Call Analysis
Six answers off one reading of the transcript, in a shape that is fixed in code rather than in a prompt. A workspace tunes what the model weighs; it never reaches the keys that come back.
One pass over the transcript
validatingChosen per vertical. Says what matters on this kind of call.
Wrapped and treated as data, never as an instruction to the model.
The keys and their shapes. Nothing a prompt can reach.
- summaryInvoice issued at the old rate; reissue and credit agreed, renewal call booked.in shape
- sentimentnegativein shape
- tonalityagent measured · customer frustrated · overall tensein shape
- intentions3 intents · 2 commitments · callback_requestedin shape
- scoresoverall 6 · communication 8 · resolution 5 · compliance 9in shape
- spam_urgencyspam 0.02 · urgency normalin shape
Every requested dimension present and in shape. Written to the call in one transaction.
Product figures from the platform’s own defaults - not customer averages
How it works.
Six requests folded into one
Summary, sentiment, tonality, intentions, scores and the voicemail spam reading used to be six separate calls, each re-sending the same words. They are assembled into a single instruction now, so the six readings agree with one another instead of being reconciled afterwards.
Three layers above the model, two of them yours
Domain guidance for the vertical goes first, workspace context is appended inside a tag derived from the workspace identifier, and the output contract is concatenated last, out of code. Editing a prompt changes what is looked for. It cannot rename a key.
The words arrive wrapped in a tag nobody can guess
Speech to text output is influenced by whoever was on the phone, so each run mints a random delimiter after the audio already exists, redacts anything matching it, and strips the invisible characters used to hide instructions from a reviewer.
Checked dimension by dimension before storage
Each requested answer is parsed against its own rule: enums lowercased then matched, numbers coerced then range checked. A miss comes back with its reason, the reading repeats once, and a second miss fails the lot rather than saving half.
Two moments in every read.
Every read passes through the same seven. Call Analysis is the lit ones, and everything either side of it is a different page in this category.
- 01Source
the call or the thread it reads
- 02Transcribe
audio into words, with speakers
- 03Read
the pass over the whole of it
- 04Judge
the score, the sentiment, the intent
- 05Extract
the fields and follow-ups pulled out
- 06Write
what lands back on the record
- 07Review
a person checking the machine
The specifics.
8 facts- Dimensions
- Summary, sentiment, tonality, intentions and scores. Spam and urgency on voicemail only
- Request
- A single text call per analysis, capped at 3500 output tokens by the orchestrator
- Malformed answer
- Rejected with a reason, repeated once, then failed. Nothing partial is saved
- Out of range
- Refused rather than trimmed. A score of 85 does not quietly become a 10
- Needs words first
- Any text dimension forces transcription; if that fails the run stops there
- Where it lands
- One row per source, unique on type and identifier, so a re-run replaces rather than adds
- Starting one
- Automatically when a recording finalises, or by hand at thirty runs an hour per person
- Not the same as
- This is the pass itself. The summary, the scores and the sentiment are its fields
What it reads, and what reads it.
Nothing here invents its input. These are where the material comes from, and where the verdict goes afterwards.
More in Intelligence
11 capabilitiesAI that proposes edits to the record - an insight becomes a field once you accept it.
The model proposes record updates with its confidence and the quote it heard them in, shown as a before-and-after you accept or reject row by row.
Positive, neutral or negative on every analysed call, stored in a column of its own so the ones that went wrong are a filter rather than a listening exercise.
Chosen words rather than a fixed list for how the agent sounded, how the customer did, and the conversation overall - the manner behind the sentiment.
Compliance scored nought to ten against what the vertical says matters, because no phrase list exists to recite. A missing reading is excluded, never passed.
What the customer was actually asking for and what the agent committed to, extracted as structured intents rather than left inside the transcript.
Callbacks and meetings pulled out with their times, a confidence figure and the words they came from, ready to accept in one click or reject in one.
Four of six speech paths return the channel each side was recorded on, so the transcript knows who spoke rather than guessing, and one control swaps a pair.
Threads summarised and read for intent and urgency, but only when a participant matches a record, so the model is never called on mail that is not yours.
Draft from an instruction or rework a message you have, landing in the composer with a suggested subject and going nowhere at all until you send it.
Describe what you track or paste a spreadsheet's columns, and get a named field group back with types, options and placeholders to approve one at a time.
A sentence becomes a grid filter and a sort order with the reading explained back, on the toolbar of every grid that carries a column menu of its own.
The rest of the platform.
Five more categories, all on the same record and the same bill. Each card names three of its capabilities, so you can tell from here whether it is worth opening.
Intelligence
See call analysis on your own floor.
Thirty minutes, your numbers and your data. We will set call analysis up live and you can decide from the thing itself rather than from this page.
14-day trial · no card · migration included