Operations agent: steps

The palette on the left of the designer groups the steps: Recording, Understanding, Reports, Deliver, Storage and Flow. Click a step on the board to set it up in the inspector; the inspector also lists what the step produces for the steps after it. Rarely needed settings are folded under More options.

Recording#

Download recording fetches the call recording into a temporary folder that is deleted when the run ends; nothing is kept. By default it takes the trigger's recording link. Max size (MB) caps the download (200 MB by default). It produces file, filename, bytes and content_type.

Understanding#

Transcribe turns a recording into text.

• Spoken language: a code such as el or en; empty detects it.
• Also provide in: extra languages; the translations are in {{transcribe.translations}}.
• Separate the speakers: splits the text into turns, {{transcribe.segments}} (speaker, start and end in seconds, text).
• Name the speakers: numbers (1, 2) or agent_customer: each turn is marked agent (your side of the call) or customer, worked out by AI from the conversation and the call direction. A voice it cannot place keeps its number and {{transcribe.speakers_note}} says why.

It produces text, language, duration_sec, translations, segments, speakers and speakers_note.

Summarise writes a summary of the transcript or any text: brief, detailed or bullets, in the language you choose, with optional extra instructions. It produces text.

Reports#

Report from questions answers your prepared questions about the transcript (for example Did the customer place an order?) and builds an online report and an Excel workbook, optionally with the transcript. The run shows each answer and where in the call it was found. It produces title, answers and the workbook for the email step.

Deliver#

Send email sends a message with the flow's values filled in.

• To, Cc, Bcc, Reply-To and Sender name.
• Email template: one of your account's branded layouts, or none to send the message as it is (see Email templates).
• Format: plain text or HTML (HTML also gets a plain-text copy).
• Add the transcript to the message, attach it as a .txt file, attach the Excel report.

Webhook calls a URL of yours at this point of the flow: method (POST, PUT, PATCH, GET, DELETE), query parameters, headers, authentication (bearer token, basic, API key), and a JSON, form or text body, every value with the flow's data filled in. With a Signing secret each call is signed so your endpoint can verify it came from us. Your endpoint confirms receipt by answering 2xx; any other answer is a failed attempt and is retried. It produces status, response and json, the parsed answer, for the steps after it.

To send a CRM the transcript split by speaker, put "transcript": {"text": "{{transcribe.text}}", "segments": "{{transcribe.segments}}"} in the body.

Both deliver steps can be previewed and tested on their own: see Designing and testing.

Storage#

S3 and AWS S3 work with buckets, files and folders on your own storage: upload a recording or a transcript, copy, list, download or delete. See Storage (S3).

Flow#

• Condition compares a value (equals, not equals, contains, greater than, less than, is empty, is not empty) and continues on the yes or no branch.
• Dispatcher sends the data to several targets at once: other operations agents of the project, each doing its own job, and webhook URLs. Targets that already succeeded are not sent again on a retry.
• Call another agent runs one other operations agent of the project with data from this flow and, by default, waits for its result (status, output).

Success and failure branches#

An arrow can leave a step on success or on failure. A failure branch lets you tell somebody when a step gives up, for example email the team when your CRM refused the transcript. In the step's retry settings, choose whether the failure branch is taken after the retries are used up or on the first failure.