Every AI language vendor publishes a version of this page, and almost all of them are written as a takedown of human interpreters. Faster, cheaper, always available, no scheduling.

That framing is why buyers in healthcare do not trust any of these pages, including ours by default. So here is the version with the limits in it, written by a company that sells one of the two things being compared. Read it with that in mind, and check the parts that matter against the rule text linked throughout.

The short answer

A qualified human interpreter is the answer wherever the law requires one, wherever the conversation carries clinical or legal weight, and wherever the language is one an automated service does not deeply support. AI phone interpretation is the better answer for routine, high-volume, immediate operational calls in supported languages.

Most practices need both. The mistake is failing to decide at all and letting the front desk improvise call by call.

Where the rule draws its line

This is the part vendor comparisons skip, and it is the most important paragraph in the article.

Under 45 CFR 92.201(c)(1), when interpretation is required a covered entity must offer a qualified interpreter. 45 CFR 92.4 defines a qualified interpreter as an interpreter who has demonstrated proficiency in speaking and understanding English and at least one other spoken language, interprets effectively, accurately and impartially without changes, omissions or additions while preserving tone and sentiment, and adheres to accepted interpreter ethics including confidentiality.

That definition describes a person. An AI service is not a qualified interpreter under it, and no amount of marketing changes that. Any vendor claiming otherwise is describing a product that does not exist.

What the rule does not do is prohibit language technology. It sets a standard for the interpretation required in a given interaction, and it sets quality requirements at 92.201(g) for audio remote interpreting, including real-time full-duplex audio over a dedicated high-speed connection with no lags or degraded quality, and adequate training for the people using it.

Side by side

Three lines in that table deserve expanding.

 

Language breadth. A large human network covers hundreds of languages and can build a relay chain through an intermediate language for speakers of languages of limited diffusion. An automated service covers a small number of deeply tuned languages. Breadth and depth are a real trade-off, and for a clinic whose patient population sits in one or two languages the trade often favors depth. For a hospital serving forty language groups it does not.

 

Ethics and impartiality. A qualified interpreter is bound by professional ethics, including confidentiality. Software is not bound by anything except your contract and your policies. That is an argument for reading the contract, the BAA and the retention terms carefully, which is what our compliance approach sets out on our side.

 

Names, numbers and identifiers. This is where the comparison turns in the other direction. Interpreted calls do not usually break on clinical vocabulary. They break on a member ID, a date of birth or an email address heard once. A service that reads identifiers back for confirmation catches errors that a rushed human exchange does not, which matters most on the scheduling desk where those details are captured.

The three call types

 

Calls that require a qualified interpreter. Informed consent. Diagnosis and prognosis. Medication changes. Behavioral health assessment. Anything a patient must understand in order to exercise a right. Do not put an AI service on these, and be wary of any vendor who suggests you can.

 

Calls where either can work, so decide by policy. Pre-visit instructions, results follow-up that only leads to booking an appointment, general insurance questions. The important thing is that the decision is written down rather than made by whoever picks up.

 

Calls where a per-minute meter is the wrong tool. Reminders, confirmations, reschedules, directions, eligibility checks. Short, frequent, and the calls a per-call minimum charges hardest. This is where an AI service earns its place, and where practices currently lose patients to hold queues and call-back promises.

 

There is a fourth category worth naming: calls neither should take alone. Languages of limited diffusion requiring a relay chain, covered in relay interpretation, and any call where the patient has asked for a particular arrangement.

What AI is genuinely better at

Four things, stated without embellishment.

 

Immediacy. No queue. The service joins a call already in progress, which changes staff behavior because there is nothing to ration.

 

Cost shape on short calls. Flat monthly pricing removes the per-minute meter, and the meter is what makes a forty-second reminder expensive. The arithmetic is in what phone interpretation actually costs a front desk.

 

Consistency of process. Read-back confirmation happens the same way on every call. Human quality varies with the individual, the queue and the hour, in both directions.

 

Coverage of the hours you are open. Evenings, Saturdays, and the two weeks your bilingual receptionist is on leave.

What humans are genuinely better at

Judgement in the moment. A qualified interpreter notices when a patient has not understood, when a family member is answering for them, when the register is wrong. That is professional skill, and it is not replicated.

 

Nuance, register and distress. Behavioral health, bereavement, safeguarding disclosures. Conversations where how something is said carries as much as what is said.

 

Language breadth, including relay. The only route for many patients.

 

Accountability. A person is bound by ethics, can be identified, and can be asked what happened on a call. That matters when something goes wrong.

What a practice running both actually looks like

The written language access procedure required by 45 CFR 92.8(d) is where this gets settled, not in a vendor meeting.

 

Name the conversation types that require a qualified human interpreter, specifically enough for a receptionist to apply without asking. Name the route for everything else. Name what happens when a patient requests a specific arrangement. Name the languages your automated route does not cover and where those calls go instead. Then train the front desk on it and record the training, which the rule requires anyway.

 

Traditional interpreter companies add a third person to the phone call. AI companies replace the employee entirely. Interpreterly does neither. The employee stays on the call and we carry the language, in Spanish and Brazilian Portuguese today, with Haitian Creole coming soon.

 

Interpreterly is a language assistance tool, not a certified interpreter service, and it does not replace a qualified human interpreter where the law requires one. Keeping your interpreter vendor is the correct decision for most practices. What changes is how much routine volume runs through a per-minute meter.

 

If you are evaluating rather than deciding, how to compare phone interpretation providers has the questions to ask, and start a free 30-day trial if you would rather hear it on real calls than read about it.

The question to ask any vendor, including us

Which calls should we not use this for?

 

An answer that names specific conversation types is a vendor thinking about your risk. An answer about how their technology handles everything is a vendor thinking about their quarter. Ask it three times to three companies and the shortlist sorts itself.

Frequently asked questions

Can AI replace human interpreters in healthcare?

No. Under 45 CFR 92.201(c)(1) a covered entity must offer a qualified interpreter when interpretation is required, and the definition at 92.4 describes a person who has demonstrated proficiency and follows interpreter ethics. AI language assistance can support routine operational communication.

 

Is an AI interpreter accurate enough for medical use?

It depends entirely on the use. For scheduling, reminders and eligibility questions, accuracy on names, numbers and identifiers matters most, and read-back confirmation addresses it. For clinical conversations, use a qualified human interpreter.

 

Is AI interpretation cheaper than a human interpreter?

The cost shapes differ. Per-minute billing with a per-call minimum is expensive on short calls and reasonable on long ones. Flat monthly pricing is the reverse. Compare all-in cost per call rather than rates.

 

What are the risks of using AI for interpretation?

The main risks are using it where a qualified interpreter is required, overestimating language coverage, and treating audio quality as somebody else’s problem when 45 CFR 92.201(g) sets requirements for it. All three are avoidable with a written policy.

Should we use both?

Most practices should. Route legally required and clinically weighty conversations to qualified human interpreters, and route routine high-volume calls to whichever option removes the hold queue. Write the split into your language access procedure.