Skip to main content
Call Analysis and Data Extraction

Extraction Templates, Run Conditions and Backfills

Start extractions from 13 ready-made templates, limit which calls they run on, test them on a real call and fill in data for past calls.

Written By voiceaiwrapper

Last updated 26 days ago

This article covers three tools that make Call Data Extractions faster to set up and cheaper to run: templates to start from, run conditions that skip calls you don't need data from, and the Test & backfill panel for checking an extraction on a real call and filling in data for calls that already happened.

If you haven't created an extraction yet, start with Capture Structured Data with Call Data Extractions.

Applies to: agents on a VAW Native Pod (an engine or provider connection).

Before you start

Start from a template

When you click New extraction (or New on an agent's Analysis tab), you can pick one of 13 templates. Each fills in a name, a description and a set of fields, including which fields are required and the choices for list fields. Change anything you like in the editor afterwards.

TemplateCategoryWhat it capturesFields
Booking detailsSchedulingWho booked, for when, and how urgent it is.patient_name, appointment_date, insurance_provider, urgency, callback_ok
Appointment rescheduleSchedulingThe old slot, the new slot, and why it moved.old_date, new_date, reason
Cancellation reasonSchedulingWhy they cancelled and whether a save offer landed.reason, save_offer_accepted
Lead qualificationSales and ordersHow good the lead is and what they are looking for.lead_quality, budget, timeline, decision_maker, objections
Property inquirySales and ordersWhich property, their budget, and whether they want a viewing.property_address, budget, bedrooms, timeline, wants_viewing
Order detailsSales and ordersThe order, the items, and any delivery issue.order_number, items, delivery_date, issue
Support issue intakeSupport and billingWhat broke, how badly, and whether it was fixed on the call.product, issue_summary, severity, resolved_on_call
Payment inquirySupport and billingThe invoice, the amount, and whether it is disputed.invoice_number, amount, dispute, promised_date
Promise to paySupport and billingHow much they promised, by when, and how.amount_promised, promise_date, payment_method
Callback requestSupport and billingWho to call back, when, and why.callback_number, best_time, reason
Contact infoPeople and feedbackName and how to reach them.full_name, phone, email, preferred_channel
Survey and NPS feedbackPeople and feedbackThe score and what drove it.nps_score, liked, disliked, would_recommend
Candidate screeningPeople and feedbackRole, experience, expectations, and interest.current_role, years_experience, salary_expectation, notice_period, interested

You can search templates by name, description, category or field name. If nothing matches, use Start from scratch.

Rename template fields to match your client's business. A dental clinic is happy with patient_name; a law firm would rather have client_name. Field names become the keys in webhook data, so settle them before you connect a CRM.

Limit which calls an extraction runs on

By default an extraction Always runs on every analysed call from the agents it's attached to. To skip calls you don't need data from, open the extraction and click Only run when.

ConditionWhat it does
Call has at least this many messagesSkips calls with fewer messages in the transcript. Each thing the caller or the agent said counts as a message.
Call lasted at least this many secondsSkips calls shorter than this.
End reasonAny end reason (default), Is one of... or Is NOT one of..., then Pick end reasons. For example, only run when the caller ended the call, or never run on calls that ended in an error.

Leave a number empty to ignore that condition. Numbers must be whole numbers of 1 or more. All conditions are optional and combined with AND, so a call must meet every condition you set. Click Done, then Save.

Call Data Extractions β†’ an extraction's editor with the Only run when panel open: Call has at least this many messages set to 4, Call lasted at least this many seconds empty (showing the Always placeholder), and End reason set to Is NOT one of... with one end reason picked.

Conditions are checked before the AI model runs, so a skipped extraction costs nothing. It isn't silently missing either: the call's Analysis tab shows it as Skipped with the reason, for example "Too few messages for this extraction (it needs at least 4)." or "This extraction only runs when the call ends with: Customer ended call."

For the full list of end reasons, see Call Statuses and End Reasons.

Test an extraction on a real call

Open the extraction, click Test & backfill, then click Preview on the latest call. It runs the extraction on the most recent ended call on this Pod and shows the result straight away: the status, the model used and the values it found. Nothing is saved to the call.

Use preview after every change to fields or descriptions, so you don't have to wait for new calls.

  • Save first. Preview runs the saved version of the extraction, not unsaved edits.
  • Conditions apply. If the latest call doesn't meet your run conditions, the preview returns no values and names the condition that wasn't met.
  • If there's no ended call yet, you'll see "No ended call to preview on yet. Make a test call first."

Fill in data for past calls with a backfill

A backfill runs the extraction again on calls that already ended. Use it after you create an extraction, attach it to an existing agent, or change its fields.

  1. Open the extraction and click Test & backfill.
  2. Under Agent, choose Any agent or one agent.
  3. Optionally, set From and To dates.
  4. Click Run backfill.

You'll see "Backfill queued. Results land on the matching calls shortly." While it runs, the Test & backfill row shows "Backfill in progress: N of N calls". Previous backfills lists each job with its status (Queued, Running, Completed or Failed), how many calls it covered and when it ran.

Call Data Extractions β†’ an extraction's editor with the Test & backfill panel open. Top: the Preview on the latest call button with a Completed preview result below it showing extracted values. Below: the Agent dropdown set to Any agent, the From and To date fields, the Run backfill button with its note about up to 100 calls, and Previous backfills listing one Completed job (for example 42/42).

What a backfill covers:

  • Up to 100 calls per backfill, newest first, from the ended calls that match your filters. For more, run several backfills with different date ranges.
  • Any agent means every agent on the Pod, including agents the extraction isn't attached to. Pick an agent to stay within one agent's calls.
  • Existing results are replaced. Calls that already have a result for this extraction get the new one.
  • Run conditions still apply. Calls that don't meet them are recorded as Skipped and cost nothing.
  • Completed means every selected call has been queued. Results then appear on each call over the next few minutes.

A backfill spends AI credits. "Each run bills its own tokens": every call it processes is a separate AI model run, added to that call's cost and charged to your credits. A backfill over 100 calls costs roughly 100 times one run. Preview first, and filter by agent and date so you only process the calls you need.

How extraction runs are billed

Every extraction run uses the agent's AI model for analysis and extractions and is charged in AI credits on top of the call's own cost:

  • After each call: one run per attached extraction, once post-call analysis completes. If analysis is skipped or turned off, attached extractions don't run.
  • Backfills: one run per processed call. A backfill runs on any ended call that matches its filters, even one whose analysis was skipped.
  • Skips cost nothing: calls that fail a run condition, or that have no transcript to read.

You can see the extraction line in each call's cost breakdown. See Credit History and Per-Call Costs.

Frequently asked questions

Not from the webapp. Run backfill only re-runs the extraction you're editing. Summary, sentiment and success results stay as they were when the call was analysed.

Open one of those calls and check its Analysis tab. The extraction may be Skipped because the call didn't meet your run conditions, or because the transcript was deleted under your retention settings. If it's still Pending or Running, wait a few minutes.

The job stopped before queuing every call, and the error appears next to it in Previous backfills. Calls already queued still get results. Run the backfill again. Calls that already got the new result are processed again, and charged again.