Skip to main content
Email addresses are hard to get right over the phone: “B” and “D” sound alike, and “gmail” is easily heard as “gmial”. With Enhanced email capture on, your agent collects the address in a short guided step. The caller spells it, every letter is written down by the platform rather than by the AI, a second listener double-checks it, and the agent reads it back word for word. The address is used only after the caller confirms it.

Turning it on

  • Dashboard: open the agent’s Configure tab, switch the header to Advanced, and turn on Enhanced email capture in the Voice & Audio section.
  • Connect: open the agent’s Advanced Settings tab (available once the agent is detached from its template) and turn on Enhanced Email Capture.
  • Your business email domains (optional): when the switch is on, add up to 5 of your own email domains, separated by commas, for example acme.com. Each domain needs a dot, and only the first 5 are kept; any entry that will not be used is listed under the box. The agent listens for these names and can suggest them when it hears something close.
  • API: set email_capture_enabled and email_capture_domains in the agent’s config. See Create Agent.
Enhanced email capture works on LiveKit agents whose language is English or Multilingual, and the caller spells in English. For other languages the switch has no effect. Notetaker and Translator agents ignore it, so neither the Dashboard nor Connect shows the switch for them.

What the caller hears

When the caller offers an email address, or the agent needs one, the agent starts the step straight away:
The caller can also say the whole address at once (“john at gmail dot com”). If they did not catch a question, they can ask (“sorry, what?”) and the agent repeats it word for word. If they ask why the email address is needed, the agent says “It’s so we can follow up with you by email.” and asks the question again.

How it works

  • Spelling is written down by the platform. It understands letters, NATO words (“B as in Bravo”), “double T”, digits (“four two”, “oh seven”), and words like “dot”, “underscore”, “dash” and “at”. Filler and polite words (“um”, “thank you”) are ignored, and a correction counts: “J O H N, sorry, J O N” and “J O H N, no, J O N” both give jon. A name said as a word stays a word: “Jay at gmail dot com” gives [email protected], while “J, O, H, N” or “Jay, Oh, Aitch, En” is spelling. If the caller says the part before the at sign as a plain word (“john”) and means it, it is accepted, and the read-back confirms it letter by letter. The AI never types or changes the address.
  • A second listener double-checks. A second speech-to-text service hears the same spelling. If the two disagree on one or two letters, the agent asks, for example, “Was that B as in Bravo, or D as in Delta?”. If they disagree more, it asks the caller to spell it again.
  • The domain is checked. It must be well formed and exist, and it is checked for a mail server. If it looks like a typo of a common provider or of one of your business domains, the agent asks “Did you mean …?”. Real domains that only look similar, such as ymail.com or yahoo.ca, are left alone. A well-known provider or one of your business domains said without its ending (“gmail”, “acme”) gets its ending, and a close miss (“gmial”) gets “Did you mean gmail dot com?”. Any other name said without its ending is asked for again, including the ending.
  • Other domains are spelled. A domain that is neither a common provider nor one of your business domains must be spelled, because a name said as a word is only the listener’s guess. The agent asks the caller to spell it one letter at a time, using a word for each letter like “T as in Tango”, then say the ending. A domain that does not exist means a letter was misheard, so the agent asks for it again the same way.
  • The read-back is exact. The part before the at sign is read in groups of four, with digits and symbols said as words. Well-known and business domains are said as words (“gmail dot com”). Any other domain is spelled out letter by letter, with the endings dot com, dot net, dot org, dot edu and dot gov said as words (“t, e, l, i… m, a, t, i… c dot com”). The read-back and the questions above are spoken word for word by the platform.
  • When it does not work out. If the caller says which part of the read-back is wrong, the agent asks again for that part; a plain “no” gets “Which part is wrong: the part before the at sign, or the domain?”. The step stops without guessing the address after two read-backs without a “yes”, after three domains that do not exist, or after too many tries (12 caller turns in the step). The agent then tells the caller it could not get the address and asks how they would like to continue, offering another way it is set up for, such as a transfer, a text message or a follow-up from your team. It does not end the call because of this, and it does not ask for the email address again on that call. The caller can also decline to give an email; the agent then carries on and asks for it again only if the caller offers it.

Sending email to the confirmed address

With Enhanced email capture on, the Send Email function sends only to the address the caller confirmed. If the agent tries to send before an address is confirmed, the send is refused and the agent collects the address first.

After the call

  • A second, slower check. After the call, the caller’s spelling of each confirmed address is listened to again, more carefully. If it sounds different from the confirmed address, the mismatch is saved with the call’s capture statistics (below) and counted in the Feature Usage card’s After-call mismatch rate. The confirmed address is never changed.
  • Statistics, not the address. Each call keeps a short record of every capture: whether it was confirmed, how many read-backs it took, how long it took, whether the two listeners agreed, the mail-server check, whether a suggestion was offered and accepted, and the after-call check. The address itself is not stored in this record.
  • Feature Usage. In the Dashboard, the Feature Usage page has an Email Capture card with the share of captures confirmed, the average time to capture, how often the two listeners disagreed, and how often the after-call check disagreed. A caller who hangs up during the step counts as not confirmed; one who declines to give an email is left out. A dash means there is nothing to measure yet.

Privacy

When PII redaction is on for the agent or for its organization, the second listener and the after-call check are switched off, so the caller’s spelling is not sent to the extra listening services. The spelling, the domain checks and the read-back work as usual.