How to Use Thoughtly Variables for Dynamic Call Personalization

How to Use Thoughtly Variables for Dynamic Call Personalization

A generic script treats every caller like a stranger. For high-volume lead conversion teams in insurance, mortgage, home services, and education, that means lost context, repeated questions, and lower booking rates. Personalization fixes it — and in Thoughtly, the building blocks are variables, metadata, and contact attributes.

This guide walks through each data layer, shows you how to configure them in the Agent Builder and Automations, and ends with a working pattern you can adapt to your own pipeline. No invented features — everything here maps directly to current Thoughtly product mechanics.

What You'll Need

• A Thoughtly account with access to the Agent Builder and Automations
• At least one deployed voice agent (inbound or outbound)
• A CRM or lead source that feeds contacts into Thoughtly (HubSpot, Salesforce, GoHighLevel, or a webhook)
• Familiarity with Speak nodes and Outcomes — see the Thoughtly docs on Outcomes and Variables if you need a refresher

Understanding Thoughtly's Three Data Layers

Personalization in Thoughtly isn't a single toggle. It runs across three distinct data layers, each scoped to a different stage of the call lifecycle. Understanding which layer to use — and when — keeps your agent logic clean and your data reliable.

Think of it this way: variables are what your agent learns during the conversation. Metadata is context you attach to a specific call before it starts. Attributes are persistent facts about the contact that survive across every call, inbound or outbound.

Step 1 — Create In-Agent Variables for Real-Time Extraction

Variables live inside Speak nodes. They extract data from the caller's response in real time, before Outcomes evaluate — which means your branching logic always sees the freshest values.

Adding a variable to a Speak node

1. Open your agent in the Agent Builder and select the Speak node where you want to capture information.
2. In the right panel, scroll to Variables and click Add variable.
3. Fill in the fields:

• Name: A short, lowercase snake_case identifier. Examples: service_needed, zip_code, insurance_type.
• Source: Choose Current speak node for precise extraction from the latest reply, or Conversation history to search the full transcript.
• Extraction instructions: Write exactly what to extract and how to format it. Be explicit about edge cases.
• Format: Text, Number, or Boolean. This enforces the output type so downstream Outcomes stay predictable.

Example: extracting service type for a home services agent

Goal: Extract the specific home service the caller needs.
If multiple services mentioned: choose the most recent, highest-priority one.
If absent or unclear: return an empty value (no placeholder text).
Normalization: lowercase, trim spaces. Use canonical names: plumbing, hvac, electrical, roofing, general.
Do not invent values.

Once extracted, this variableVariableA named value the voice agent stores during a conversation — caller name, intent, qualifying answers — and uses to drive routing and post-call actions.Read full definition → is immediately available in Outcomes. A rule-based Outcome can route emergency plumbing calls to a live dispatcher while scheduling non-urgent requests for the next available window.

Re-extraction and confirmation

Each time the agent visits a node that defines a variable, it retries extraction and overwrites the previous value if it finds a better one. This is useful for confirmation flows — ask the caller to repeat their email, and the variable updates to the confirmed version. Use Current speak node as the source for confirmation nodes to avoid pulling older, unverified mentions from the conversation history.

Step 2 — Pass Metadata into Calls for Pre-Call Context

Metadata is a per-call context layer. It travels with the call and is discarded afterward. Use it when you want the agent to know something before the conversation starts — without permanently attaching it to the contact record.

Setting metadata in an automation

1. Open the Automation that triggers your outbound call.
2. In the Call Phone Number or Call Contact step, find the Metadata field.
3. Add a JSON object with the context you want to pass:

{
  "campaign": "summer_enrollment",
  "lead_source": "facebook_form",
  "intent": "schedule_tour",
  "program_interest": "nursing"
}

Referencing metadata inside the agent

Inside any Speak node prompt, reference metadata with the {{metadata.VARIABLE_NAME}} syntax. For example:

The caller submitted a form expressing interest in {{metadata.program_interest}}.
Their lead source is {{metadata.lead_source}}.
If the lead source is "facebook_form", acknowledge that they filled out a form online
before asking qualification questions.

Common metadata use cases

• Campaign labels: Tag each call with the campaign name so post-call reporting can segment by source.
• A/B testing: Pass a variant flag and adjust the agent's opening line or qualification flow accordingly.
• Intent signals: If your form captures what the lead wants (schedule a tour, get a quote, request a callback), pass it as metadata so the agent skips redundant questions.
• Priority flags: Assign a priority level based on lead score so high-intent leads get transferred to a live rep immediately.

Remember: metadata does not persist after the call. If you need a value to carry over to future calls, copy it into an Attribute in a post-call automation step.

Step 3 — Use Contact Attributes for Persistent Personalization

Attributes live on the Thoughtly contact record and persist across every call. They are the right choice for data that should influence future interactions — preferred language, region, plan tier, consent timestamps, or lifetime value.

Setting attributes in an automation

Use the Create or Update Contact step or the Add Attributes to Contact step in any automation. Example payload for a mortgage lead:

{
  "region": "southeast",
  "loan_type": "conventional",
  "pre_approved": true,
  "preferred_language": "spanish",
  "last_called_at": "2026-06-10T14:30:00Z"
}

Referencing attributes inside the agent

Attributes use a longer reference path: {{metadata.system.contact.attributes.ATTRIBUTE_NAME}}. In a Speak node prompt:

This contact is in the {{metadata.system.contact.attributes.region}} region.
Their preferred language is {{metadata.system.contact.attributes.preferred_language}}.
If preferred_language is "spanish", greet the caller in Spanish and continue
the conversation in Spanish unless they switch to English.

When to use attributes vs. metadata

Step 4 — Reference Variables in Prompts and Outcomes

The real power of variables shows up when you combine extraction, metadata, and attributes inside prompts and outcome rules. Here is a practical pattern for an insurance lead qualification agent.

Dynamic prompt example

Imagine an inbound agent for an insurance brokerage. Before the call connects, an On Inbound Call automation looks up the contact and passes metadata. The agent's advanced prompt references all three data layers:

The caller is {{metadata.system.contact.attributes.first_name}} from
{{metadata.system.contact.attributes.region}}.
They previously expressed interest in {{metadata.intent}} coverage.

Greet them by name. If intent is "auto", ask about their current coverage and
whether they are comparing quotes. If intent is "home", ask about property type
and desired coverage amount. If intent is unknown, ask what type of insurance
they are looking for.

Using variables in rule-based outcomes

After a Speak node extracts a variable, you can branch on its value with rule-based Outcomes. For example:

• If coverage_amount >= 500000 → route to a senior agent via Transfer node
• If zip_code matches your service area → continue qualification
• If email is empty → loop back to re-ask (self-loop pattern)
• If consent_to_record == false → end the call with a polite opt-out message

Rule-based Outcomes evaluate from top to bottom, and the first match wins. Place your most specific rules above more general ones to avoid misrouting.

Using variables in Actions

Mid-call Actions can reference extracted variables in their payloads. A webhook Action, for example, can include the caller's extracted zip code and service type in the request body so your backend can return real-time availability or pricing without a separate lookup.

Step 5 — Write Extracted Data Back to Your CRM

Variables extracted during the call are available to post-call automations. An On Call Completed automation can write every captured field back to your CRMCRMThe system of record for leads, contacts, deals, and activity. Thoughtly reads from and writes to your CRM continuously.Read full definition → as structured data — no manual entry, no transcriptionSpeech-to-Text (STT)The system that turns the caller's speech into text the agent can reason over.Read full definition → errors.

Example post-call automation flow

1. Trigger: On Call Completed (scoped to your qualification agent)
2. Step 1 — Update CRM record: Map extracted variables to CRM fields. For a HubSpot integration: service_needed → Service Interest, zip_code → Zip Code, coverage_amount → Coverage Amount.
3. Step 2 — Update contact attributes: Write last_called_at, last_outcome, and any persistent values back to the Thoughtly contact for future calls.
4. Step 3 — Send follow-up: If the outcome was interested but not ready to book, trigger an SMS or email follow-up referencing the specific details the caller mentioned.

This closed loop — extract during the call, write back after — is what makes AI-driven lead conversion measurably better than manual follow-up. Every call produces structured, accurate data regardless of which agent handles it.

Common Mistakes

• Vague extraction instructions. "Get their info" will produce inconsistent results. Write explicit instructions with format requirements, edge-case handling, and examples of correct output.
• Using Conversation history when Current speak node is safer. If you just asked a confirmation question, use Current speak node to avoid pulling an older, unverified value from earlier in the transcript.
• Storing temporary data in Attributes. Campaign variant flags and one-time intent signals belong in Metadata. Attributes should be reserved for facts that matter on the next call.
• Not handling empty values. Always plan what happens when a variable comes back empty. A rule-based Outcome with a self-loop that re-asks, or a default branch that skips the field, prevents dead-end conversations.
• Inconsistent naming. Stick to lowercase snake_case for all variable names (zip_code, not ZipCode or Zip Code). This keeps rule-based Outcomes and Action payloads readable and less error-prone.
• Over-extracting PII. Only capture the personally identifiable information you actually need. If you do not use the caller's full address downstream, do not extract it.

Measuring Success

Personalization should improve measurable outcomes, not just feel better. Track these metrics before and after implementing variable-driven personalization:

• Qualification rate: Percentage of calls that produce a fully qualified lead. Personalized agents that skip redundant questions and adapt to context should qualify leads faster.
• Booking rate: Percentage of qualified leads that book an appointment or next step. Context-aware agents that reference the caller's specific interest convert at higher rates.
• Average handle time: Personalized calls that skip already-known information should be shorter without sacrificing quality.
• CRM field completion: Percentage of post-call CRM records with all required fields populated. Structured variable extraction should push this toward 100 percent.
• Repeat-call deflection: Contacts who call back should not need to re-explain their situation. Persistent Attributes reduce repeat context-gathering.

Use Thoughtly's Analytics dashboard to segment call outcomes by agent, campaign (via metadata), and contact attributes. Compare conversion rates between personalized and non-personalized cohorts to quantify the lift.

Frequently Asked Questions

Can I use variables from one Speak node in another node's prompt?

Yes. Once a variable is extracted at any node, it remains available for the rest of the call. Reference it by name in any subsequent Speak node prompt, Outcome rule, or Action payload.

What happens if a variable extraction fails?

The variable returns empty. Build a fallback into your Outcomes — either a self-loop that re-asks, a default branch that continues without the field, or a Transfer to a human if the data is critical.

Can I personalize SMS and email follow-ups with the same variables?

Yes. Post-call automations triggered by On Call Completed have access to all variables extracted during the call. Map them into your Send SMS or Send Email step templates for consistent, personalized follow-up.

How do I test variable extraction before going live?

Use the Test Agent text chat in the Agent Builder. Send representative caller messages and check that variables extract correctly. Then use Call Me to verify the full voice experience. Keep a list of 10–15 common caller phrases per branch and run them after each edit.

Is there a limit to how many variables I can add per node?

Thoughtly does not impose a hard limit, but keep it practical. Each variable adds extraction overhead. Group related fields logically across nodes — ask for name and email in one node, service details in another — rather than extracting everything in a single Speak node.

Sources and Further Reading

Thoughtly Variables documentation — full reference for in-agent variable extraction, formats, and best practices.

Attributes vs. Metadata in Automations — detailed guide to the two data layers that carry context into and out of calls.

Thoughtly Agent Builder Overview — how nodes, outcomes, variables, and actions fit together.

Thoughtly Automations: Actions and Steps — reference for automation steps including Call Phone Number, Call Contact, and CRM integrations.

How to Use Thoughtly Workflows for Multi-Step Lead Qualification — related guide on building qualification flows with Speak nodes, Variables, and Outcomes.

How to Write Lead-Qualification Prompts Without Overfitting — companion guide for writing prompts that work well with variable-driven personalization.

How to Sync AI Conversations Back to HubSpot — step-by-step CRM writeback setup for HubSpot users.

Thoughtly — AI agents that convert the leads companies are already getting by calling, texting, and emailing until someone is ready to talk.