Let AI propose.
Let the process decide.
| Step | Who or what owns it | Reason |
|---|---|---|
| Identify a callback request in free text | AI can propose | The wording varies from message to message. |
| Check whether contact is permitted | Explicit rules | A model should not invent permission. |
| Choose the account owner | CRM mapping rule | Ownership is a business decision already recorded. |
| Resolve “next Thursday” without a timezone | Clarification or review | The date cannot be safely assumed from the sentence alone. |
| Confirm an appointment | Calendar response | A conversational claim is not proof of a saved booking. |
Evaluate difficult examples before activation: conflicting dates, a shared company inbox, a message about two projects, and a request outside the service area. Record both incorrect results and the amount of review work left for your team.
Include the cost of review.
A cheap model call is not the full cost if every result needs correction. Compare the automated preparation plus human review against the existing task. If a rule solves the step reliably, an AI model may add no useful value.
Ask what judgment the step requires
Some work is repetitive because the same rule applies every time: when an approved job changes status, create a task. Other work is repetitive but involves interpretation: read a free-form email and identify the job, request, and missing information.
The first situation may need a straightforward integration. The second may benefit from AI, provided the system handles uncertain interpretations and checks the result. Adding AI to a deterministic step can add cost and variability without improving the outcome.
Put AI inside a controlled process
A useful pattern is to have AI suggest structured information, validate that information, and route uncertain results to a person. The surrounding workflow still needs ordinary rules for permissions, required fields, duplicate handling, and destination updates.
- Extract: pull proposed details from unstructured text.
- Validate: check required values and valid record identifiers.
- Review: send ambiguous matches or consequential actions to a person.
- Act: perform only the action the workflow permits.
- Record: keep enough context to explain and correct the outcome.
Treat outside content as input
An email, document, or customer message can contain irrelevant or misleading instructions. It should be treated as business data, not permission to change what the automation is allowed to do. The workflow should restrict available actions and avoid exposing broad account access.
Keep sensitive information out of a model call unless the task needs it and the handling has been agreed. Review the relevant provider settings and your customer obligations before processing real records.
Test difficult examples before relying on it
A few neat demonstrations are not enough. Use representative historical examples with permission, including incomplete messages, multiple jobs with similar names, contradictory details, and unexpected attachments. Define what counts as an acceptable answer before evaluating results.
Track incorrect matches and manual-review rates as well as successful processing. A high review rate may still be useful if the workflow saves preparation time, but the remaining labor belongs in the cost calculation.
Keep a person accountable
AI does not remove the need for an operational owner. Someone still needs to review exceptions, approve changes, and decide what to do when the result is wrong. Start with a limited task and expand based on observed performance.
GoodHandoff uses AI when it helps the actual process. The proposal identifies where interpretation occurs, where rules apply, and where your team retains judgment.