Key takeaways
- Language need is large: 9.0% of US residents aged 5 and over speak English less than very well[1].
- AI can help with administrative calls in a patient's language, such as hours, booking and reminders, if the practice tests it in each language.
- Section 1557 rules still require a qualified interpreter when interpretation is needed[2], and human review of machine-translated critical content[2].
- Design the line so any clinical or complex call in another language goes straight to a qualified interpreter, with no extra steps for the patient.
How many patients need language support
More than one in five people aged 5 and older in the US, 22%, spoke a language other than English at home in 2017 to 2021[3]. Not all of them need language help on the phone, but many do: 9.0% of US residents aged 5 and over speak English less than very well, according to the 2024 American Community Survey[1]. In some communities the share is far higher.
Spanish dominates nationally. Among people who spoke a language other than English at home, about 61% spoke Spanish[3]. The remaining share is spread across many languages, which is exactly why staffing a bilingual front desk for every patient is hard, and why AI language support is attractive.
- Spanish 61%
- All other languages 39%
Source: U.S. Census Bureau, 2017-2021[3]
Show the data
| Item | Value |
|---|---|
| Spanish | 61% |
| All other languages | 39% |
Physicians are already experimenting. In the AMA's 2026 survey, 18% of physicians said they use AI for translation services[4]. The question for the front office is where that kind of support is appropriate on a phone line, and where it is not.
What federal rules require
Section 1557 of the Affordable Care Act applies to health programs and activities that receive federal financial assistance from HHS, among others; check with counsel whether and how it covers your practice. Its rules set two requirements that shape any AI language tool.
- Covered entities must offer a qualified interpreter when interpretation services are required for patients with limited English proficiency[2].
- If a covered entity uses machine translation for critical or technical patient content, that translation must be reviewed by a qualified human translator[2].
- Covered entities must also make reasonable efforts to identify and mitigate discrimination risks from patient care decision support tools, including AI, that use factors such as national origin[5].
In plain terms: AI can help a patient book a visit or hear the office hours in their language. It cannot stand in for an interpreter when the conversation is about their health, and a machine-translated consent form or after-visit instruction needs a qualified human check before it is used.
Where AI language support fits
| Interaction | AI language support? | What else is needed |
|---|---|---|
| Office hours, directions, holiday closures | Appropriate | Answers reviewed in each language before go-live |
| Booking, rescheduling, confirming | Appropriate with testing | Read-back of date and time in the patient's language; text confirmation in that language |
| Appointment reminders | Appropriate | Templates translated and reviewed by a qualified translator |
| Insurance and billing questions | Capture and route | Staff or interpreter for anything beyond capture |
| Symptoms, medications, results | Not appropriate | Immediate route to a qualified interpreter with clinical staff |
| Consent, financial policy, care instructions | Not appropriate for unreviewed machine translation | Qualified human translation or review |
Designing a multilingual phone line
Offer the choice early
The first seconds of the greeting offer the most common languages in your patient population, in those languages, and the AI also responds if a caller simply starts speaking another language.
Confirm the language
The assistant confirms the language it will use and gives the option to switch or reach a person.
Stay administrative
In every language, the assistant handles only the tasks approved for that language, using answers your practice has reviewed.
Escalate to interpreters fast
Any clinical topic, any confusion and any request for a person triggers a warm transfer to staff with interpreter access, passing the language and reason so the patient does not repeat it.
Confirm in writing
Bookings and changes are confirmed by text in the patient's language using reviewed templates.
Record the preference
The patient's preferred language is saved to the record so staff, reminders and interpreters are ready next time.
Greeting opening, example for an English and Spanish line: “Thanks for calling. You've reached the practice's automated assistant. Para español, diga español.” Then, in Spanish: “Puedo ayudarle con citas y preguntas sobre la oficina. Para hablar con una persona y un intérprete, diga persona.”
Before you switch on another language
Decide which languages to support from your own data, not national averages. Pull preferred language from your practice management system, count interpreter requests from the last few months, and ask front desk staff which languages they hear on the phone that the system does not record. The languages that appear most often are your first candidates; everything else routes to staff with interpreter access from day one.
Then make sure the human side is ready, because the AI line will send more callers to it, not fewer, once patients learn they can reach the practice in their language.
- Confirm your telephone interpreter service can be conferenced into a live transfer, and test the connection from the AI line, not just from a desk phone.
- Agree how staff will see the caller's language before they pick up, so they can start the interpreter connection straight away.
- Check that bilingual staff who take calls in another language have had their language skills assessed, since qualified interpretation is a defined standard, not just fluency.
- Have reminder and confirmation templates translated and reviewed by a qualified translator before the first patient receives them.
- Prepare a short note for patients, in each language, explaining what the assistant can help with and how to reach a person and an interpreter.
Finally, give the plan an owner. Language access tends to slip when it belongs to everyone; one named person who reviews non-English calls and keeps translated answers current makes the difference between a feature that works and one that quietly degrades.
Testing in every language you offer
An AI that performs well in English can perform much worse in another language, especially with regional accents, mixed-language speech and medical vocabulary. Test each language separately, with people who speak it natively.
- Run fictional calls for each approved task in each language, including callers who switch languages mid-sentence.
- Include emergency phrases in every language and confirm the fixed emergency message plays correctly in that language.
- Have a qualified reviewer check every approved answer and template in each language before go-live.
- Test names, addresses and dates of birth read back in each language; these are where mishearing happens.
- Include non-English calls in your weekly review sample, reviewed by someone qualified in that language.
- Retest after vendor updates; language performance can change without notice.
Where it goes wrong
- The AI keeps a patient in an automated flow when they need an interpreter, because the clinical trigger words were only set up in English.
- Machine-translated reminders contain an error nobody reviewed, such as the wrong day or location.
- A family member interprets on a call about a clinical issue because it was easier than connecting an interpreter. Qualified interpreters exist for good reasons, including accuracy and privacy.
- The language preference is captured but never reaches the record, so the next call starts in English again.
- Reviewers cannot understand non-English calls, so errors in those languages go undetected for months.
Example: a fictional community practice
Example: a fictional community health practice serves many Spanish-speaking and some Vietnamese-speaking patients. It turns on AI language support in Spanish for office questions, booking and reminders, after a qualified translator reviews every answer and template. Vietnamese callers are routed straight to staff with telephone interpreter access while the practice evaluates the AI's Vietnamese performance.
Example, continued: in week two, the Spanish-language review finds that the assistant translated the clinic's Saturday walk-in hours using the wrong time format. The answer is fixed and the template rechecked. The practice also discovers that Spanish clinical keywords were missing one common phrase for shortness of breath, adds it, and reruns its emergency test calls in both languages. After two months, the share of Spanish-speaking callers who reach their goal without waiting for a bilingual staff member has grown, and staff interpreter calls are focused on clinical conversations.
Measuring language access
- Calls by language and outcome, compared with English calls.
- Time to reach an interpreter when one is needed.
- Clinical calls in other languages that stayed in an automated flow, which should be zero.
- Errors found in non-English calls during review.
- Patients whose preferred language is recorded correctly.
For error handling across all languages, see when AI gets it wrong. Using call data to spot language gaps is covered in call analytics.
Frequently asked questions
Can AI replace medical interpreters on phone calls?
No. Section 1557 rules require a qualified interpreter when interpretation is needed for patients with limited English proficiency. AI can help with administrative calls in other languages, but clinical conversations need qualified interpreters.
Can we use machine translation for patient documents?
Under Section 1557 rules, machine translation of critical or technical patient content must be reviewed by a qualified human translator. Consent forms and care instructions are clear examples.
Which languages should a practice offer on its AI phone line?
Start with the languages most common among your own patients, test each one with native speakers, and route all other languages directly to staff with interpreter access.
References
U.S. Census Bureau (2024). GCT1603: Percent of People 5 Years and Over Who Speak English Less Than 'Very Well' (2024 ACS 1-year geographic comparison table) (opens in a new tab) ↑
Legal Information Institute, Cornell Law School (Electronic Code of Federal Regulations text) (current). 45 CFR § 92.201 - Meaningful access for individuals with limited English proficiency (opens in a new tab) ↑
U.S. Census Bureau (2017-2021). New Data on Detailed Languages Spoken at Home and the Ability to Speak English (press release CB25-TPS.40) (opens in a new tab) ↑
American Medical Association (2026). More than 80% of physicians use AI professionally: AMA survey (opens in a new tab) ↑
Legal Information Institute, Cornell Law School (Electronic Code of Federal Regulations text) (current). 45 CFR § 92.210 - Nondiscrimination in the use of patient care decision support tools (opens in a new tab) ↑