Post-call extraction turns a voice call into CRM fields. When an AI call ends, a model reads the transcript and fills the fields you defined (interest level, budget, preferred date, promised payment). Edesy then pushes those values to your CRM via workflow integrations, to Google Sheets, or to any system through a webhook, so nobody has to read transcripts.
Extract your first fields: Try Edesy free — ₹50 free credit, no card needed.
The problem with transcripts
A transcript is a record, not data. A sales manager can't filter 500 transcripts for "budget above ₹50 lakh, wants a site visit this weekend". A CRM can, if those values are in fields. Extraction is the step that turns one into the other.
This post is the practical how-to. If you are building your own pipeline and want the architecture, see post-call data extraction with LLMs. For what to do with the CRM after the call, see post-call CRM automation.
Step 1: Decide what to extract
Start from the decision someone will make with the data. If nobody will filter or act on a field, don't extract it.
| Use case | Fields worth extracting | Who uses it |
|---|---|---|
| Real estate lead | Budget, location, property type, site-visit date, timeline | Sales team |
| EMI reminder | Promised date, amount promised, reason for delay, wants link | Collections desk |
| Appointment reminder | Confirmed, reschedule requested, new slot | Front desk |
| Survey | Score, main complaint, would recommend | Quality team |
| Solar lead | Roof type, monthly bill, owns property, best time to visit | Field team |
Five to eight fields per agent is a good limit. More than that and the call either gets long or fields come back empty.
Step 2: Use a template, then adjust
Edesy has extraction templates and presets for common call types. Pick the closest one and edit it. For each field, set:
- A clear name that matches the CRM field you will map it to.
- A type: text, number, date, yes/no, or a choice from a list.
- A description that tells the model what counts. "Budget in rupees as a number; if the caller gives a range, use the upper value" is far better than "budget".
- Allowed values for choice fields, such as hot / warm / cold. These keep CRM picklists clean.
See the post-call extraction feature for the full list of field options.
Step 3: Make sure the agent asks
The most common reason for an empty field is that the caller was never asked. Extraction can only find what was said. Check that your prompt asks for every field you extract. If you used Copilot to build the agent, ask it to add the questions for your fields.
For values that must be exact, use the collect data tool during the call. The agent asks, repeats the value back, and the caller confirms. Dates and amounts benefit most from this. The full list of in-call tools is on the mid-call actions page.
Step 4: Send the data to your CRM
There are three routes, depending on your stack.
| Route | Best for | How it works |
|---|---|---|
| Workflow integration | Kylas, Zoho, HubSpot, Salesforce, Pipedrive, LeadSquared, Edesy CRM | A workflow with the data extracted trigger and a CRM update step |
| Google Sheets | Small teams, quick reporting | Google Sheets update/append step in a workflow |
| Webhook or API | Your own system | Subscribe to call.extracted, or add a webhook step |
Workflow integrations are the simplest path for supported CRMs: choose the trigger, map each extracted field to a CRM field, and add conditions. See CRM integration and the workflow builder.
For your own backend, a webhook subscription sends the extracted data when it is ready. Payloads are
signed with HMAC in the x-edesy-signature header, and failed deliveries are retried up to five times.
Details are in the webhook subscriptions docs.
Step 5: Act on the values
Once the data is in fields, use conditions in the same workflow:
- Hot lead with a site-visit date: update the CRM stage and send a WhatsApp confirmation.
- Promised payment date: schedule a reminder call for the day after, if no payment arrives.
- Reschedule requested: book a new calendar slot or schedule a callback.
- Negative survey score: email the quality team with the summary.
Lead qualification outcomes can also run outcome actions directly: CRM update, email, callback, workflow or webhook.
Step 6: Check quality every week
Extraction quality drifts as scripts change. A weekly check takes ten minutes:
- Open per-agent extraction analytics and look for fields that are often empty.
- Read three or four transcripts where a field is empty or odd.
- Fix the cause: the question, the field description, or the allowed values.
- Regenerate extraction for recent calls if you need the corrected values in the CRM.
Common fixes:
- Dates come back in mixed formats: say the format in the description ("YYYY-MM-DD; convert 'next Friday' using the call date").
- Amounts include words: set the type to number and say "digits only".
- Choice fields get new values: list the allowed values and add "if unclear, use unknown".
A worked example
A real estate agent calls 1,000 portal leads. It extracts budget, location, property type and site-visit date. A workflow on data extracted checks: if a site-visit date exists, it updates the CRM stage to visit scheduled, sends a WhatsApp with the address, and books the slot on Google Calendar. If the lead is cold, it only updates the CRM and adds a note with the auto summary.
The sales team opens the CRM in the morning and sees a filtered list of visits, not 1,000 transcripts. Ready-made agents for this use case are in templates.
What it costs
Extraction is part of each call. Calls cost from ₹4/min on plans and credit packs (₹6/min pay as you go), plus telephony; see the pricing page. Sign-up includes ₹50 free credit, no card needed.