Feature · Multilingual

PL / EN / CZ on one number — or three. When which?

Multilingual voice AI sounds like a feature for global brands. Most often it actually solves a CEE-specific problem: PL store also selling to CZ and SK, sometimes DE/UA. We show language detection, why mid-call automatic switching is a bad decision, when per-market numbers are worth it.

CallBotAgent Team30 Apr 2026~7 min readFeature · Multilingual
Phone receiver with three flags radiating: PL, EN, CZ. Below: number +48 ••• 234 and caption '1 number · 3 languages · CEE-first'
TL;DR

Multilingual voice AI in CallBotAgent has two paths: (1) one number serving multiple languages via phone prefix detection — early access for Scale+; (2) separate number per market with dedicated identity — Branded Phone Number per market (Scale+). Choice depends on scale: up to 1 extra market — option 1, above — option 2.

  • PL + 1 extra language = option 1 (one multi-lang number), Scale+ early access.
  • PL + 2-5 EU markets = option 2 (separate number per market), Scale+ Branded Phone.
  • We don't recommend automatic mid-call switching — more harm than good.
  • Language detection based on phone prefix, not speech recognition.
How detection works

Phone prefix = call language. No NLU, no ML.

AI Voice Agent recognizes customer language from phone prefix in incoming/outgoing call, not speech analysis. Deliberately simple — ML language detection in first 2 seconds peaks at ~85% accuracy in best systems, drops to ~70% in field conditions (noise, weak connection). Not good enough for confirmation calls.

Mapping is deterministic: +48 → PL, +44/+1 → EN, +420 → CZ, +421 → SK, +49 → DE. Detection happens in 0.2s — before AI even starts greeting. Customer with +420 hears 'Dobrý den' from second one.

When detection fails: mobile registered in different country (+48 person living in UK speaking English), virtual VoIP, business with Stanford-Cisco prefix. AI defaults to store language with fallback 'May I continue in English?'.

Horizontal language detection flow: incoming call +420 → AI capsule with magnifier 'Prefix detection' → emerald CZ flag 'Dobrý den' (0.2s · 0.4s)
Anti-pattern

Why automatic language switching is a bad decision

Temptation seems obvious: if AI recognizes language from number, it could also recognize from speech and switch dynamically. Conceptually elegant. Operationally destructive.

Three reasons: (1) mid-call switch loses context — AI was collecting info in one language, in second has to recover, +30s call time. (2) Customer disorientation — thinks they're talking to different people, loses trust. (3) Scripts must be perfectly synced in both languages, multiplying error surface.

Product rule: call language is set in first 0.2s and doesn't change. Only exception: customer explicitly asks 'can we switch to English?' in first 5 seconds — then AI restarts in English with fresh context.

One vs N numbers

When one multi-lang number, when separate per market

One multi-lang number for 1 extra market (PL + CZ). Plus: cheaper, simpler maintenance, one marketing campaign. Minus: no local identity (+48 PL calling CZ customer = trust friction), weaker per-market analytics attribution.

Separate per market (Branded Phone Number, Scale+) for 2-5 markets. Plus: local identity per market (CZ customer sees +420 and answers 73% more often than +48), CNAM (caller name display) configurable per market, clean attribution.

Above 5 markets — Enterprise with custom SIP per market, multi-cloud routing. Conversation with our team, not self-serve.

Split-screen comparison: '1 multi-lang number' (one phone with 3 flag chips) vs '3 numbers per market' (three phones, each one flag)
Panel configuration

Multilingual in panel: choice per plan

Plan Starter: 1 call language (usually PL). Multilingual not available. Plan Growth: PL + 1 extra. Plan Scale: PL + 4 + early access DE/FR/UK native. Plan Enterprise: 22+ via telephony layer.

Scripts per language entered separately. Versioning + A/B test in Scale+ allows comparing per-language variants. Cross-sell map per language too (supplement A → cross-sell B in PL, but → C in CZ — depending on market preference).

Conclusions

Four things to remember

1

Language detection = prefix, not ML. +48 PL, +420 CZ. 99%+ accuracy, 0.2s. ML speech: 70-85%.

2

Language fixed at start, not mid-call. Dynamic switch = lost context, customer disorientation, +30s.

3

1 extra market = one number. 2-5 = separate number. Empirical threshold. Local CNAM picked up 73% more often.

4

Multilingual = Growth+ (1 extra), Scale+ (4+). Starter has one language. Enterprise = 22+ via telephony layer.

7-day trial with real credits

Selling cross-border?Test multilingual (Growth+).

1,000 trial credits (~20 confirmation calls) · full access to your chosen plan · cancel anytime. Card required · no charge during trial · auto-activates on day 8.

Cancel anytimeEmail support includedGDPR · EU data residencyCard required · $0 during trial