The path that works: pick one high-value field, enable in suggested mode, watch accept rate for four weeks, switch to applied if accept rate is above 85 percent, expand. The path that fails: enable autofill on every field in applied mode and discover the data quality problem two months later when reports start looking off.
- Identify one or two high-value pilot fields
Pick a picklist field where data entry is currently spotty and where the value correlates with other record data. Account.Industry and Opportunity.Type are common pilots.
- Confirm training data volume
Open Setup, Einstein Autofill. Check that the pilot field has at least 1,000 historical records with the field populated. Below that, predictions are noisy and the pilot will fail.
- Enable autofill in suggested mode
Toggle autofill on for the pilot field, mode Suggested. The model trains in the background; predictions appear on the next record-create or record-edit after training completes.
- Pilot with one team for four weeks
Five to ten users in the pilot team see suggestions on the field. The rest of the org sees the field unchanged. Watch the Insights report for accept rate, edit rate, and clear rate.
- Decide on applied mode based on accept rate
Above 85 percent accept rate, applied mode is safe. 60 to 85 percent, leave in suggested. Below 60 percent, the field is wrong for autofill; revert and pick a different field.
- Expand to additional fields one at a time
Repeat the pilot pattern per field. Resist the urge to batch-enable; per-field rollout catches field-specific issues before they become org-wide complaints.
- Schedule the monthly Insights review
Pull Einstein Autofill Insights monthly. Watch for fields whose accept rate drops over time. Drift signals retraining is needed or the underlying data shape has shifted.
Per-object list of fields autofill is active on. Roll out one at a time, not in batches.
Suggested shows a pre-filled value the user can accept or edit; applied writes the value on save without showing it. Start with Suggested.
Filter the historical records the model trains on (date range, record type, owner). Useful when older data has different patterns than current data.
The Einstein Autofill permission set that grants users access to suggestions on the enabled fields.
How often the model retrains against fresh data. Defaults to weekly; can be tuned for fields where the underlying pattern shifts faster.
- Applied mode without piloting suggested mode produces wrong data that users never see. Always pilot suggested first.
- Fewer than 1,000 historical records with the field populated produces noisy predictions. The page warns; admins still ignore the warning regularly.
- Text fields rarely benefit from autofill. Predictions are too generic to be useful, and users learn to ignore them, which trains poor habits.
- Validation rules still apply to autofilled values. A predicted value that fails a validation rule blocks the save, which is desirable but confusing if undocumented.
- Record history shows the saving user, not Einstein, as the author of an applied autofill value. Compliance teams may want a separate audit trail signal documented.