AI customer support agent: how to start a pilot

A support team doesn't lose time only on typing. It loses time in three places: reconstructing the customer's request, looking up the approved information and drafting the reply. An AI customer support agent can help with exactly these steps.

That is why the pilot starts from a well-defined task. It doesn't try to automate the entire conversation with the customer. You pick one type of request, describe it thoroughly and check whether the agent handles it correctly.

The pressure to adopt AI is real, but so is the resistance. According to a 2024 Gartner survey, 60% of customer service leaders say they are under pressure to adopt AI. The same survey shows that 64% of customers would prefer companies not to use AI in customer service. A narrow, controlled pilot helps you manage both pressures.

The first requests you give an AI customer support agent should have clear procedures. For each type, there must be a description of what an acceptable answer looks like. If no one on the team can say what a good answer looks like, the agent has nothing to go by.

Deliberately leave out of the first version:

• requests involving unclear commercial terms;
• requests that involve commitments to the customer;
• requests for which you don't yet have a written procedure.

If you are still choosing the process, read our guide on how to choose the first process for an AI agent in your CRM. The core principle stays the same: preparing information is kept separate from any action that affects the customer. A verifiable proposal is often a better starting point than execution without approval.

During the pilot, an AI customer support agent works only with approved content. That means the knowledge base defined for the pilot, not every document the company has. The clearer the source, the easier the review.

If two documents contradict each other, the decision does not belong to the AI. It belongs to the process owner, who corrects or withdraws the wrong document. It pays to do this exercise before the pilot, not during it.

When information is missing, the agent asks for human review. It doesn't fill the gap with assumptions. This rule is written explicitly into the agent's guardrails, not left open to interpretation.

The rule matters from a legal standpoint too. In Moffatt v. Air Canada (2024), the company was held liable for incorrect information given by the chatbot on its website. The American Bar Association's analysis notes that the tribunal rejected the idea that the chatbot was a separate entity.

In the first version, an AI customer support agent sends nothing on its own. It drafts, and a person authorises sending. Drafting and authorisation are separate steps, and each reply is approved individually.

What a verifiable output contains

The agent's output has three parts:

• the proposed reply to the customer;
• the information used, meaning the sources from the knowledge base;
• what needs checking before sending.

This way, the approver doesn't redo the work. They can see where each statement comes from and where the uncertain points are.

Human accountability, defined in advance

Name the people who approve, who own the knowledge base and who can stop the agent. The NIST AI RMF requires the AI system's tasks to be defined and human oversight to be documented. It also calls for designated people who can disengage or deactivate the system.

NIST further recommends testing before deployment and regularly afterwards. The system must be able to fail safely when it goes beyond the limits of its knowledge. Asking for human review instead of guessing is exactly that kind of mechanism. NIST makes clear that its recommendations are not a checklist, but they help you avoid overlooking the important questions.

Escalation hands the request over to a person. Along with it go the context, the drafted proposal and any open questions. The colleague who takes over doesn't have to start from scratch.

Customers value this highly. The same Gartner survey shows that their main fear is that it will become harder to reach a human. Gartner recommends that the AI tells the customer it is connecting them to a person when it can't find a solution, and that the conversation continues where it left off.

Transparency also has a legal dimension. Article 50 of the AI Act applies from 2 August 2026 and requires systems that interact directly with people to be designed so that users know they are talking to an AI, unless this is obvious. According to Baker Botts, violations can lead to fines of up to €15 million or 3% of worldwide turnover. Check the official text with your legal counsel.

Integrating an AI customer support agent is limited to what the pilot requires. That might be an API, channels, customer history or message sending. Each element is confirmed technically, not assumed.

Evaluation is based on criteria agreed in advance:

The last criterion is often decisive. If review means redoing the work entirely, the operational benefit needs to be reconsidered. Expansion is discussed only once the team understands the results, the limits and the reasons for human intervention.

In the Aurora catalogue, the role associated with support is Sofia, the customer support agent, with the title “Customer Support Coordinator” and the API ID sofia. The association indicates a role, not a guarantee. Availability, limits and integration are reconfirmed for your case.

Sofia's page describes several limits: she doesn't promise compensation, discounts or new deadlines without approval, and she immediately hands complaints and legal threats over to a person. During setup, the guardrails and handover rules are defined with you and tested in a sandbox.

Prepare the following for the conversation:

• the request type included and the one deliberately excluded;
• the approved knowledge base and its owner;
• the output format: reply, sources, points to check;
• who approves each reply and who handles escalations;
• the integrations that are strictly necessary;
• the evaluation criteria.

On cost: the pricing page lists €699 per agent per month, excluding VAT, plus a one-off setup fee of €4,499 per company. You can start with a single agent. You receive the exact scope and price in a written proposal.

If the process involves sensitive data, you can request an NDA before sharing details. Once your list is ready, book a conversation with the AuroraAISystems team. Describe the process, the tools involved, who does what and roughly how many requests you handle per week.