AI in Inbound CX

Every AI feature here is off unless an organization turns it on.

The AI assistant block, in a call flow

An organization can put a block in a call flow that talks to the caller, tries to answer them, and hands over to a person when it cannot. The organization writes the instructions and the opening line. It is disabled on every account by default.

Voicemail transcription

Voicemail is transcribed to text automatically by Twilio’s speech recognition. That is machine transcription rather than a language model, but it is automated processing of what a caller said, so it belongs on this page.

Transcripts of recorded calls

On the Pro and Business plans an organization can switch on transcripts of the calls it records. It is off until an administrator turns it on, it reaches recorded calls with a caller and never calls between colleagues, and the transcription is Twilio’s rather than a language model. We store the text encrypted, Twilio’s copy is deleted once ours is stored, and a transcript is deleted with its recording.

AI summaries of recorded calls

Where an organization has turned transcripts on, a language model at Anthropic reads the transcript of a recorded call and writes a short summary of it: what the call was about, why the caller rang, a suggested outcome and any follow-ups. It also carries one word for how the call ended, out of positive, neutral, negative and mixed; that word is the model’s reading of the conversation and not a score of anything. The portal marks the summary as written by AI and asks the reader to check it before acting on it. It is stored encrypted and deleted with the transcript.

AI drafts of call evaluations

On the Business plan, an organization that has turned transcripts on can also let a language model draft an evaluation of a call against a scorecard the organization wrote. It is off until an administrator switches it on and picks a scorecard, a draft may quote only what the transcript contains, and a person has to read and save it before it counts as an evaluation. Nothing is scored by AI alone.

Speech recognition in menus

A call flow can let a caller say a word instead of pressing a key. Again Twilio’s recognition, and again only where an organization built a flow that uses it.

Nothing else in Inbound CX uses AI. There is no chatbot and nothing that reads your messages. Nothing reads a recording either, beyond the transcription above, where an organization has turned it on. We do not use customer data to train models, ours or anyone else’s, and no third party is permitted to train on it either.

A caller who reaches the AI assistant hears that they are speaking to an automated assistant before they say anything. The disclosure is spoken first, ahead of whatever opening line the organization wrote, in the language the flow is speaking.

It is added by us and an organization cannot edit it, shorten it or switch it off. That is deliberate: an organization writes both the instructions and the greeting, so a disclosure that lived in either would be one that could be deleted, and a language model can decline to follow an instruction in any case. Putting it in the audio itself is the only version that is actually guaranteed to be heard.

The model is also instructed to say plainly that it is automated if a caller asks, and never to claim or imply otherwise. A caller who asks for a person gets one.

When the AI assistant is handling a call, what the caller says is converted to text and sent to Anthropic, together with the last few turns of that conversation, so a reply can be generated. Nothing else goes with it: no call history, no contact record, no recording, no other caller’s conversation.

Where an organization has turned on transcripts and their summaries, the transcript of a recorded call goes to Anthropic in the same way, so that a summary of it can be written. Nothing else goes with it.

The conversation is held for the duration of the call and discarded with the call-flow session, which expires within hours. Anthropic is listed on the sub-processors page with its status. Content sent through the API is not used to train models.

The organization that built the flow is the controller of what its callers say. If you are a caller and want to know what was done with your conversation, ask the business you called first.

Nothing about a person. The assistant answers questions and routes calls. It does not decide whether someone gets credit, a job, a service, a price or a refund, and it does not score, rank or profile any caller. An AI draft of a call evaluation is a suggestion to a supervisor, who reads it, changes what they disagree with and saves it before it counts. There is no automated decision with a legal or similarly significant effect on an individual anywhere in Inbound CX.

Every AI conversation has two ways out, and one of them is a person: if the caller asks for a human, if the assistant cannot help, if it errors, or if the exchange runs past a limit the organization set, the call goes to a human. That limit is at most twelve exchanges and defaults to six.

If Quebec’s Law 25 gives you the right to an explanation of automated processing, or to have a person review it, the route out described above is that person, and sales@inboundcx.com will tell you what was used.

A language model can be confidently wrong, and one reading instructions written by an organization can be wrong in that organization’s specific way. An organization that turns this on is responsible for what its assistant tells callers, exactly as it would be for what an employee told them. Do not use it for anything where a wrong answer is expensive: quotes, eligibility, medical or legal questions, or anything an escalation to a person would handle better.

Write to sales@inboundcx.com.