/ for enterprise /
Security, sustainability, scalability. Risk gated on your rules.
wxrks is the translation management system enterprises run their AI on. Content enters from your systems, is translated on the models you choose from your own memories and terminology, gated per segment against thresholds you set, and returned with a record. wxrks supplies the framework. The configuration is yours.
- SOC 2 Type 2 since 2020
- Forrester Wave TMS Q3 2025
- AWS US or EU hosting
- G2 4.6 / 5
/ 300+ companies run translation on wxrks /
- Assurance
- SOC 2 Type 2, held continuously since 2020
- Hosting
- AWS, us-east-1 or eu-west-1, single-tenant available
- Models
- Eight AI providers and eight MT engines, your keys, stateless
- Structure
- Unlimited users and organizational units under one account
/ the structure /
Four pillars, one platform, your configuration
Security
Built for security review. SOC 2 Type 2 since 2020, AWS in the US or Ireland, SSO, SCIM, MFA. Stateless models and no training on your content, written into the DPA.
Security and dataSustainability
A configuration that outlives any model. Memories, terminology and decisions live in your assets, not in a fine-tuned model. Switching provider is an edit to one row.
SustainabilityScalability
Predictable at one page and at a hundred million words. Account, Organization and Organizational Unit keep brands, regions and suppliers apart. No fixed throughput limit.
ScalabilityRisk
Quality as gated risk management. Every segment scored with a named finding and a severity. Above threshold passes and locks. Below it is fixed or routed to a named person.
How the gate runs/ security /
What your security team receives, and what they find
Security review is a precondition, not a phase. Send the questionnaire and we return it while your content runs.
Under NDA
- Current SOC 2 Type 2 report
- Penetration test executive summary
- Policy set with acknowledgement record
- CAIQ-Lite
- Architecture and data flow diagrams
- Subprocessor register and DPA
- Assurance
- SOC 2 Type 2 examined annually by an independent CPA firm, held continuously since 2020. Continuous control monitoring. Customer-requested penetration tests permitted.
- Hosting
- Amazon Web Services only. us-east-1 or eu-west-1 at your election, across multiple availability zones. EU accounts keep databases, keys and content in Europe. Dedicated single-tenant environments on the same release train.
- Isolation
- Row-level security bound to the account, file storage separated by account, all file access logged. Organizational units as a second boundary inside it.
- Identity
- SSO with Okta, OneLogin and Microsoft Entra ID. SCIM provisioning and deprovisioning. MFA. Four roles, permission groups across seventeen categories, scoped to the unit.
- Encryption
- TLS 1.2 or higher in transit with HSTS. AES-256 at rest with keys in AWS KMS. Provider credentials you store are encrypted and cannot be read back.
- AI and content
- Every model runs stateless. No customer content trains any model, by wxrks or by any provider. Provider list published, 30 days' notice of change. Written into the DPA.
- Availability
- 99.9 percent monthly, committed with service credits. Recovery time and recovery point objectives of four hours. Restore tested annually. Public status page.
- Deletion
- Export in TMX, TBX, XLIFF, XLSX and CSV at any time. Deletion within 30 days of a verified request, full purge within 60 days of termination, to NIST SP 800-88 Rev 1.
/ sustainability /
Models change every quarter. The configuration does not.
Your memories, terminology, instructions and reviewer decisions are applied at request time. Nothing is fine-tuned, so nothing is lost when the model changes.
Switch without re-implementation
Changing the model behind an operation is an edit to one row. No retraining, no lost reviewer decision.
A different model per operation
Translate, Proofread, Smells, Fix Tags, Review Assessment and term learning each run where you choose, with reasoning effort set per row.
Tokens spent where you decide
Exact and in-context matches never reach a model. Repeated segments propagate once. A request carries the segment, its matches and its terms, not the document.
Reproducible
Identical inputs yield identical outputs. Run the same reference set before and after a switch and compare edit distance.
AI providers
- OpenAI
- Azure OpenAI
- Anthropic
- Google Gemini
- Google Vertex AI
- Groq
- OpenRouter
- Cerebras
- Any base URL
Machine translation
- Amazon Translate
- Google Cloud Translation
- Microsoft Translator
- DeepL
- Intento
- TransPerfect
- ODIN
- Google AutoML
/ scalability /
The same behaviour for a one-line request and a global launch
One Account holds Organizations, each holding Organizational Units. Settings cascade down and the most specific wins. Each unit has its own memories, glossaries, workflows, models, price list, members and audit trail.
People
Unlimited users. Thousands of requesters, reviewers, linguists and suppliers in one instance, provisioned from your directory through SCIM. Suppliers set up in under a day.
Volume
No fixed throughput limit. Files up to 2 GB. Large work units split across suppliers and reassemble. Connectors create projects, rules route them, quotes and assignments run on arrival.
Visibility
Progress per language at project and portfolio level. Cost accruing to the unit that raised the work. Dashboards by supplier, language, content type, unit and date. Export and API for your own BI.
Connectors wxrks maintains
- Adobe Experience Manager
- Contentful
- Drupal
- WordPress
- Webflow
- Optimizely
- Kentico
- Marketo
- Eloqua
- Knak
- Zendesk
- ServiceNow
- Pendo
- Figma
- GitHub
- GitLab
- Bitbucket
- Phrase
- QuickBooks
- PayPal
Every action has an API call
- REST API
- Webhooks
- CLI
- MCP server
- Pipelines
- Translation Portal
- Vendor Portal
- Work Packages
/ risk /
Risk is gated while translation happens, not after
In a conventional workflow every review round is a human pass, so a file gets one review and whatever it misses ships. Here evaluation is a machine operation. It runs as many times as your rules require, at no added cost or turnaround.
A manager who reads no Thai sees that a page carried two major terminology findings that were resolved and one minor register finding that was accepted, with the reasoning attached.
- 1
Ingest and assemble context
Content arrives through a connector, the API, the CLI or the portal. Relevant context is retrieved semantically from your memories, glossaries, reference material and prior decisions.
- 2
Translate in your voice
One suggestion per segment with your terminology and precedent already inside it, on the model you assigned. A Sous-Chef agent carries the register for that brand or unit. Nothing is trained.
- 3
Score every segment
Smells returns a named finding, a severity from Info to Blocker and a short explanation. Severity becomes a confidence score per segment and an aggregate per file. Error types map onto MQM dimensions.
- 4
Apply your rules
Per content tier, deliver, fix deterministic issues and deliver, or fix what is deterministic and route the rest. Grammar, punctuation and tag fixes apply automatically. Ambiguous source stays open for a person.
- 5
Route to the right reviewer
Open segments go to the most eligible reviewer by subject, sensitivity and load. A reviewer sees the twelve segments that need judgement, each issue named, not four thousand words against source.
- 6
Re-evaluate before release
The completed file is scored again against your delivery thresholds. If it fails, it goes back. The loop adds no cost and no turnaround.
- 7
Deliver, record, learn
Translations return to the source system. Every segment carries its provenance, machine, memory or human, into the exported file. A reviewer's correction changes the next suggestion in the same session.
/ who holds the controls /
wxrks supplies the framework. The configuration is yours.
wxrks does not choose your configuration, does not tune it toward a result, and takes no position on which setting is appropriate for your content. Human review is available at every step and applies where you place it.
Models
Which providers run, for which operations, units and language pairs. Your keys or wxrks defaults. Public models can be restricted for content you tag as sensitive.
Assets
Which memories, glossaries and reference material apply, with what priority, per unit and per workflow step. Owned by you and exportable at any time.
Thresholds
The confidence level per content tier, and what happens above and below it. Custom review categories replace or extend the defaults.
Workflow
The step catalog, its names and order, per account and per unit. Which steps a person performs and who that person is.
Structure
Brands, regions, business units and suppliers as organizational units with their own assets, models, prices, members and audit trail.
Environment
Hosting region, identity provider, retention, single-tenant or shared, and what your security team receives under NDA.
/ evidence /
Measured on production programs
One published study and the results named customers report on their own configurations. What your program sees depends on your content, languages and baseline, which is what the proof of concept measures.
More effective than traditional machine translation post-editing, measured as translation edit rate against a neural MT baseline. Average edit distance fell from 18.23% to 11.17%.
95%
Content through the agentic framework with no human review step, on CCEP's own thresholds. 41,000 employees, 31 countries, 155 languages.
70%
Turnaround reduction across 42 languages, with cost down more than half year on year since 2018.
70%
Lower translation technology spend across 30 markets, with a 40% efficiency gain.
50%Automation AnywhereMore than half off translation cost, requests raised from Microsoft Teams by people who are not translators.
60%
Reduction in time to market on training content.
Improvement over neural MT post-editing, by language
Translation edit rate, same study. English source, 22 June to 23 December 2023. Context-sensitive output outperformed in every language pair tested.
It doesn’t ask us to supervise a machine — it gives our linguists a semantic exoskeleton. It allows us to move from a defensive stance of fixing mistakes to a progressive stance of amplifying creativity.
When localisation is framed primarily as a clean-up step for machine generated output, linguists risk being positioned as fixers rather than experts. The future isn’t a compromise between human and machine but rather, a synthesis.
Companies just run their content through LLMs and think the job’s done. No workflows, no audit trail, no TMs or glossaries, no consistency. And frighteningly frequently: no expert human oversight.
