Let AI in customer service independently prepare work that is easy to verify, but leave decisions to an employee when an action has major consequences, is difficult to reverse or relies on uncertain information. Human-in-the-loop means more than an “approve” button. The employee needs to be able to assess the context, source, proposed action and possible consequences.
A model can summarise a long customer enquiry, find relevant knowledge and draft a response. That is different work from cancelling an order, refunding money or promising an exception. Deliberately separating these tasks lets you benefit from speed without quietly transferring decision-making authority to a language model.
The right boundary depends on consequences and reversibility
“Can AI do this?” is not the best design question. Many actions can technically be automated. The relevant question is: “What happens if this outcome is wrong, and how easily can we put it right?”
Assess each application along four dimensions:
- Impact: does the action affect money, delivery, rights, safety or an important customer relationship?
- Reversibility: can an error be reversed completely and quickly?
- Certainty: are identity, source data and policy rules unambiguous?
- Degree of exception: does the situation follow a known route or require interpretation?
The higher the impact and uncertainty, and the lower the reversibility, the stronger human oversight needs to be.
Four levels of autonomy
| Level | AI’s role | Employee’s role | Example |
|---|---|---|---|
| 1. Support | Summarise, search and organise | Creates the response and decision independently | Summarising a long conversation |
| 2. Propose | Drafts a response or follow-up action | Checks, edits and sends | Preparing return instructions |
| 3. Execute after approval | Prepares a defined system action | Assesses the context and explicitly approves | Changing a delivery date |
| 4. Execute automatically | Handles a predictable, low-risk step | Reviews samples and exceptions | Internal categorisation or acknowledgement of receipt |
An application does not need to stay at one level permanently. Start lower, measure corrections and errors, and increase autonomy only for a tightly defined scenario that is demonstrably manageable.
Human oversight needs to be meaningful
An employee who must click through ten proposals in five seconds may formally provide oversight, but offers little protection. A good review screen therefore shows:
- the original enquiry, without replacing it with the AI summary;
- the correct customer, order or case file;
- the knowledge sources used and current system facts;
- which information is missing or contradictory;
- the proposed text and, separately, the proposed action;
- which systems will change because of the action;
- how the action can be stopped or reversed.
Yindle Desk can present that service decision as one case. If an approved action needs to update data across several applications, Yindle Connect can execute it in a controlled and traceable way.
Check identity and permissions outside the model
A language model must not decide whether a sender actually belongs to an order. A customer may accidentally quote the wrong number, a message may have been forwarded or input may deliberately mislead. Identity and permission checks therefore belong in fixed application logic.
For an order-related action, use at least:
- a reliable link between the sender and customer profile;
- a check that the process object belongs to that customer;
- a server-side list of permitted actions;
- a fresh check immediately before execution;
- an audit entry recording initiation, approval and outcome.
The model can propose “change delivery date”, but the application determines whether the action exists, whether the user is authorised and whether the change is still possible at that moment.
Separate the response, decision and execution
Many automations unnecessarily combine these three parts:
- Response: what do we tell the customer?
- Decision: what does the organisation allow in this situation?
- Construction: which change is made in which systems?
An employee may approve the response but reject the action, or vice versa. Allowing separate approval makes clear where a correction belongs. A text edit calls for better knowledge or tone; a rejected action may indicate missing process rules or permissions.
Make clear when the customer is speaking with AI
There is an important distinction between AI drafting something for an employee in the background and a system communicating directly with a customer. Transparency requirements apply to AI that interacts directly with people. The European Commission explains under Article 50 of the AI Act that people must know from the outset when they are communicating directly with an AI system, unless that is obvious.
Transparency is sensible regardless of the precise legal classification. Explain what the assistant can and cannot do, offer a recognisable route to an employee and avoid suggesting a human identity when a system responds independently. Have the specific application reviewed legally.
A practical approval framework
Use the same five checks for each proposed action:
- Is the link correct? Does this concern the right customer and process object?
- Are the facts current? Was the status checked again immediately before execution?
- Is this action permitted? Does it fit policy, permissions and any statutory rights?
- Are the consequences visible? Does the employee know which systems and agreements will change?
- Has recovery been designed? Can partial or incorrect execution be detected and recovered?
If any answer is unknown, the case does not continue automatically. It receives an owner who can supply the missing information.
Learn from corrections without automating blindly
Record more than whether a proposal was approved. Note what the employee changed and why. Group corrections, for example, by missing knowledge, wrong source, incorrect intent, tone, incomplete context or unsafe action.
Many small tone changes call for a different improvement from one rare but serious incorrect order action. Use both frequency and impact. A controlled approach to automation starts with observing alongside the team, evaluating and only then introducing limited autonomy.
Checklist for meaningful human-in-the-loop oversight
- The employee sees the original input and relevant sources.
- The text proposal and system action can be assessed separately.
- Fixed logic checks identity, permissions and current status.
- Approval is a deliberate action, not an unclear default.
- Rejecting and editing are easy and recorded as feedback.
- High-risk and irreversible actions cannot quietly happen automatically.
- The customer can reach an employee when direct AI interaction is insufficient.
Give AI a clear role, not unlimited authority
Human-in-the-loop works when people and systems each do what they do well. AI processes language, finds context and prepares repeatable work. The employee understands exceptions, weighs consequences and takes responsibility for decisions with real customer impact.
Want to determine a responsible level of autonomy for one service workflow? Design the approval model with Yindle, including sources, permissions, recovery and measurable feedback.