standard for Tier 2 and Tier 3 service levels), the workflow is seamless within Into23 Workspace. After AI translation is complete, the content is automatically routed to a qualified linguist from Into23's network based on language pair, domain expertise, and availability. The linguist reviews the AI output in a dedicated verification interface, making corrections and annotations. Once verification is complete, the verified output replaces the raw AI translation in the project, and an LQA report is automatically generated. The entire process — from AI translation through Into23 Verify+ to final delivery — is tracked in Into23 Workspace with timestamps, reviewer identification, and quality metrics. Clients can monitor verification progress in real time through the dashboard.

Yes — Into23 Workspace is currently in beta with limited enterprise pilot slots available for Q1-Q2 2026, offering early access, dedicated onboarding support, preferential pricing, and direct input into feature development priorities. Beta access is ideal for enterprises currently managing multi-engine translation workflows manually or those looking to consolidate their translation technology stack. To request beta access, visit our staging page or contact us directly at [email protected]. We evaluate beta applications based on translation volume, language pair requirements, and alignment with our platform roadmap.

Into23 offers three service tiers: Tier 1 AI Basic ($0.03-$0.08/word) for internal documents and lower-risk content with AI-only translation; Tier 2 Into23 Verify+ ($0.08-$0.15/word) for client-facing and regulated content where every word is AI-translated then verified by a certified linguist; and Tier 3 Enterprise AI Ops (annual partnerships from $300K-$3M) for large-scale programs with dedicated teams, custom engine training, and SLA-backed delivery. All tiers include access to the Into23 Workspace platform, Translation Memory leverage discounts, and LQA reporting.

Into23's onboarding follows a structured process designed to set up your translation program for long-term success. Phase 1 (Discovery, 1-2 weeks): We assess your content types, language pairs, volume projections, quality requirements, and existing assets (TMs, glossaries, style guides). Phase 2 (Setup, 1-2 weeks): We configure your Into23 Workspace workspace, import existing TMs and termbases, set up engine preferences, and establish quality benchmarks. Phase 3 (Pilot, 2-4 weeks): We run a controlled pilot project to validate quality, turnaround, and workflow integration before scaling. Phase 4 (Scale): Based on pilot results, we ramp to full production volume with continuous optimization. For Enterprise AI Ops clients, onboarding includes dedicated project manager assignment, custom engine evaluation, and integration engineering support.

Into23 serves enterprise clients across several key verticals. Legal: contracts, litigation support, regulatory filings, patent documentation — where accuracy is non-negotiable and terminology precision is critical. Life Sciences: clinical trial documentation, regulatory submissions, patient-facing materials, medical device labeling — where compliance requirements (FDA, EMA) demand rigorous quality processes. eCommerce: product descriptions, marketing campaigns, customer support content — where speed and volume are paramount for global marketplace expansion. Consumer Electronics: user manuals, UI strings, marketing materials, support documentation — where technical accuracy meets consumer accessibility. AI & Technology: training data, model evaluation, safety testing — where our AI Data Services capabilities complement our translation expertise. Our APAC base gives us particular strength in CJK (Chinese, Japanese, Korean) language pairs.

Into23's standard SLAs include: AI-only translation delivered within 2-4 hours, Into23 Verify+ within 24-48 hours, minimum LQA quality score of 95 for verified translations, 99.5% Into23 Workspace platform uptime, and support ticket response within 4 business hours. Enterprise AI Ops clients receive enhanced SLAs with dedicated support channels, escalation procedures, and financial remedies for SLA breaches. All SLAs are documented in service agreements and tracked through Into23 Workspace's reporting dashboard, with customization available based on specific client requirements.

Yes. Our AI-first approach means we can deliver initial translations within minutes for most content types and language pairs. For AI-only (Tier 1) projects, turnaround is essentially real-time for reasonable volumes. For AI + Verification (Tier 2) rush projects, we maintain a network of linguists across time zones to provide 24/7 coverage. Rush surcharges apply (typically 25-50% depending on turnaround requirement and language pair), but the AI-first workflow means even rush verified translations are significantly faster than traditional human-only approaches. Enterprise AI Ops clients can pre-arrange dedicated linguist availability for anticipated rush periods (product launches, regulatory deadlines, marketing campaigns).

Scalability is a core advantage of Into23's AI-first model. Our Into23 Workspace platform can process millions of words per day through its multi-engine architecture, so there is no bottleneck waiting for human translators to become available for the initial translation pass. For the Into23 Verify+ layer, we maintain a scalable network of certified linguists across all supported language pairs, with capacity planning based on client volume forecasts. For Enterprise AI Ops clients, we pre-assign dedicated linguist teams sized to handle projected volumes with surge capacity for peaks. Our infrastructure scales on Alibaba Cloud, and our engine selection algorithm distributes load across multiple providers to avoid single-provider rate limits. We have successfully handled projects exceeding 10 million words per month across multiple language pairs.

Into23 is headquartered in Hong Kong with offices in Beijing, Shanghai, Bangalore, and Tokyo — providing strong coverage across Asian languages, time zones, and markets while serving enterprise clients across North America, Europe, and Asia-Pacific. Our APAC footprint gives us direct access to key Asian markets and large pools of multilingual talent for both translation and AI data services. We have particular strength serving companies headquartered in NA/EU that are scaling operations in APAC markets, with our Hong Kong headquarters positioning us at the crossroads of East-West business.

Into23 is an AI-native language operations partner built from the ground up around an AI-first architecture, unlike traditional LSPs that have bolted AI onto legacy workflows. Key differentiators include the Into23 Workspace platform providing unified access to 17+ engines, AI with Into23 Verify+ on every output, best-fit engine selection by language pair and domain, and dual capability in both AI translation and AI data services from a single partner. Our APAC operational base provides cost advantages and CJK language expertise that Western-headquartered providers often lack.