Unbabel

What is Unbabel?

AI-plus-human translation layer purpose-built for multilingual customer support. It sits inside the helpdesk so agents write in their own language and customers receive fluent translated replies, with human editors in the loop where quality thresholds demand it.

Main features

• Widn.AI translation models
• real-time translation inside Zendesk/Salesforce/Intercom tickets and chat
• human-in-the-loop editing triggered by quality scoring
• MT Quality Estimation (COMET-based) that predicts translation quality without a reference
• customer-specific glossaries and tone
• Unbabel LangOps analytics
• multilingual chat and email.

Practical use cases

• Letting an English-speaking support team serve customers in 30 languages
• translating a knowledge base at scale
• meeting SLA in markets without hiring native-speaking agents
• measuring translation quality objectively before a reply goes out.

Who is it for?

• Customer support and CX leaders at international companies; BPOs; e-commerce and travel operations serving many markets.

Pricing, free access & limits

Usage-based (per word or per ticket) with enterprise contracts; quality tier affects price

Plans described in the source

Pricing is quoted per customer and is usage-based - typically per translated word, with rates varying by quality tier (pure MT is cheapest, MT plus human editing costs more) and by language pair. Self-serve entry tiers have existed at low monthly rates for small volumes. Enterprise contracts are annual with volume commitments. There is no public price list - request a quote.

Usage limits to consider

Volume commitments and per-word overage; human-in-the-loop turnaround adds latency compared with pure MT; quality tiers determine cost and speed.

Strengths & tradeoffs

These points summarize the supplied research, rather than hands-on test results.

Strengths noted in the source

• The quality-estimation layer is genuinely differentiated - it predicts whether a translation is good enough to send without a reference translation, which is what lets the human-in-the-loop economics work
• deep, native helpdesk integrations
• support-domain tuning outperforms generic MT on ticket language.

Limitations to consider

• No public pricing at all, which makes evaluation slow
• human-in-the-loop adds latency unacceptable for some live-chat scenarios
• enterprise sales motion with long cycles
• overkill and unaffordable for small teams
• narrow focus means it does not serve document or software localisation well.

Platforms, languages & integrations

Platforms

• Web app plus helpdesk-embedded integrations
• API

Languages

• 30+ languages with strong support-domain tuning; Arabic and French supported

API access

Yes - Unbabel API for programmatic translation and quality estimation

Integrations and exports

• Zendesk, Salesforce Service Cloud, Intercom, Freshdesk, Front, Sitecore, Contentful, custom via API

Evaluation notes

Ask specifically for per-word rates by quality tier and by language pair during the sales process - the blended average hides large variation. Latency on human-edited tiers should be tested against your live-chat SLA.

Explore similar tools

Other tools in AI Translation Tools. Compare their specific features and limits for your task.

Frequently asked questions

Sources & editorial notes

Based on the directory editor’s supplied research. The source does not provide a verification date for this row. Importing a record does not independently verify every product claim.

• Official product website
• Official pricing