Frequently asked questions

No. Translation moves between languages. Lahja moves between accents within the same language — an Egyptian agent’s Arabic becomes Saudi Arabic that a Saudi customer would actually use. It changes vocabulary, phrasing and register, not meaning.

No, and there is no setting that enables it. Every reply is shown to an agent, who can edit it freely; the exact text on screen when they press send is what the customer receives.

Egyptian, Saudi Neutral, Gulf Neutral, Emirati, Modern Standard Arabic, English (US) and English (UK) ship by default. You can add more — a code, a name, and the wording rules for it.

Free text you write per accent: the wording your brand uses and the wording it never uses. It is sent to the AI with every single transformation for that accent, so the output follows your house style rather than generic dialect.

Yes. When an agent edits a reply before sending, the approved wording is stored for that accent and fed into later transformations, so the same correction is not needed twice.

They are checked explicitly. After each transformation Lahja verifies that digits, order codes, links and phone numbers survived intact — treating Arabic-Indic and Latin digits as equivalent — and warns the agent if anything looks altered.

Instagram Direct, Facebook Messenger and WhatsApp Business. All three use official Meta APIs — no scraping and no shared passwords.

An Instagram professional account, a Meta developer app, and “Allow access to messages” enabled in Instagram’s connected-tools settings. No Facebook Page is required. You can connect an account you own straight away; serving other companies’ accounts needs Meta’s App Review.

Meta’s rule, not ours: a business may reply freely for 24 hours after a customer’s last message. Outside that window the platform rejects the send, and Lahja shows you the rejection instead of marking it delivered.

Neither is possible today. TikTok has no public messaging API, and Snapchat’s only messaging endpoint is for brand-to-creator collaboration, not customer service. We would rather say so than list them as “coming soon”.

It is metered per workspace — tokens, latency and USD cost per call, visible in Reports. A typical reply on the fast setting costs a fraction of a cent.

Yes, in Settings → AI cost & speed. Review detail controls how much the AI writes (reply only, plus a change list, plus a back-translation), while Model trades dialect nuance for speed. Fast settings answer in around two seconds; the fullest setting takes noticeably longer because it writes far more.

Yes. Turn off “Translate incoming customer messages” and no model call is made for inbound messages at all — the accent is still detected locally, at no cost.

Yes. Settings → “Agents see” can restrict agents to conversations assigned to them plus unassigned ones. Supervisors, admins and owners always see everything.

Team → invite by email and role. They set their own password through a link that expires in 7 days and can be revoked at any time.

No. See the privacy policy.

No. Every query is scoped to one workspace, and a connected channel account is locked to a single workspace by a database constraint, so two businesses cannot receive the same customer’s messages.

See Delete your data. Requests are completed within 30 days.