Before acting and after
There is oversight before, when a person approves the answer or the task before it reaches the customer, and oversight after, when a person reviews a sample of what the AI already did.
The first slows the agent down and the second does not. Most teams combine both depending on the risk of each topic.
Why it matters
It is the answer to the question every decision maker asks, can I trust the AI with my customers?
There is no need to review everything. There is a need to know which tasks it never does alone, who reads what it answered and what happens when it gets something wrong. Without that, trust depends on nothing going wrong.
Where a person goes
In customer service, human oversight sits at four points.
- Before a sensitive task. A refund, a cancellation or a change of account holder waits for a person.
- When the AI does not know. It hands the conversation over instead of answering.
- In the daily review. A person reads a sample of answers and flags the wrong ones.
- Before a change goes live. New instructions are tested with real questions.
How much oversight, by risk
Oversight is spread by topic.
Reviewing everything before it is sent removes the point of having an agent.
| Topic | Oversight | Example |
|---|---|---|
| How-to question | Review a sample, afterwards | How to export a report |
| Question about account data | Review a sample and every failure | The status of an invoice |
| Task involving money or a contract | Approval before doing it | A refund |
| Sensitive topic | Always a person | A complaint |
What the EU AI Act says
Article 14 of Regulation (EU) 2024/1689 requires high-risk AI systems to be designed so that they can be effectively overseen by natural persons while they are in use.
Whether a specific use is high-risk is for the company’s legal advisers to assess.
Human in the loop in Intake
In Intake, an action marked as requiring human approval is not run by the AI, which hands the conversation to a person on the team.
On every answer, “Why it answered this way” shows the sources it read and the team marks whether it was right. Instructions, actions and guided responses are tested in draft before they are applied.
