AI Receptionist Accessibility: Older Callers and Relay Calls

The person who demoed your AI receptionist spoke clearly, at a normal pace, from a quiet room, on a good connection, using words the system was built to expect. Almost nobody who calls a medical practice does all five of those things.

A meaningful share of your callers are over 50, and hearing loss becomes common in that group. Others speak with an accent, or quietly, or with a speech difference following a stroke. Research puts current voice assistants at roughly 50 to 60% accuracy on impaired speech, against the 90 to 95% generally treated as the floor for acceptable performance on typical speech. That gap is not a rounding error, and it lands on the patients least able to work around it.

There is also a legal dimension that vendor material never mentions. For a healthcare provider, communicating effectively with patients who have disabilities is an obligation rather than a courtesy, and it specifically covers how you answer the phone. This post covers where automation genuinely struggles, what the obligation requires, and what to design so that no caller reaches a dead end.

Your callers are older than the demo assumed

Age changes how people manage automated phone systems, and not only through hearing loss. Research on age-related decline in recognising machine speech points to changes in processing the rhythm and stress of speech rather than to hearing thresholds alone. Someone who hears perfectly well can still struggle with a synthetic voice moving at conversational pace.

Published research figures. The gap lands on your most vulnerable callers.

The practical effect is that older callers take longer to respond, pause mid-sentence more often, and are more likely to give a partial answer while they think. A system tuned for speed reads those pauses as the end of a turn and starts talking over the person, which is where the call unravels.

It compounds with the synthetic voice itself. Speech generated by software carries slightly different rhythm and stress from human speech, and those are exactly the cues an older listener relies on most when the audio is imperfect. A voice that sounds fine in a quiet demo can be genuinely hard work over a mobile connection.

Where recognition falls apart

Speech recognition does not fail evenly. It fails at the edges of what it was trained on, and those edges are populated by exactly the patients who most need a phone to work. Understanding where it breaks lets you decide which calls automation should not be trying to handle at all.

It is worth being fair about the comparison. A busy human front desk also mishears surnames, also struggles with a poor connection, and also puts people on hold. The difference is that a person notices confusion and adapts within a sentence, while a system does what it was configured to do until something stops it. Designing that stop is the whole job.

The pattern in that right-hand column is worth noticing. Almost every remedy is either patience or a person. Neither is expensive, and neither is the default configuration.

Relay calls are an obligation, not an edge case

A relay call is one where an operator sits between a deaf or speech-disabled patient and your office, typing what you say and speaking what the patient types. It arrives on an ordinary phone line, opens with the operator explaining the service, and moves in slow, formal turns that do not resemble a normal conversation at all.

A relay call arrives on an ordinary line and behaves nothing like one.

For a healthcare provider this is not a category you may decline to serve. Effective communication with patients who have disabilities is placed directly on the provider, and covered entities are required to respond to relay-service calls in the same way they respond to any other call. A system that hangs up on an operator, or loops them through a menu they cannot navigate, is a compliance problem rather than a usability one.

The safe default is simple. Detect the pattern, stop trying to automate, and route to a person, because none of the efficiency the system was bought for applies to a conversation that runs at typing speed. Nothing about this is legal advice, and a practice with a meaningful deaf or hard-of-hearing population should have its own counsel look at the setup.

Design choices that help every caller

The useful thing about designing for the hardest callers is that almost every choice improves the line for everyone. None of these are accessibility features bolted on the side. They are just better phone design, and most practices could ask for them tomorrow.

None of these are accessibility extras. They are better phone design.

Plain wording is the cheapest of them. A system that asks "are you an established patient with us" will lose people who have never heard the phrase. "Have you been seen here before" gets the same information from everyone. Practices write their phone scripts in the vocabulary of the office rather than the vocabulary of the caller, and automation makes that habit much more expensive.

The biggest single lever is pace. Synthetic voices default to a brisk conversational speed that sounds natural to a 35-year-old and fast to a 75-year-old. Slowing it slightly costs a few seconds per call and lifts comprehension across the whole population.

The second is one question at a time. Systems love to compress: name, date of birth, and reason for calling in a single sentence. Any caller who is thinking, or who did not catch the first part, now has to unpick which piece the system missed. Ask, confirm, then ask again.

The third is reading back. Names, dates, and numbers should be repeated to the caller before anything is booked. This matters everywhere, and it is what turns a misheard surname from a chart problem into a two-second correction. The broader version of that question is covered in what to listen for on a vendor demo.

The route to a person has to survive

Every accessibility failure ends in the same place: the caller needs a human and cannot get to one. So the route to a person has to work without depending on the caller being understood. That means pressing zero always works, saying almost any variation of asking for help works, and silence does not end the call.

Two failed attempts is a reasonable ceiling before the system stops trying. Systems configured to retry three or four times are not being thorough, they are trapping someone. A caller who has already been misheard twice has learned that this will not work, and every further attempt spends goodwill you cannot get back.

There is a related trap in how systems handle a caller who keeps saying the same thing. Repetition is a signal of failure, not of a caller who needs to be prompted more clearly. If the same phrase arrives twice, the right response is a person, not a rephrased question.

Silence deserves its own rule. A caller who says nothing may be deciding, may be deaf, may be waiting on a relay operator, or may have put the phone down to find an insurance card. Treating silence as a hang-up is the most common way these systems fail a vulnerable caller, and the fix is to offer a person rather than to disconnect. How that handoff should be built is covered in designing the escalation path.

Test with the callers who actually struggle

Staff testing tells you nothing here, because your team knows what the system expects and unconsciously speaks to it. Testing with real patients who find phones difficult is uncomfortable and it is the only way to find out what happens.

Recruiting for this is easier than it sounds. Most practices have a handful of long-standing patients who would be glad to help and who fall squarely in the group you are worried about. Ask directly, explain why, and make clear that the system failing is the useful outcome rather than an embarrassment.

Ask two or three willing patients to call and book a routine appointment, then listen to the recordings with them if your state's rules allow it, which is worth checking first in what recording consent requires. What you hear in those calls will change the configuration more than any vendor checklist.

Where to start

Three checks, none of which needs the vendor's help. Call your own line and say nothing for 20 seconds. Call again and deliberately mumble a surname twice. Then have somebody place a relay call through 711 and see how far it gets.

Three checks, and every one of them ends somewhere a caller can use.

Walk the routing once. A relay operator or text relay goes to a person immediately, treated exactly as any other call. Two failed attempts at understanding ends the automated portion. A request for a person, or a pressed zero, transfers without further questions. And everything that does continue reads names and numbers back before acting on them.

None of this makes automation unsuitable for a practice with older patients. It makes the escape routes the part you configure first, rather than the part you discover after a complaint. If you want the size of what your phone line is costing before you change anything, the free Growth Leak Audit works from your own numbers, and the boundaries of what these systems should handle at all are in what an AI receptionist actually does.