AI + Expert Workflow | Angel Translation

The Shift

How AI has changed the workflow

Traditional Process

Manual-only workflow

Client submits files

Input

Human translation (full document)

Manual proofreading

Delivery

Output

Angel AI + Expert Process

Hybrid 5-step delivery system

Client submits files

Input

AI draft generation

Terminology validation

AI + DB

Domain expert review

Human

LISA QA audit + delivery

Output

AI changes efficiency.Not accountability. — The expert review and quality audit steps are non-negotiable, regardless of AI involvement.

The Right Tool for the Right Job

When AI is enough. When it isn't.

AI Translation Excels At

Speed and scale — for the right content type

✓Internal reference documents and drafts

✓High-volume repetitive content (e.g. product descriptions)

✓E-commerce listings and basic product specs

✓Initial drafts that will be reviewed before use

Expert Review Is Non-Negotiable For

When errors have business, legal, or safety consequences

● Semiconductor datasheets, EDA software, patent filings

● FDA/CE/NMPA regulatory submissions and IFU instructions

● Legal contracts, IP agreements, cross-border compliance

● Aerospace manuals, AMM documentation, safety specifications

● Brand campaigns, literary content, cultural localization

When translation affects business outcomes — regulatory approval, product safety, legal liability — AI alone is not enough.

Talk to a Specialist →

Different languages, different engines

Which model does what

Models differ sharply by language and by register, so we assign them per stage rather than running one engine across the whole pipeline.

Model assignment by stage
Where it is used	Model	Why this one
Chinese-language translation and review	Qwen MT	Trained heavily on Chinese; handles Chinese sentence structure and domain phrasing more reliably.
Foreign-language translation and review	Google translation model	Widest language coverage, and steadier on smaller languages than general-purpose chat models.
Creative content in foreign languages	Anthropic	Finer control of tone, rhythm and register — for copy that has to read as though written natively.
Regulatory and legal review	OpenAI	Holds reference and conditional relationships across long clauses; in this material a structural error costs more than an awkward word.

Model selection for the remaining stages varies with the language, the register and the confidentiality terms of the project, and is not listed exhaustively. Two things we will state plainly: models run inside our controlled delivery process, directed by people, and client documents are not used to train any public or third-party model. See Confidentiality & Security.

The Angel Method

Our 8-step delivery system

Every project runs this pipeline, with the review and QA depth set by the service level.

Client Submission Kickoff

Files, glossaries, reference materials, and target language specs are uploaded. A dedicated project manager is assigned immediately.

Response within 4 hours

AI Draft Generation AI Layer

Enterprise-grade LLM processes the content, grounded in our industry terminology assets built over 26+ years. Initial draft completed at 4× human speed.

Speed ×4 vs. traditional

Terminology Matching AI + DB

Every key term is cross-referenced against the client's project glossary and our 26-year industry terminology database. Inconsistencies flagged automatically before human review begins.

Term match rate verified

04
Domain Expert Review Human QA
A linguist with domain-specific academic qualifications (engineering, medical, legal, aerospace) reviews the full document. At this step, domain intent, compliance precision, cultural register, and contextual accuracy are verified line by line.
Most critical step — not automated

LISA QA Audit Human QA

An independent QA review applies the LISA QA 3.1 standard. Defect classification by severity. Compliance-critical projects receive formal QC report and written quality guarantee.

Composite error rate per GB/T 19682: L2 ≤1.5‰ (meets national standard) · L3 ≤1‰ · L4 ≤0.5‰

Format Restoration + Delivery Output

Translated content restored to original format (PDF, Word, InDesign, DITA, XML, software strings). Multi-language simultaneous release supported. QC report attached.

Original format preserved

Terminology Extraction TM Asset

All new approved terminology pairs are extracted and added to the client's dedicated term base. Project glossary delivered alongside translation. This asset compounds in value with every project.

Permanent terminology asset built

Continuous Optimization Long-term

Each completed project builds up that client's terminology assets and reference base. The accumulated terminology and style preferences reduce time and error rates on subsequent projects — without their content ever entering model training.

Each project improves the next

Service Architecture

Four service levels. One quality system.

Angel AI Review

AI draft + basic QC

Best for

Internal reference · High-volume repetitive content

→LLM full-auto draft

→Terminology consistency check

→Basic format preservation

Save 70%+ vs. traditional

Angel Pro

AI + linguist polish

Best for

Technical docs · Product manuals · Marketing

→AI draft + terminology validation

→Native linguist review + polish

→Style consistency enforced

Save 40–50% vs. traditional

★ Recommended

Angel Premium

Expert-led delivery

Best for

Compliance files · Medical · Legal · Whitepapers

→AI-assisted + domain expert full review

→LISA QA audit + QC report

→Written quality guarantee available

→Supports FDA / CE / ISO submissions

Comparable to traditional human cost

Angel Global

Publication-grade

Best for

Publishing · Film · Games · Brand launch

→Expert translation + creative adaptation

→Subtitles · Dubbing · Distribution support

→Cultural adaptation

20–50% premium · Significant creative value

Find Your Fit

Different projects. Different solutions.

Not all content is the same. Not all projects need the same solution.

L1 · Angel AI Review

AI Can Handle This

Speed over precision, internal use

Product reviews and user comments

FAQ pages and help center content

Internal email and memos

E-commerce product listings

Meeting notes and summaries

L2 · Angel Pro

AI + Expert Review

Quality matters, budget matters too

Product website and app copy

Marketing brochures and decks

Product operation manuals

Training and onboarding materials

Press releases and company news

L3 · Angel Premium

Expert-Led · Compliance-Ready

When errors have real consequences

Semiconductor datasheets and EDA software

FDA / CE / NMPA regulatory filings

Legal contracts and patent filings

Medical device IFU instructions

Aerospace AMM and safety documentation

Technical whitepapers for global audiences

L4 · Angel Global

Publication Grade

Brand, culture, creative impact

Film and drama subtitle localization

Animation and manga adaptation

Video game script and UI

Brand campaigns and slogans

Web novel and literary translation

Quality System

How we guarantee quality

Terminology Management

Client-specific term bases enforced across every project. Industry-specific terminology accumulated and continuously refined. No two clients share the same term base.

26-year term database · Per-client glossaries

Domain Expert Assignment

Every project is matched to a linguist with relevant domain experience, and to a specialist academic background where the material requires one. Semiconductor work goes to EE graduates. Medical filings go to biomedical specialists. Not interchangeable.

5,000+ domain-background specialists

LISA QA Audit Standard

An independent QA audit is applied to every project; its depth and reporting follow the service level. Defects are classified by severity and category. A QC report is delivered with the final files at L3 and above.

ISO 9001:2015 · LISA QA 3.1 · Written guarantee

Data Security + NDA

NDA available for all projects. Enterprise-grade file encryption in transit and at rest. Client content is processed only through approved enterprise AI services with data retention disabled, and is never used to train public models. Full audit trail maintained.

NDA · Encryption · Confidentiality audit trail

Accountability

Where the liability lands

A translation released without competent expert review raises the same three questions no matter who or what produced it: is it fit for purpose, who is accountable, and who pays when it is wrong. Independent industry research reached the same conclusion this year — when a translation error contributes to a bad outcome, the claim tends to travel to whichever party in the chain is least protected, not necessarily to the one who decided to skip the review.

A client instruction is not a shield

Asking for unreviewed output is a decision to accept a lower level of assurance. It does not, by itself, settle who answers for the consequences — and an instruction nobody wrote down is worth little once a dispute starts.

Review is the default, not an upsell

Two of our eight steps — expert review and quality review — stay with people. The composite error rate each service level commits to is measured on their output, not on the AI draft. We do not remove them quietly to hit a price.

The destination sets the level

A regulatory submission, an IFU, a patent claim and a product page do not carry the same exposure. We match the service level to where the file ends up, and say so before the work starts rather than after.

Source: CSA Research, September 2026. Summarised in our own words.

Frequently Asked

Questions about AI + expert translation

For all client projects, we use enterprise-grade API access to LLM providers with data retention turned off — your content is never used for model training. For highly sensitive projects, we can use on-premise AI deployment. NDAs are available for all projects. Full encryption in transit and at rest.

AI can generate a fluent first draft, but it cannot verify whether a technical term is being used correctly in its specific engineering context, whether a regulatory phrase meets FDA requirements, or whether a brand message resonates culturally. More importantly, AI tools do not carry legal accountability. When a medical device registration is rejected or a semiconductor spec is misunderstood, someone must be accountable. We are.

Every client has a dedicated term base that is applied at both the AI draft stage (automated matching) and the expert review stage (manual verification). Terms approved in one document are automatically enforced in all subsequent documents for that client. We also deliver a project glossary at completion — a permanent asset that grows in value over time.

Yes. Clients can specify a human-only workflow (Angel Premium or Angel Global service levels). In these cases, AI is not used for translation generation — our team works directly from source. Terminology databases and QA tools are still applied. This option is particularly common for brand-sensitive and literary content.

We use enterprise API access to leading models, combined with our proprietary industry corpus for terminology grounding. The AI layer is a tool in our workflow — it does not determine final output quality. Our quality system (expert review + LISA QA audit) remains consistent regardless of which AI engine is used.

Standard projects: 2–5 business days depending on volume and language pair. Compliance-critical projects (FDA filings, CE certification packs): typically 5–10 business days to allow full expert review and QC. Urgent delivery options are available — contact us with your timeline and we will assess feasibility. Our AI-assisted workflow typically reduces delivery time by 40% compared to traditional processes.

Pricing is based on source word count, language pair, content type, and service level (L1–L4). AI-assisted tiers (L1, L2) offer significant savings vs. traditional rates. Expert-led tiers (L3, L4) are priced comparably to traditional high-quality translation while delivering faster turnaround. Repeat projects benefit from compounding TM savings. Upload your files for an instant quote.

Yes. NDAs are standard for all enterprise engagements. Our translators and reviewers are also bound by confidentiality agreements. Files are encrypted in transit and at rest. We do not share client content with third parties. For classified or government-adjacent content, please contact us to discuss our specialized handling protocols.

Yes — and this is where our model creates the most value. Over time, your dedicated term base grows, AI accuracy for your specific content improves, and your project manager builds deep knowledge of your brand, products, and regulatory context. Our average client relationship exceeds 3 years. Our longest active relationships span over 10 years.

230+ languages. Key regulated-market language pairs: ZH↔EN, ZH↔DE (EU CE certification), ZH↔JA (Asia-Pacific), ZH↔KO, ZH↔FR, ZH↔AR (Middle East). For rare language pairs or specialized combinations, please contact us — our network covers most commercially significant language combinations with qualified domain experts.

How this process applies by industry

Semiconductor →Life Science →Aerospace & UAV →Smart Manufacturing →Content Globalization →

How L1–L4 tiers are priced →

Ready to Start