Call Analysis
Define what the AI extracts from every call — presets, custom fields, and per-agent scopes.
Every completed call is read by an AI and turned into structured data: a summary, the caller's sentiment, the topics discussed, how the call ended, follow-up actions — and whatever else you tell it to extract. The Call Analysis page is where you define that last part.
If Conversations is the archive of what was said, Call Analysis is the definition of what to pull out of it.
In this guide
- What every call produces, with no configuration at all
- Getting to the page
- Preset fields — the seven that ship enabled
- Adding your own fields
- Field types
- Scopes: account-wide and per-agent
- The 15-field limit
- Where results appear
- Suggestions to start from
- Rules that keep results usable
- Troubleshooting
1. What every call produces
Analysis runs on every completed call automatically. Without you configuring anything, each call yields:
| Field | What it is |
|---|---|
| Summary | A 2–4 sentence plain-language account of what happened and how it ended. |
| Sentiment | The caller's overall sentiment: positive, neutral, negative, frustrated or angry. |
| Topics | Up to five short topic tags. |
| Outcome | One short phrase for how the call ended — "booked appointment", "requested callback", "not interested", "information provided". |
| Action items | Up to five short follow-up actions for the business. |
| Disposition | The classified verdict on what answered — see Dispositions. |
| Memory facts | Durable facts about the caller, written to their caller profile. |
| Callback request | If they asked to be rung back — see Callbacks. |
Transcripts in any language are analysed fully, and the summary is written in the transcript's primary language.
Sentiment is inferred from wording, not audio. Analysis reads the text transcript, so it judges tone from language, word choice, escalation, complaints and satisfaction rather than from the sound of the voice.
2. Getting to the page
- Open the Voice AI hub.
- Click Call Analysis in the left navigation (the sliders icon).
- The page opens at
…/voice/call-analysis.
Per-call analysis is available on every tier, Launch included. Live monitoring, AI scoring and cross-call sentiment analytics are Growth and above.
3. Preset fields
Every account starts with a preset field set. These are real, editable fields — not hardcoded behaviour — so you can rename their prompts, change their type, disable them or remove them entirely.
| Field | What the AI is asked to extract | Default |
|---|---|---|
| Name | The caller's full name, and any preferred name they gave. | On |
| Telephone | Any phone or contact number the caller provided. | On |
| Any email address the caller volunteered, e.g. for confirmations. | On | |
| Address | The caller's address or structured location details. | On |
| Preferences | Stated preferences — appointment times, service preferences, habits. | On |
| Complaints | Any complaints or issues raised that should be flagged for follow-up. | On |
| Payment references | Invoice or payment references the caller mentioned. (Sensitive) | Off |
Payment references is opt-in. It is off by default because it captures financial identifiers; switch it on deliberately.
Call history is not a field here. The memory system always tracks who called and when, on every plan. You do not need to — and cannot — configure it on this page.
4. Adding your own fields
Click Add field and give it:
- A label — what the result is called wherever it appears. Up to 80 characters.
- A prompt — the question the AI answers about the call, written as an instruction. Up to 300 characters.
- A type — how the answer should be shaped.
The prompt is what does the work. Write it as a specific question about the call, not a category name:
| Weak | Better |
|---|---|
| "Budget" | "Did the caller mention a budget or price expectation, and what was it?" |
| "Competitor" | "Did the caller mention a competitor, and which one?" |
| "Quality" | "Rate the lead quality from 1 to 10 based on interest and fit." |
5. Field types
| Type | Result shape | Use it for |
|---|---|---|
| Text | A string | Names, free-form answers, single values |
| Yes / No | A boolean | "Did they ask for X?" |
| Number | A number | Scores, quantities, budgets |
| List | An array of strings | Objections, requested services, anything with several answers |
Choosing the right type matters downstream: Call Actions guards compare against these values, and a number stored as text cannot be compared with >.
6. Scopes
Fields are defined at two levels:
Account-wide — the default set, applied to every agent that has no override of its own.
Per-agent — an override for a single agent. Use it where one agent has a job the rest don't: a collections agent asking for a promise-to-pay date, a triage agent asking about symptoms.
When an agent has its own set, it is used in place of the account set (with the account set — or the presets — as the base it was built from). A single Save in the page header flushes every scope that has unsaved changes, so you can edit the account set and an agent override in one pass.
You can also copy one scope's fields into another, or replace a scope's fields wholesale, from the scope controls.
7. The 15-field limit
At most 15 enabled fields per scope. Beyond that, extra fields are silently dropped at analysis time, so the page shows the count and warns you before you cross it.
Labels are capped at 80 characters and prompts at 300. If you need more than 15 things extracted, the usual answer is a per-agent scope rather than one enormous account-wide set.
8. Where results appear
Custom fields are recorded on new calls only — enabling a field does not retroactively analyse calls that already happened.
Results show up:
- On the call's detail panel in Conversations, under the call summary.
- On the caller's profile, where facts accumulate across calls.
- As inputs to Call Actions guards, so an action can fire only when a field has a particular value.
- In Reports and CSV exports.
- Through the API.
9. Suggestions to start from
The page offers one-click starter fields. They are ordinary fields once added — edit them freely.
| Suggestion | Type |
|---|---|
| Competitor mentioned | Text |
| Lead quality (1–10) | Number |
| Callback requested | Yes / No |
| Product interest | Text |
| Objections raised | List |
| Budget mentioned | Text |
| Ready to buy | Yes / No |
10. Rules that keep results usable
- One question per field. A prompt asking two things gets one muddled answer.
- Say what counts as evidence. "Did they mention a budget?" is better than "budget", and "only if they gave a figure or a range" is better still.
- Prefer Yes/No and Number for anything you will filter or act on. Text is fine for reading; it is awkward to build a guard against.
- Disable rather than delete while you are experimenting — a disabled field keeps its prompt for later.
- Don't ask for something the transcript cannot contain. Analysis reads text. It cannot hear an accent, judge audio quality or see a screen.
11. Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
| A new field is blank on every call | It was added after those calls ran | Fields apply to new calls only |
| A field is always empty | The prompt is too vague, or the information genuinely isn't in the call | Rewrite the prompt as a specific question |
| Some fields silently stopped recording | More than 15 fields are enabled in that scope | Disable some; the page shows the count |
| An agent ignores the account fields | It has a per-agent override | Open the agent's scope on this page |
| Numbers come back as text | The field type is Text | Change the type to Number |
| Nothing is analysed at all | Very short calls with no transcript produce no analysis | Check the call has a transcript in Conversations |
Related guides
- Call Actions — making something happen because of what was extracted
- Conversations & Recordings — where results are read
- Caller Profiles & Memory — how facts accumulate across calls
- Dispositions — the classified call outcome