The multilingual AI customer support agents worth shortlisting in 2026 are Fin (now part of Salesforce), Decagon and Lorikeet, with Sierra, Zendesk AI agents, Ada, Maven AGI and NiCE Cognigy close behind for specific setups. Published coverage ranges from 45+ languages (Fin) to 100+ (Cognigy), but the number that matters is how each agent detects language, which channels each language works on, and what happens when a customer writes in a language you have not set up.

This guide compares the eight on those questions, using only what each vendor publishes.

Key takeaways

• Language counts are not like for like. Zendesk lists 80 languages its AI agents fully support and says others may work without guaranteed fluency (Zendesk); Ada supports 60 on chat and email but only a subset on voice (Ada docs).

• Detection rules differ. Fin detects language once per conversation and keeps it (Fin help); Decagon and Sierra say their voice agents switch language mid-conversation.

• Demand is real. 22% of US residents aged 5 and over spoke a language other than English at home in 2017 to 2021 (US Census Bureau).

• Per-resolution pricing does not charge by language at Fin ($0.99 per outcome) or Lorikeet ($0.99 Start, $0.90 Scale per chat, email or SMS resolution).

• Test in your top non-English languages before launch, on the channels you actually run, with real tickets.

Which multilingual AI customer support agents are worth shortlisting?

Three products cover most buyers. Lorikeet runs the same agent across email, voice and chat, and its customer Taptap Send supports 15 languages across more than 80 receiving markets (Taptap Send story). Fin publishes 45+ languages and can answer in any configured language from content written in one language using real-time translation (Fin help). Decagon says its voice agents support 70+ languages with automatic detection and language switching (Decagon Voice). If you already run Zendesk, its AI agents list 80 fully supported languages and are the lowest-friction start (Zendesk).

How we chose and checked these tools

We included AI agents that resolve customer conversations (not translation plug-ins for human agents) and that publish language support on their own website or documentation. We ranked them by fit for a support leader asking for multilingual AI agents: breadth of published languages, how clearly the vendor documents detection and fallback, coverage on voice as well as text, and whether pricing is published. Every vendor fact was checked on the vendor's own pages on 30 October 2026, and each claim links to its source. Where a vendor does not publish something, the table says "Not published".

Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim.

The competing guides we read, including Fin's own ranking (fin.ai), rank their publisher first. We have put Lorikeet third because two vendors publish broader language coverage than we do.

Comparison table

1. Fin (now part of Salesforce): broadest self-serve setup for Intercom teams

Fin is the AI agent Intercom built; Salesforce completed its acquisition of Fin on 10 September 2026 (Salesforce). Fin's own guide lists 45+ languages across chat, email, voice, SMS, WhatsApp and social channels (fin.ai). Its help centre explains the mechanics: you pick supported languages in settings, Fin searches content in the customer's language first, and with real-time translation turned on it answers in any supported language from content written in one (Fin help). You can also force a language per customer, which suits region-specific or compliance-driven requirements.

Best for: teams on Intercom, or on another helpdesk Fin supports, who want to add languages themselves without a vendor project.

Limits: Fin's help centre says it detects the language once per conversation and keeps all answers and workflows in that language, and it only responds in languages you have configured. Pricing: $0.99 per outcome with a 50 outcome monthly minimum (Fin pricing).

2. Decagon: widest published voice language list

Decagon's voice page says its agents support customers in 70+ languages with automatic detection and language switching, run on Chord, its own speech model, and understand callers across accents and background noise (Decagon Voice). The same page describes outbound calling and handoff to human agents with a summary, and agent behaviour is governed by Agent Operating Procedures (AOPs), policies and brand guidelines.

Best for: large consumer brands with heavy phone volume in many languages and the budget for an enterprise rollout.

Limits: Decagon does not publish pricing, and its voice page does not break the 70+ figure down by channel, so ask which languages are supported on chat and email as well as voice. Pricing: Not published.

3. Lorikeet: multilingual resolution for regulated and complex support

Lorikeet runs one agent across chat, email, voice and SMS, with workflows that call your own systems to complete tasks rather than only answer questions. Remittance company Taptap Send uses it across email, voice and chat with 15 languages supported, on European-hosted infrastructure, and the agent applies the right policy for each geography (Taptap Send story). Channels can greet each customer in their own language from a language field on the customer profile (docs). For voice, Lorikeet recommends a separate agent per language, with a multilingual mode for callers who switch languages within one call (voice page).

Best for: fintech, health and insurance teams that need the same guardrails and actions in every language, plus published per-resolution pricing.

Limits: Lorikeet does not publish a language list, so confirm your languages on a trial. Its docs say multilingual voice mode adds detection latency on every turn (voice quickstart), and conversation summaries are generated in English only (SDK FAQ). Pricing: Start $2,100 a month and Scale $5,100 a month paid annually, plus $0.99 or $0.90 per chat, email or SMS resolution and $1.50 or $1.20 per voice resolution; escalations are not charged (pricing).

4. Sierra: enterprise brands that want voice and chat from one agent

Sierra's product page says its agents create consistent experiences across voice, chat, email and WhatsApp in 59 languages (Sierra). Its voice page says the agent speaks over 55 languages with mid-conversation language switching, handles interruptions and accents, and routes each turn to the model best suited to the language and tone (Sierra Voice). Sierra also takes card and ACH payments over the phone through PCI Level 1 infrastructure, which matters for billing calls in any language.

Best for: large enterprises that want one vendor for voice and digital channels with payments on the call.

Limits: no published price, and the two pages give different counts (59 across channels, over 55 for voice), so ask for the list per channel. Pricing: Not published.

5. Zendesk AI agents: the default if you already run Zendesk

Zendesk lists 80 languages that its AI agents fully support for use cases, generative replies and agentic AI, and says languages outside the list may work but fluency is not guaranteed (Zendesk). The same article says transliteration is not supported: Russian has to be typed in Cyrillic, for example. An admin must add and activate each language; some languages get automatic AI translation of the agent's replies, others need manual translations before they can be activated (Zendesk setup).

Best for: Zendesk Suite customers who want multilingual automation in the helpdesk they already pay for.

Limits: manual-translation languages need admin work before launch, and customers who type in Latin script for a non-Latin language will not be understood. Pricing: not stated on the languages pages; check your Zendesk plan.

6. Ada: clear documentation on language fallback

Ada's documentation says its AI agent supports 60 languages on web chat, messaging and email, with responses generated natively in the customer's language and knowledge ingested in any of the 60 (Ada docs). It is unusually explicit about edge cases: English is always enabled and is the fallback when the agent cannot work out the language, email conversations always start in English until the customer replies in another language, and voice supports a subset of the 60.

Best for: mid-market and enterprise teams whose volume is mostly chat and email.

Limits: narrower voice coverage than text, and the Ada dashboard is English only. Pricing: Not published.

7. Maven AGI: voice in any language on an existing contact centre

Maven AGI says Maven Voice resolves support calls in real time, in any language, across any channel, integrates with your contact centre stack, handles interruptions, and adapts across accents (Maven Voice). It hands off to human agents with full context.

Best for: enterprises adding AI voice to an existing contact centre platform.

Limits: "any language" is not a list; ask for the languages that have been tested in production on your channels. Pricing: Not published.

8. NiCE Cognigy: largest published language count

Cognigy, now branded NiCE Cognigy, says its platform supports 100+ languages out of the box and 110+ prebuilt integrations (Cognigy). It is built for large contact centres running many agents and flows.

Best for: enterprise contact centres with a team to build and run flows.

Limits: the homepage does not say which channels or features each language covers. Pricing: Not published.

How many of your customers need another language?

Probably more than your ticket tags show. The US Census Bureau reports that 22% of people aged 5 and over spoke a language other than English at home in 2017 to 2021; about 61% of them spoke Spanish, and 62% also spoke English "very well" (US Census Bureau). That last figure is the case for letting customers choose: many bilingual customers will write in English but switch when a problem gets complicated.

In Europe, the EU has 24 official languages (European Union), and Eurostat found that in 2022 a quarter (25.3%) of working-age adults knew no foreign language (Eurostat). An English-only agent cannot serve those customers at all.

Why do language counts mislead?

Because each vendor counts something different. Zendesk counts languages its AI fully supports, Ada counts languages on chat and email, Decagon's 70+ sits on its voice page, and Maven says "any language". Regional variants add another layer: the W3C notes that language tags such as fr-CA combine a language subtag with a region subtag, as defined in BCP 47 (W3C), so "French" may or may not mean Canadian French in a vendor's settings. Ask for a matrix of languages by channel and by feature (answers, actions, voice, knowledge ingestion) instead of one number.

Why is multilingual voice harder than chat?

Because timing is unforgiving in every language. A study of turn-taking across 10 languages, published in PNAS, found that speakers in all of them avoid overlapping talk and minimise silence between turns, with average gaps that vary by language only within 250 milliseconds of the cross-language mean (Stivers et al., PNAS 2009). Detecting the language on every turn adds delay, which is why Lorikeet's docs recommend a fixed language per voice agent unless callers truly switch mid-call. Vendors that switch mid-call, such as Sierra and Decagon, are worth testing on your own recordings.

What still needs a human?

• Complaints, hardship and vulnerable customers. Route these to people, and make sure escalation keeps the customer's language and the full transcript.

• Legal and regulated wording. Disclosures and terms should be translated and signed off by people who read the language, not generated freshly each time.

• Quality review in each language. Someone fluent should review samples in every live language; a blended quality score hides the weakest market.

• Unsupported languages. Decide the fallback in advance: English, a human queue, or a message saying which languages you support.

Questions to ask before you pick

1. Which languages are supported on each channel we run, and for answers, actions and voice separately?

2. Does the agent detect language once or on every message, and can we force a language per customer?

3. If our knowledge base is only in English, does the agent translate it, and how is product terminology handled?

4. What happens when a customer writes in a language we have not enabled?

5. Can we test real tickets in our top three non-English languages before launch, and see results per language after?

6. Does pricing change by language, channel or seat?

For a step-by-step rollout plan, see how to offer multilingual AI support; for voice timing, see our voice latency guide; and if you are weighing Fin or Zendesk's own agents, see Fin alternatives and Zendesk AI compared. To test your own languages on real tickets, start a 30-day free trial.