"No, that's M as in Mary": why the best voice agents know when to stop talking

Anyone who has called an automated line knows this moment. You say your last name. The system reads back something close but wrong. You spell it. It gets one letter wrong. You spell it again, slower, using "B as in boy." Three or four attempts later you either give up and accept the typo, or you hang up.

Few things make a caller lose patience faster, and it's usually over the one detail that matters most. A misspelled name breaks the confirmation email. A wrong digit in a policy number sends the claim to the wrong file. A mangled street name sends a technician to the wrong house.

The fix isn't a voice agent that insists harder. It's a voice agent that knows when voice is the wrong tool for the moment, and offers the caller an easier way to finish.

Why some details are hard to get right by voice

Speech recognition has improved enormously, and most of a call goes fine. The trouble concentrates in a few places:

• Names. There's no dictionary for them. "Aoife," "Nguyen," "Siobhan," and "Kaczmarek" don't sound the way they're spelled, and plenty of common names have several spellings: Jon, John, Jonn.
• Letters that sound alike. B, D, E, P, T, and V blur together on a phone line, and so do M and N, and F and S. Phone audio drops most of the high frequencies that tell them apart.
• Accents. Callers speak with every accent there is, and so does whoever set up the agent. A detail that's perfectly clear to the caller can be ambiguous to the listener, human or not.
• Background noise. Kitchens, truck cabs, job sites, and busy streets are where a lot of real calls happen. Noise eats consonants first, which are exactly what spelling depends on.
• Long strings. Email addresses, reference numbers, VINs, and policy numbers have no meaning to anchor on. One missed character and the whole thing is wrong.

Human receptionists run into the same problems. The difference is that a good one notices when it isn't working and says, "Tell you what, just text it to me." A voice agent should do the same.

First, say it right: pronunciation dictionaries

Getting names right cuts both ways. An agent that mispronounces your business name, your street, or the product a caller is asking about sounds careless, and callers notice. A pronunciation dictionary fixes that. You list the words the agent is likely to get wrong, like your company name, local place names, staff names, product names, and industry terms, and tell it how each one is said.

It's a small thing that makes a big difference. A clinic in Worcester wants "WUSS-ter," not "war-SESS-ter." A law firm wants the founding partner's name said the way the partner says it.

Know when to stop spelling

Confirming details is good practice. Reading back a spelling or a phone number takes seconds and prevents a failed follow-up. But confirmation has limits. If the first read-back is wrong and the second is still wrong, a third try over the same noisy line is unlikely to go better, and the caller's patience is already running out.

A well-crafted agent treats that as a signal to change channels, and it doesn't wait for things to go wrong when it can see them coming. Some details are worth moving to text before anyone tries to spell them:

• Email addresses, almost always.
• Long reference, policy, account, or order numbers.
• Unusual names and street names, especially on a noisy line.
• Anything the caller is reading off a card or a document anyway.

"That's a tricky one to get right over the phone. Can I text you a quick link so you can type it in?"

That one sentence turns a frustrating loop into a five-second task. The caller types it in, the agent reads it straight from the reply, and the call carries on.

Continue by text or email, while the caller is still on the line

Moving to text or email isn't only for spelling. Once the caller has a message from you, it opens up things that voice can't do at all:

Type the detail instead of saying it

An email address, a name, or a reference number, typed exactly as the caller has it. No read-back needed, and nothing to mishear.

Send a photo or a document

Some things are easier to show than describe. A photo of an insurance card or a driver's license. The leak under the sink, or the model label on the furnace. The router's blinking lights. A receipt, a damaged item, a signed form. The caller snaps it and sends it, and it lands on the same record as the call, so your team isn't matching up attachments later.

Finish a form from a link

Some processes simply need a form: a longer intake questionnaire, an authorization, a signature, a payment. Reading twenty questions aloud is slow for everyone. The agent can cover the part that benefits from a conversation, then send a link so the caller can finish the rest on their phone, in their own time or right away while the agent waits.

The call keeps doing what it's good at: understanding what the caller needs, answering questions, and deciding what happens next. The text or email handles what voice is bad at. Everything ends up on one structured record, whichever way it arrived.

Let the caller choose

The point isn't to push everyone off the phone. Some people would rather spell their name than open a text, and some can't easily do both at once because they're driving or their hands are full. A good agent offers, and takes no for an answer.

• Offer, don't insist. "Would it be easier to text it?" is a question, not a redirect.
• Let callers pick text or email. Some people don't want texts from businesses, and some are calling from a landline.
• Keep the conversation going. The caller can stay on the line while they reply, and the agent confirms once it arrives.
• Don't lose what was already said. Anything collected by voice stays on the record. The text or email only fills the gaps.
• Have a fallback. If the caller can't reply right now, the agent takes what it has and can still send the link so they can finish later.

Texting also has rules of its own. Ask on the call, send only after a clear yes, say who the message is from, and make it easy to stop. We cover that in consent, disclosure, and trust, and our own approach is in our SMS terms.

What it sounds like on a real call

Here's a home-services call on a construction site, with a compressor running in the background:

1. The caller explains the problem, and the agent asks the usual questions: what's wrong, how urgent it is, and where the job is.
2. The street name comes through garbled. The agent reads it back, and the caller corrects one part of it.
3. Rather than try again, the agent offers: "I want to make sure the tech finds you. Can I text you a link to type the address and send a photo of the unit?"
4. The caller says yes, taps the link, types the address, and sends a photo of the model label. The agent confirms both have arrived.
5. The agent books the visit, and the technician's record has the exact address, the model number from the photo, and a summary of the call.

Nobody spelled anything more than once. The caller spent less time on the phone, and the business got better information than a perfectly transcribed call would have given it.

Test it before your callers do

You can check all of this in a few minutes. Call your agent and:

1. Listen for mispronounced names and places, and add them to the pronunciation dictionary.
2. Give it a name it's likely to misspell, and see whether it offers to take it by text after a failed read-back.
3. Call from somewhere noisy, like a car, a kitchen, or a street.
4. Ask to send a photo, and check that it lands on the call's record.
5. Say no to the text, and make sure the agent carries on by voice without pushing.

Testing like this is part of how HiloVoice agents are created. Pronunciation dictionaries, continuing by text or email, photos and documents, and links to forms are all built in, and you choose when your agent offers them.