Practice guide

Communicating with multilingual families: what a translate button does not solve

Machine translation solved the easy 80% of this problem about five years ago. The remaining 20% is where districts get into trouble, and it is almost entirely about terminology, register and review rather than about the engine.

Last reviewed 2026-08-04 ยท Kastr is pre-launch; we publish dated status rather than logos.

School terms a generic engine renders badly for US-district Spanish
EnglishCorrect US-district SpanishWhat a generic engine often producesWhy it matters
School districtdistrito escolarcomarca escolar, junta escolarNames your legal entity. Families search for it.
School counselorconsejero escolarabogado, asesor legalSuggests a lawyer. Alarms families.
Parent-teacher conferenceconferencia de padres y maestrosconferencia de prensa, reunión de la facultadReads as a press event rather than an appointment.
Attendance noticeaviso de ausenciasaviso de asistenciaInverts the meaning — reads as praise for attending.
Early dismissalsalida tempranadespido anticipadoDespido means being fired from a job.
Shelter in placerefugiarse en el lugarrefugio en el sitioEmergency instruction. Ambiguity is unacceptable.
Free and reduced lunchalmuerzo gratis o a precio reducidoalmuerzo libreLibre means unrestricted, not without cost.
Field tripexcursión escolarviaje de campoLiteral and meaningless in context.
Guardiantutor legal / encargadoguardiánGuardián reads as a security guard.
Grade (year group)gradocalificaciónConfuses year group with a mark.

Register is the part nobody plans for

US school communication has a house style: direct, slightly formal, instruction-first. Machine translation preserves meaning and discards register, so an English message that reads as calm and official can arrive as either bureaucratic or oddly casual depending on the target language.

The effect is largest in languages with grammatical formality distinctions. Spanish versus usted is the obvious case: a district notice addressed informally reads as presumptuous to some families and friendly to others, and the engine picks without knowing which. Vietnamese, Korean and Arabic each carry their own version of this problem, and in each case the engine's default is a guess.

You cannot fix this per message. You fix it once, by deciding your district's convention — formal address, instruction in the first sentence, no idiom — and by having a human who speaks the language read the output for the message types that matter. That is a staffing decision more than a software one.

Which messages should never go out machine-translated unreviewed

Two categories, and they are not negotiable:

  • Emergency and safety instructions. Lockdown, shelter in place, evacuation, reunification. The cost of an ambiguous verb is measured in minutes at the worst possible moment. Pre-translate these as approved templates, reviewed by a human, before the day you need them — not in the moment.
  • Legally operative notices. Truancy notices, special-education procedural safeguards, discipline notifications, anything with a deadline or a right attached. These frequently have statutory language requirements in your state, and a machine rendering may not satisfy them regardless of quality.

Everything else — newsletters, event reminders, schedule changes, general updates — is a reasonable use of machine translation with a spot-check routine rather than a full review.

The preview step is the whole point. Kastr renders your draft into up to five target languages through DeepL inside the composer, before the message exists as a send. That turns a review from a process nobody has time for into a two-minute read by the bilingual staff member who is already in the building. What it does not include: a district glossary to pin preferred terminology, and a "see original" footer on the delivered message. Both have been claimed on our own older marketing pages and neither exists.

How many languages, honestly

There is no single true number, so here are all of them. Kastr's composer exposes nine target languages. The family profile offers ten. The migration importer maps thirteen. DeepL's own supported target set is around thirty. Translation is DeepL-powered, and the honest way to state it is by surface rather than by a headline count.

This matters because the category is full of large round numbers. A previous version of our own site claimed 47 languages, which was unsupportable, and we removed it. If a vendor quotes you a specific large figure, ask two questions: which languages, and where in the product can a staff member select them. The gap between "the engine supports it" and "a user can pick it" is where those numbers come from.

For a language your platform cannot handle, the answer is a plan rather than a feature: identify the households, decide which message types get human translation, contract a translator on retainer for those, and record the household's language properly so the gap is visible instead of silent.

A language-access plan that fits on one page

  1. Count your home languages properly. Not from the ELL programme — from household language on the enrolment record. Those are different populations, and the second is the one you must communicate with.
  2. Set a threshold. Many districts use something like 20 households or 5% of enrolment as the line above which a language gets full translation of every message. Below it, targeted translation of essential notices.
  3. Pre-translate the emergency set, reviewed by a human, stored as templates. Six to ten messages covers most of what you will ever need at speed.
  4. Name a reviewer per language. A person, not a role. If you cannot name one, that language is machine-only and you should know it.
  5. Decide the channel per language group. Reading fluency and speaking fluency diverge, so voice in the home language sometimes reaches a household that text does not. Ask; do not assume.
  6. Record preferred language on the contact record, and check it against what each send actually went out in. That comparison is one of the five reach numbers worth reporting to a board.

Questions people actually ask

How many languages can we actually send in?

In Kastr, the composer exposes nine target languages, the family profile ten, and the importer maps thirteen; DeepL's supported target set is around thirty. There is no single accurate number, which is why we describe it as DeepL-powered rather than quoting a count. Ask any vendor which languages and where in the interface they are selectable.

Is machine translation good enough for a school message?

For routine messages, generally yes, with a spot-check routine. For emergency instructions and legally operative notices, no — not because the quality is poor but because the failure cost is asymmetric. The workable middle is machine translation with a human preview step, which is why previewing the translation before the message sends matters more than the engine's benchmark scores.

Which messages should never be machine translated?

Emergency and safety instructions, and legally operative notices such as truancy letters, special-education procedural safeguards and discipline notifications. Pre-translate the emergency set as human-reviewed templates before you need them, and check your state's requirements on statutory notices — some specify how translation must be produced.

What do we do for a language the platform does not support?

Record the household language accurately so the gap is visible, decide which message types those families get in their language, and contract a human translator for that subset. The failure mode is not the missing language — it is a district that never counted the households and therefore never noticed they receive nothing they can read.

Does translation store our families' personal information?

In Kastr, translations are cached against a hash of the source text, target and source language, so the same string is never paid for twice. The cache holds no index of who a message was about or who it went to — a row is retrievable only if you already possess the exact source text. If the translation provider is unavailable, the message passes through untranslated rather than failing silently, and the target language is still recorded so reporting stays accurate.

One price. Every feature. Locked for three years.

$3.50 per student per year under 5,000 students. No tiers, no add-on modules, no per-message fees. Published on the site because you should not have to book a call to learn a price.