Configure AI auto-fill and backfill CRM data

Mark fields AI-fillable, configure field-task behavior, run backfills, and verify AI-written CRM updates.

AI auto-fill lets GTM Engine update selected CRM fields from transcripts, meetings, emails, research, and other record context. Backfill lets you run those field tasks across existing records, not just new activity.

What You'll Build

You will configure one AI-fillable field, test it on a record, and then backfill existing records.

When to Use This

Use this when you want GTM Engine to maintain fields such as:

Account angle or summary.
Deal risk assessment.
Forecast notes.
Competitors.
Next best action.
Buying committee notes.
Meeting summary or follow-up context.

Estimated time: 15-30 minutes.

Concepts used: Workflows, Tasks, Variables, Batch History.

Prerequisites

CRM field mapping is complete in Settings → Fields.
Email, calendar, or call recorder data is connected if the field depends on conversation context.
You have permission to configure fields and run backfills.
You know which record type the field belongs to: account, contact, or opportunity.

1. Choose the Field

Open Settings → Fields and pick the field you want GTM Engine to maintain.

Good first fields are:

Text or long-text fields, such as account_angle or deal_risk_summary.
Single-select fields with clear allowed values, such as forecast category.
Multi-select fields where the valid options are already defined.
Date or number fields where the expected output is unambiguous.

Avoid starting with a field that has unclear business rules. If people on your team disagree about the desired value, document that decision before asking AI to fill it.

2. Mark the Field AI-Fillable

In the field configuration, enable AI auto-fill for the field and configure the field task instructions.

The field task should answer:

What context should the model use?
What value should it produce?
What should it do when there is not enough evidence?
Are there allowed values or formatting rules?

For structured fields, keep the instructions strict. For example:

Choose one forecast category from the allowed values only. If the evidence is weak, keep the current value.

3. Pick or Confirm the Model

New field tasks use the configured Default field task model unless you choose a specific model on the task.

Admins can adjust the default model from the workflow field-task model configuration area. Existing field tasks keep the model they were saved with until an admin migrates them.

Use a stronger model for ambiguous fields that require synthesis. Use a faster model for simple extraction or classification.

4. Test on One Record

Before backfilling many records, test the field task on a single representative record.

Check:

The generated value is in the correct format.
The model does not invent evidence.
Single-select values match allowed options.
Dates are real dates.
Multi-select values use existing options.
The field remains unchanged when context is insufficient.

If the output is wrong, tighten the instructions before running a backfill.

5. Run Backfill

When the test looks good, enable Backfill existing records and choose the relevant scope.

Use Save & Backfill or Confirm & Backfill depending on the dialog state. GTM Engine creates a trackable batch so you can follow progress in Batch History.

For opportunities, use available filters when you only want to backfill a subset, such as open pipeline or a specific stage category.

6. Verify Results

After the backfill starts:

Open Batch History and watch progress.
Open a few updated records.
Confirm the field value matches the evidence.
Check whether failures are concentrated on one record type, field, or missing context.
Rerun failed records only after you understand the cause.

Success Check

You are done when:

The field task is saved.
A single-record test produces the expected value.
Backfill starts as a batch.
Updated records show sensible field values.
Failed records, if any, have understandable errors.

Common Pitfalls

Running a large backfill before testing one record.
Asking the model to infer a value your team has not defined clearly.
Forgetting that existing tasks keep their saved model until migrated.
Letting the model invent enum values instead of choosing from allowed options.
Backfilling records that do not have the transcript, email, or activity context the field needs.